HSE SCIENTIFIC JOURNAL 4...2019/02/26 · 4(46)-2018 HSE SCIENTIFIC JOURNAL Publisher: National...
Transcript of HSE SCIENTIFIC JOURNAL 4...2019/02/26 · 4(46)-2018 HSE SCIENTIFIC JOURNAL Publisher: National...
-
№4(46)-2018
HSE SCIENTIFIC JOURNAL
Publisher: National Research University Higher School of Economics
Subscription index in the Rospechat catalog –
80870
The journal is published quarterly
The journal is included into the list of peer reviewed scientific editions established by the Supreme Certification
Commission of the Russian Federation
Editor-in-Chief:A. Golosov
Deputy Editor-in-ChiefS. Maltseva
Y. Koucheryavy
Computer Making-up:O. Bogdanovich
Website Administration:I. Khrustaleva
Address: 33, Kirpichnaya Street, Moscow,
105187, Russian Federation
Tel./fax: +7 (495) 771-32-38http://bijournal.hse.ru
E-mail: [email protected]
Circulation: English version – 300 copies, Russian version – 300 copies,
online versions in English and Russian – open access
Printed in HSE Printing House3, Kochnovsky Proezd, Moscow,
Russian Federation
© National Research University Higher School of Economics
Data analysis and intelligence systems
Yu.P. Yekhlakov, E.I. Gribkov
User opinion extraction model concerning consumer properties of products based on a recurrent neural network ........................................................ 7
T.S. Stankevich
The use of convolutional neural networks to forecast the dynamics of spreading forest fires in real time ......................... 17
Information systems and technologies in business
E.V. Vasilieva, V.N. Pulyaeva, V.A. Yudina
Digital competence development of state civil servants in the Russian Federation ............................................... 28
S.P. Novikov, O.V. Mikheenko, N.A. Kulagina, O.D. Kazakov
Digital registry of professional competences of the population drawing on distributed registries and smart contracts technologies.................................................. 43
M.A. Shtefan, J.M. Elizarova
Investment project efficiency and risk evaluation: an integrated approach ................................................................ 54
Modeling of social and economic systems
G.L. Beklaryan
Decision support system for sustainable economic development of the Far Eastern Federal District ........................... 66
-
BUSINESS INFORMATICS No. 4(46) – 2018
2
ABOUT THE JOURNAL
Business Informatics is a peer reviewed interdisciplinary academic journal published since 2007 by National Research University Higher School of Economics (HSE), Moscow, Russian Federation. The journal is administered by School of Business Informatics.
The journal is published quarterly.
The mission of the journal is to develop business informatics as a new field within both information
technologies and management. It provides dissemination of latest technical and methodological
developments, promotes new competences and provides a framework for discussion in the field of
application of modern IT solutions in business, management and economics.
The journal publishes papers in the areas of, but not limited to:
data analysis and intelligence systems
information systems and technologies in business
mathematical methods and algorithms of business informatics
software engineering
Internet technologies
business processes modeling and analysis
standardization, certification, quality, innovations
legal aspects of business informatics
decision making and business intelligence
modeling of social and economic systems
information security.
The journal is included into the list of peer reviewed scientific editions established by the Supreme
Certification Commission of the Russian Federation.
The journal is included into Russian Science Citation Index (RSCI) database on the Web of
Science platform.
International Standard Serial Number (ISSN): 2587-814X (in English), 1998-0663 (in Russian).
Editor-in-Chief: Dr. Alexey Golosov – President of FORS Development Center, Moscow,
Russian Federation.
-
EDITOR-IN-CHIEFAlexey Golosov FORS Development Center, Moscow, Russia
DEPUTY EDITOR-IN-CHIEFSvetlana Maltseva National Research University Higher School of Economics, Moscow, Russia
Yevgeni Koucheryavy Tampere University of Technology, Tampere, Finland
EDITORIAL BOARDHabib Abdulrab National Institute of Applied Sciences, Rouen, France
Sergey Avdoshin National Research University Higher School of Economics, Moscow, Russia
Andranik Akopov National Research University Higher School of Economics, Moscow, Russia
Fuad Aleskerov National Research University Higher School of Economics, Moscow, Russia
Alexander Afanasyev Institute for Information Transmission Problems (Kharkevich Institute), Russian Academy of Sciences, Moscow, Russia
Anton Afanasyev Central Economics and Mathematics Institute, Russian Academy of Sciences, Moscow, Russia
Eduard Babkin National Research University Higher School of Economics, Nizhny Novgorod, Russia
Sergey Balandin Finnish-Russian University Cooperation in Telecommunications (FRUCT), Helsinki, Finland
Vladimir BarakhninInstitute of Computational Technologies, Siberian Branch of Russian Academy of Sciences, Novosibirsk, Russia
Alexander Baranov Federal Tax Service, Moscow, Russia
Jorg BeckerUniversity of Munster, Munster, Germany
Vladimir Belov Ryazan State Radio Engineering University, Ryazan, Russia
Alexander Chkhartishvili V.A. Trapeznikov Institute of Control Sciences, Russian Academy of Sciences, Moscow, Russia
Vladimir Efimushkin Central Research Institute of Communications, Moscow, Russia
Tatiana Gavrilova Saint-Petersburg University, St. Petersburg, Russia
Herv GlotinUniversity of Toulon, La Garde, France
Andrey Gribov CyberPlat Company, Moscow, Russia
Alexander Gromoff National Research University Higher School of Economics, Moscow, Russia
Vladimir Gurvich Rutgers, The State University of New Jersey, Rutgers, USA
Laurence Jacobs University of Zurich, Zurich, Switzerland
Liliya Demidova Ryazan State Radio Engineering University, Ryazan, Russia
EDITORIAL BOARD
Iosif Diskin Russian Public Opinion Research Center, Moscow, Russia
Nikolay Ilyin Federal Security Guard of the Russian Federation, Moscow, Russia
Dmitry Isaev National Research University Higher School of Economics, Moscow, Russia
Alexander Ivannikov Institute for Design Problems in Microelectronics, Russian Academy of Sciences, Moscow, Russia
Valery Kalyagin National Research University Higher School of Economics, Nizhny Novgorod, Russia
Tatiana Kravchenko National Research University Higher School of Economics, Moscow, Russia
Sergei Kuznetsov National Research University Higher School of Economics, Moscow, Russia
Kwei-Jay LinNagoya Institute of Technology, Nagoya, Japan
Mikhail Lugachev Lomonosov Moscow State University, Moscow, Russia
Peter Major UN Commission on Science and Technology for Development, Geneva, Switzerland
Boris Mirkin National Research University Higher School of Economics, Moscow, Russia
Vadim Mottl Tula State University, Tula, Russia
Dmitry Nazarov Ural State University of Economics, Ekaterinburg, Russia
Dmitry Palchunov Novosibirsk State University, Novosibirsk, Russia
Panagote (Panos) Pardalos University of Florida, Gainesville, USA
scar PastorPolytechnic University of Valencia, Valencia, Spain
Joachim Posegga University of Passau, Passau, Germany
Kurt Sandkuhl University of Rostock, Rostock, Germany
Yuriy Shmidt Far Eastern Federal University, Vladivostok, Russia
Christine Strauss University of Vienna, Vienna, Austria
Ali Sunyaev Karlsruhe Institute of Technology, Karlsruhe, Germany
Victor Taratukhin University of Munster, Munster, Germany
Jos TriboletUniversidade de Lisboa, Lisbon, Portugal
Olga Tsukanova Saint-Petersburg National Research University of Information Technologies, Mechanics and Optics, St. Petersburg, Russia
Mikhail Ulyanov V.A. Trapeznikov Institute of Control Sciences, Russian Academy of Sciences, Moscow, Russia
Raissa Uskenbayeva International Information Technology University, Almaty, Kazakhstan
Marcus Westner Regensburg University of Applied Sciences, Regensburg, Germany
-
BUSINESS INFORMATICS No. 4(46) – 2018
4
ABOUT THE HIGHER SCHOOLOF ECONOMICS
Consistently ranked as one of Russia’s top universities, the Higher School of Economics (HSE) is a leader in Russian education and one of the preeminent economics and social sciences universities in Eastern Europe and Eurasia.
Having rapidly grown into a well-renowned research university over two decades, HSE sets itself apart with its international presence and cooperation.
Our faculty, researchers, and students represent over 50 countries, and are dedicated to maintaining the highest academic standards. Our newly adopted structural reforms support both HSE’s drive to internationalize and the groundbreaking research of our faculty, researchers, and students.
Now a dynamic university with four campuses, HSE is a leader in combining Russian educational traditions with the best international teaching and research practices. HSE offers outstanding educational programs from secondary school to doctoral studies, with top departments and research centers in a number of international fields.
Since 2013, HSE has been a member of the 5-100 Russian Academic Excellence Project, a highly selective government program aimed at boosting the international competitiveness of Russian universities.
-
BUSINESS INFORMATICS No. 4(46) – 2018
5
ABOUT THE SCHOOL OF BUSINESS INFORMATICS
The School of Business Informatics is one of the leading divisions of HSE’s Faculty of Business and Management. The School offers students diverse courses taught by full-time HSE instructors and invited business practitioners. Students are also given the opportunity to carry out fundamental and applied projects at various
academic centers and laboratories.
Within the undergraduate program, students participate each year in different case-
competitions (PWC, E&Y, Deloitte, Cisco, Google, CIMA, Microsoft Imagine CUP,
IBM Smarter Planet, GMC etc.) and some of them are usually as being best students by
IBM, Microsoft, SAP, etc. Students also have an opportunity to participate in exchange
programs with the University of Passau, the University of Munster, the University of
Business and Economics in Vienna, the Seoul National University of Science and
Technology, the Radbound University Nijmegen and various summer schools (Hong
Kong, Israel etc.). Graduates successfully continue their studies in Russia and abroad,
start their own businesses and are employed in high-skilled positions in IT companies.
There are four graduate programs provided by the School:
Business Informatics
E-Business;
Information Security Management;
Big Data Systems.
The School’s activities are aimed at achieving greater integration into the global
education and research community. A member of the European Research Center for
Information Systems (ERCIS), the School cooperates with leading universities and
research institutions around the world through academic exchange programs and
participation in international educational and research projects.
-
BUSINESS INFORMATICS No. 4(46) – 2018
7
User opinion extraction model concerning consumer properties of products based on a recurrent neural network1
Yuri P. YekhlakovProfessor, Department of Data Processing Automation Tomsk State University of Control Systems and RadioelectronicsAddress: 40, Prospect Lenina, Tomsk, 634050, RussiaE-mail: [email protected]
Egor I. GribkovDoctoral Student, Department of Data Processing Automation Tomsk State University of Control Systems and RadioelectronicsAddress: 40, Prospect Lenina, Tomsk, 634050, RussiaE-mail: [email protected]
Abstract
This article off ers a long short-term memory (LSTM) based structured prediction model taking into account existing approaches to sequence tagging tasks and allowing for extraction of user opinions from reviews. We propose a model confi guration and state transition rules which allow us to use past predictions of the model alongside sentence features. We create a body of annotated user reviews about mobile phones from Amazon for model training and evaluation. The model trained on reviews corpus with recommended hyperparameter values. Experiment shows that the proposed model has a 4.51% increase in the F1 score for aspects detection and a 5.44% increase for aspect descriptions compared to the conditional random fi eld (CRF) model with the use of LSTM when F1 spans are matched strictly.
The extraction of user opinions on mobile phones from reviews outside of the collected corpus was conducted as practical confi rmation of the proposed model. In addition, opinions from other product categories like skin care products, TVs and tablets were extracted. The examples show that the model can successfully extract user opinions from diff erent kinds of reviews. The results obtained can be useful for computational linguists and machine learning professionals, heads and managers of online stores for consumer preference determination, product recommendations and for providing rich catalog searching tools.
Key words: user feedback; deep learning; machine learning; natural language processing; opinion processing.
Citation: Yekhlakov Yu.P., Gribkov E.I. (2018) User opinion extraction model concerning consumerproperties of products based on a recurrent neural network. Business Informatics, no. 4 (46), pp. 7–16.DOI: 10.17323/1998-0663.2018.4.7.16
1 This study was conducted under government order of the Ministry of Education and Science of Russia, project No. 8.8184.2017/8.9
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
8
Introduction
Potential consumers face a challenging choice when looking for complicated technical devices (for example, a tel-ephone, a refrigerator or a TV). The existence of a large number of manufacturers and model lines on the market, the variability of possible specifications of goods lead to a compound growth in the number of possible options for consumer properties of similar products. The consumer either randomly chooses a product relying on a brand name or makes a decision guided by advertising. With the rapid growth of the internet and social networks, most of pragmatic consumers are guided not only by advertisements, but also by user opinions of the consumer properties that become appar-ent over time.
Manufacturers are also interested in better understanding of their clients: which goods do they prefer, what pros and cons of product fea-tures do they notice. Based on this data, deci-sion makers can create a product assortment, provide individual selection of goods and ser-vices, make special offers to clients and do other activities to raise loyalty of clients and increase competitiveness.
There are a great number of sources that can provide you with user opinions about goods. They can be thematic forums, review articles and videos and social network communities. Chain stores let their clients provide reviews of goods they bought on their sites. Aggrega-tors like “Yandex.Market” can ease the search of this kind of information by collecting user comments in one place along with the ability to rate the usefulness of the information con-tained.
However, the majority of such platforms only solve a problem of gathering informa-tion in one place, without analyzing and gen-eralizing information automatically. Users are forced to study and analyze the reviews on their own, which can be problematic consid-
ering the large amount of information. Natu-ral language processing (NLP) methods based on machine learning can deliver a well-devel-oped solution and practical representation of information for analysis of the user opinion extraction task concerning consumer proper-ties of products.
Nowadays, most of the papers devoted to user review processing are based on senti-ment analysis of the text. At the same time, sentiment is considered as an attribute of the whole text or its large parts (paragraphs and sentences) [1, 2], which is not sufficient for user opinion extraction. Research in the area of aspect-oriented sentiment analysis is devoted to the problems of searching for def-inite aspect mention of products (consumer properties and features) and determining the user attitude to them in general, by means of placing them into one of the categories: good, bad, neutral, unknown [3, 4]. The most devel-oped problem statement of defining user atti-tude towards a product is a detailed sentiment analysis [5], where it is suggested to label aspects and opinions in review texts, where the user expresses sentiments about the given aspect. There are a number of works where conditional random fields [3, 6, 7] are used for detailed sentiment analysis. Syntactic ana-lyzer results (part-of-speech tagging, depend-ency trees, immediate constituent trees etc.) are used as inputs for segmentation. Such analyzers are not available for a vast number of natural languages. In addition, the accu-racy of these analyzers depends on the nature of training data. Moreover, the given model uses special glossaries: emotional, sentiment etc. Recent advances in deep learning based NLP statistic modeling allows us to avoid using extra features when training the models [8, 9]. Such models on their own learn neces-sary features for problem solving during train-ing.
In this paper, we offer a long short-term memory (LSTM) model for user opinion
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
9
extraction which does not require the pres-ence of such analyzers and relies on pre-trained vector representation). For training and quality testing, we use the labeled corpus (dataset) of mobile phone reviews marked by us.
1. Construction of the training sample for user opinions
extraction from texts
There are few datasets for fine-grained sen-timent analysis available at this moment. So, [10] describes the USAGE corpus – an anno-tated set of 800 Amazon user reviews from 8 categories about rather simple electronic devices like toasters and coffee machines. The proposed review texts annotation scheme consists of the following entities:
aspect is an important product property or mention of it (with indication if aspect belongs to the product which is described in reviews);
description is a text that contains the user’s opinion about aspect (with sentiment);
coreference resolution is used for cases where aspect refers to an entity in another sentence in the review;
aspect–description link – for grouping together related aspects and descriptions.
Although the USAGE corpus is available, the authors decided to make an additional annotated corpus. This decision was moti-vated by two factors. First, the desire to eval-uate the quality of the opinion extraction algorithms on the more complex and featured packed products (so we annotated reviews about mobile phones). Second, due to human resources limitations, we decided to use a simpler annotation scheme compared to the one used in USAGE (we dropped coreference resolution and annotated only full aspect–description pairs).
Annotation was done on the user reviews
corpus from the Amazon online store pre-sented in [11]. This corpus contains 143 mil-lion review texts about 25 product categories written during the period from May 1996 to July 2014 along with metadata about product title, identifier and product description, prod-uct category, brand, price, author identifier and user rating. The annotators were asked to mark aspect and description spans within review texts. In contrast to the USAGE data-set, we gave strict instructions to mark only full aspect-description pairs that together form user opinions. Below “opinion” will be referred as a pair
,
where – opinion’s aspect starting with a word on position and ending with a word on position ;
D (d
begin, d
end ) – opinion’s description starting
with a word on position dbegin
and ending with a word on position d
end.
Herein spans from different opinions should not intersect. Thus, an annotated sentence from a review text with opinion O
1 consist-
ing of aspect with associated span A (1, 2) and description with associated span D (4, 5) can be presented this way:
3,232 reviews were annotated in total. The annotated corpus contains 9,344 opinions, 1,994 unique aspects, 5,124 unique descrip-tions. The quantitative description of the anno-tated corpus is presented in Table 1.
2. User opinion extraction model
Tasks of the user opinion extraction model can be presented as a sequence tagging task where for each element of input sequence a class label (tag) should be determined. This requires spans of opinions to be reshaped as
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
10
a tag sequence. One of the most popular ways to represent span as a sequence is an IOB for-mat [12]. It uses three different tag types: O – absence of any particular entity; B-X – begin-ning of entity of type X; I-X – continuation of entity of type X. This paper proposes two types of entities – aspects and descriptions denoted with labels “Aspect” and “Description”. Then the set of possible tags Y will contain the fol-lowing elements: O – absence of any par-ticular entity; B-Aspect – beginning of the aspect; I-Aspect – continuation of aspect; B-Description – beginning of the description; I-Description – continuation of description. This way we can uniquely associate a sentence containing a set of opinions with the sequence of tags as shown in the Figure 1.
As a formal technique for solving this prob-lem, we proposed to use recurrent neural net-works (RNNs). This kind of neural networks is widely used to solve a broad variety of machine learning tasks like natural language mode-ling [13], part-of-speech tagging [8], sequence
classification [14], audio recognition [15], time series forecasting [16], etc.
The input of RNN at each time step t is the next element of an arbitrary sequence which is transformed into a sequence of outputs by recurrent relationships between the sequence of hidden states:
,
where xt – current input;
– previous hidden state;
U and W – input and recurrent transforma-tion matrices;
b – bias;
f – nonlinear activation function.
Recurrent connections between the hidden state allows for transferring contextual infor-mation about a sequence under processing and use of this information when predicting out-puts h
t . This way we can see h
t as an interme-
diate representation of sequence that accu-mulates information about the preceding steps
Table 1. Quantitative description of mobile phones user reviews corpus
Quantitative description Value
Number of reviews 3,232
Number of opinions in corpus 9,344
Number of unique aspects 1,994
Number of unique descriptions 5,124
Fig. 1. Correspondence between opinion spans and IOB tag sequence
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
11
of processing. In this paper, we use the LSTM recurrent neural network [17] because it is less exposed to the vanishing gradient problem [18] compared to simple RNN:
(1)
A classifier for word tag determination is used in the following way: to each output value we apply linear transformation and the softmax function, resulting in a possible tag probability distribution:
(2)
From the general RNN definition, it follows that information propagates in left-to-right order within the network. In some cases, it can be useful to know the context of subsequent words for correct classification of the current word. In order to allow the use of information from both directions, in each step of the pre-diction a bidirectional neural network is used, one of which processes input sequence in left-to-right order and other processes it in right-to-left order, after which the hidden states corresponding to the same position are con-catenated:
It should be noted that in structured predic-tion tasks (which include the task of sequence tagging) there are dependencies between tags in the output sequence. Therefore, models that don’t take these dependencies into account can produce ill-formed tag sequences. For exam-ple, when predicting tag sequence in the IOB
format sequence of predictions the O I-Aspect is wrong because the tag I-Aspect can only fol-low after the corresponding B-Aspect tag.
To model correlations between different pre-dictions within the same sequence, it is pro-posed to use the conditional random field (CRF) model that was proposed in [19]:
(3)
where A – matrix with probability of transition from tag y
i to tag y
i+1;
– probability of tag on position i ;
W {w1, w
2, ..., w
n } – input sequence;
y {y1,
..., y
n } – predicted sequence.
Then the probability of the sequence of pre-dictions y is evaluated as follows:
(4)
where summation in the denominator happens by all possible sequences y.
During training of the conditional random filed model, the log-probability of the true tag sequence is maximized:
(5)
An optimal sequence of predictions can be computed using dynamic programming. In doing so, the optimal sequence of predictions should correspond to the maximum of expres-sion as follows:
Using the model of conditional random field allows us to predict a globally optimal structure only in case the linear structure is considered and only local features are used for each node of prediction. This limitation led to the devel-
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
12
opment of structured prediction methods that can handle more complex structures (trees, for example) and use non-local features (like word classification results from previous steps) [20, 21]. Accordingly, for word tag prediction the following expression is proposed:
(6)
where ct – model configuration at the moment t;
– function for mapping model configu-ration c
t to feature set.
Applying the expression (6) for the user opin-ion extraction task in practice requires that we define the form of configuration c and state-to-features mapping function . In our work, we use inspiration from the dependency pars-ing model presented in [22]. Configuration is defined as 4-tuple c = (S, B, l, Y ) consisting of the buffer B that holds unprocessed elements of input sequence, the stack S that holds words from all currently found entities, the tag of last found entity l (Aspect or Description) and par-tially constructed output sequence Y. Feature vector is formed at each step t which is used to determine tag of current word and change model configuration according to this tag. The model configuration transition rules presented in Table 2. The semicolon symbol in the table denotes sequence concatenation.
The following form of is proposed based on model configuration c
t and configuration
transition rules:
where Bi – i-th element of buffer B;
– j-th element of stack S on the step t ;
– k-th row of matrix E;
– n-th element of predicted tags sequence on the step t.
We use hidden states from the last layer of multilayer bidirectional LSTM as the elements of the buffer. In this regard, every element of the buffer will contain information not only about the word at the corresponding position, but also about the preceding and subsequent context. The authors assume that information about words in a stack of found entities will contribute to the accuracy of a starting posi-tion detection for subsequent entities. For example, the “battery life” aspect found may give a hint to the model that the next words “is perfect” are the description. In addition, an extra hint is the tag of the previously found entity . Rows of matrix serve as features for l. Input for LSTM are pretrained vector representations of words obtained with the use of the FastText model [23]. This work uses vectors2.
Table 2.Model configuration transition rules
Precondition
B – y bt ; S t y I – y bt ; S t l t
S t l t
2 Source: https://github.com/plasticityai/magnitude
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
13
When calculating tag probabilities only tags that obey precondi-
tions from Table 2 are considered. This makes it possible to avoid an ill-formed prediction sequence.
3. Sequence tagging algorithm
Based on the foregoing, the sequence tagging algorithm can be framed as follows.
Let the input text be given as a sequence of words W {w
1, w
2, ..., w
n}. We should deter-
mine the output sequence of word tags Y {y
1, y
2 ..., y
n}. The Adam optimization method
[24] with parameters lr = 0.001, = 0.9, = 0.999 and gradient clipping at 3.0 is used to
train the model.
Step 1. Initialize model state as B = , C = , l = 0, Y = , t = 0.
Step 2. Fill buffer with hidden states from input sequence processed by LSTM network:
.
Step 3. If t < n, then go to step 4, otherwise go to Step 6.
Step 4. Make the feature vector and determine the tag for current position in buffer:
.
Step 4. Change S, i, Y according to the rules from Table 2 depending on the tag .
Step 5. t = t + 1, go to Step 3.
Step 6. End.
The results obtained from processing mobile phone reviews from the Amazon online store by proposed model and algorithm are described in Table 3. For example, from the sentence “The screen is fantastically large while the overall dimensions of the phone are manageable for those without giant hands” opinions “screen is fantastically large” and “dimensions of the phone are manageable” were extracted.
Furthermore based on the annotated data-set, user opinions on other product categories were processed (Table 4). The results obtained suggest that the model has shown good results both for extracting opinions on mobile phones and on products of other categories.
4. Experimental evaluation of the model
Experimental evaluation of the proposed model was done in comparison with the bidi-rectional CRF–LSTM model without char-acter features from [22]. Models were trained with the backpropagation method. Parameters were optimized by Adam [24] with the follow-ing parameters: lr = 0.001, = 0.9, = 0.99 and gradient clipping at 3.0.
Table 3. Opinions extracted from mobile phone reviews
Product Opinions
Sony Xperia XAPhone is awesome; phone is easy to use; phone is perfect for those who need extra storage; battery life is mediocre; battery life is absolutely terrible compared; no great sound
Apple iPhone 6S It didn’t work properly from the beginning; it’s a decent; bad charger; worry free product
Huawei P20Phone is a flagship performer; phone stopped receiving phone calls; phone is absolutely amazing for the price; the screen is fantastically large; camera is simply amazing; fantastic camera; camera produces great photos
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
14
The dataset described in section 1 was used for user opinions extraction. Because of the small sample size, estimation was done by 5-fold cross-validation. To exclude the influ-ence of the lexicon of train sample during testing, we split data at the documents level.
The proposed model and CRF-based one both use 2-layer bidirectional LSTM with 100 dimensional hidden state; the input vec-tor size is 100. We used the following hyper-parameters for mapping (t): C
BT = 2, C
ST = 4,
.
During testing, we tracked a set of criteria common for such kinds of tasks: precision, recall and F1-measure which determines the
overall quality of the model. Values of crite-ria were calculated in two ways: strict – when the match counted only if the found span is the same as true span; soft – when match counted if found and true spans have at least one com-mon word.
Analysis of Tables 5 and 6 reveals that the proposed model shows better performance than the Bi-LSTM-CRF model for both ways of criteria calculation. For strict match-ing, absolute improvement in the aspect span detection is 4.51%; in the description span detection – 5.44%. For soft matching, abso-lute improvement in the aspect span detection is 3.77%; in the description span detection – 3.52%.
Table 4. Opinions extracted
from reviews of other product types
Product Opinions
EltaMD PM Therapy Facial Moisturizer
is a great night cream product; product highly emollient without being greasy; product recommended by my dermatologist; very moisturized skin feel
Samsung UN55MU6500 Curved 55-Inch 4K Ultra HD Smart LED TV
Easy setup TV; TV is not worth the money or aggravation; wonderful picture; picture is so clear; remote is easy to use; remote is ergonomic and a breeze to use; color is unbelieveable
Fire 7 Tablet with AlexaSmall tablet; amazing little tablet; screen does not react well to water; screen is freezing; battery doesn’t hold a charge; battery dies to quickly; charging port its weakness
Table 5.Results of extracting opinions
from mobile phones dataset (strict matching)
ModelAspects Description
R P F1 R P F1
Bi-LSTM + CRF 39.20 50.58 44.17 41.30 54.03 46.82
Proposed model 47.87 49.51 48.68 49.70 52.93 52.26
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
15
Conclusion
The suggested approach as a combina-tion of the annotated dataset and an LSTM-based structured prediction model allows us to extract from review texts opinions on con-sumer properties of both the product as a whole and its features. The developed RNN-based structured prediction model is capable of using non-local features for entities prediction and does not require additional syntactic features.
The model trained on dataset has shown bet-ter results compared to the CRF-based model: F1 for aspect extraction is higher by 4.51%, F1 for description extraction is higher by 5.44%.
The experiments carried out have revealed that extra features in the base model have positive effect on the results.
The results obtained can be useful for NLP and computational linguist specialists, for the business community when selling goods, pro-viding services and developing their consumer properties.
In follow-on papers we will discuss questions on the prediction of links between aspects and descriptions for better opinion extraction and also incorporating information about opinion sentiment, as well as indication if the aspect corresponds to the product under review and coreference resolution.
Table 5.Results of extracting opinions
from mobile phones dataset(strict matching)
ModelAspects Description
R P F1 R P F1
Bi-LSTM + CRF 53.93 63.05 57.83 56.09 64.93 60.19
Proposed model 61.08 62.74 61.9 62.49 64.98 52.26
References
1. Sadegh M., Ibrahim R., Othman Z.A. (2012 Opinion mining and sentiment analysis: A survey. International Journal of Computers & Technology, vol. 2, no. 3, pp. 171–178.
2. Zhang L., Wang S., Liu B. (2018) Deep learning for sentiment analysis: A survey. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, vol. 8, no. 4, pp. 1942–4787.
3. Pontiki M., Galanis D., Papageorgiou H., Androutsopoulos I., Manandhar S., Mohammad Al-S., Al-Ayyoub M., Zhao Y., Qin B., De Clercq O. (2016) SemEval-2016 task 5: Aspect based sentiment analysis. Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval–2016). San Diego, CA, USA, 16–17 June 2016, pp. 19–30.
4. Jo Y., Oh A.H. (2011) Aspect and sentiment unification model for online review analysis. Proceedings of the Fourth ACM International Conference on Web Search and Data Mining (WSDM 2011). Hong Kong, China, 9–12 February 2011, pp. 815–824.
5. Zirn C., Niepert M., Stuckenschmidt H., Strube M. (2011) Fine-grained sentiment analysis with structural features. Proceedings of 5th International Joint Conference on Natural Language Processing (IJCNLP 2011). Chiang Mai, Thailand, 8–13 November 2011, pp. 336–344.
6. Yang B., Cardie C. (2012) Extracting opinion expressions with semi-Markov conditional random fields. Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP–CoNLL 2012). Jeju Island, Korea, 12–14 July 2012, pp. 1335–1345.
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
16
7. Yang B., Cardie C. (2013) Joint inference for fine-grained opinion extraction. Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics. Sofia, Bulgaria, 4–9 August 2013. Vol. 1, pp. 1640–1649.
8. Collobert R., Weston J., Bottou L., Karlen M., Kavukcuoglu K., Kuksa P. (2011) Natural language processing (almost) from scratch. Journal of Machine Learning Research, no. 12, pp. 2493–2537.
9. Zhai F., Potdar S., Xiang B., Zhou B. (2017) Neural models for sequence chunking. Proceedings of the Thirty-First Conference on Artificial Intelligence (AAAI-17). San Francisco, CA, USA, 4–9 February 2017, pp. 3365–3371.
10. Klinger R., Cimiano P. (2014) The USAGE review corpus for fine-grained, multi-lingual opinion analysis. Proceedings of the Language Resources and Evaluation Conference (LREC 2014). Reykjavik, Iceland, 26–31 May 2014, pp. 2211–2218.
11. He R., McAuley J. (2016) Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering. Proceedings of the 25th International Conference on World Wide Web (WWW 2016). Montreal, Canada, 11–15 April 2016, pp. 507–517.
12. Sang E.F., Veenstra J. (1999) Representing text chunks. Proceedings of the Ninth Conference on European Chapter of the Association for Computational Linguistics (EACL 1999). Bergen, Norway, 8–12 June 1999, pp. 173–179.
13. Mikolov T., Karafiat M., Burget L., Cernocky J., Khudanpur S. (2010) Recurrent neural network based language model. Proceedings of the Eleventh Annual Conference of the International Speech Communication Association (INTERSPEECH 2010). Makuhari, Chiba, Japan, 26–30 September 2010, vol. 2, pp. 1045–1048.
14. Ghosh M., Sanyal G. (2018) Document modeling with hierarchical deep learning approach for sentiment classification. Proceedings of the 2nd International Conference on Digital Signal Processing (ICDSP 2018). Tokyo, Japan, 25–27 February 2018, pp. 181–185.
15. Graves A., Jaitly N., Mohamed A. (2013) Hybrid speech recognition with deep bidirectional LSTM. Proceedings of the IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU 2013). Olomouc, Czech Republic, 8–12 December 2013, pp. 273–278.
16. Guo T., Xu Z., Yao X., Chen H., Aberer K., Funaya K. (2016) Robust online time series prediction with recurrent neural networks. Proceedings of the IEEE International Conference on Data Science and Advanced Ana-lytics (DSAA 2016). Montreal, Canada, 17–19 October 2016, pp. 816–825.
17. Hochreiter S., Schmidhuber J. (1997) Long short-term memory. Neural Computation, vol. 9, no. 8, pp. 1735–1780.
18. Bengio Y., Simard P., Frasconi P. (1994) Learning long-term dependencies with gradient descent is difficult. IEEE Transactions on Neural Networks, vol. 5, no. 2, pp. 157–166.
19. Lafferty J., McCallum A., Pereira F.C.N. (2001) Conditional random fields: Probabilistic models for segmenting and labeling sequence data. Proceedings of the 18th International Conference on Machine Learning (ICML 2001). Williamstown, MA, USA, 28 June – 1 July 2001, pp. 282–289.
20. Chen D., Manning C.D. (2014) A fast and accurate dependency parser using neural networks. Proceedings of the 19th Conference on Empirical Methods in Natural Language Processing (EMNLP 2014). Doha, Qatar, 25–29 October 2014, pp. 740–750.
21. Dyer C., Ballesteros M., Ling W., Matthews A., Smith N.A. (2015) Transition-based dependency parsing with stack long short-term memory. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing. Beijing, China, 27–31 July 2015, vol. 1, pp. 334–343.
22. Lample G., Ballesteros M., Subramanian S., Kawakami K., Dyer C. (2016) Neural architectures for named entity recognition. Proceedings of the 15th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL HTL 2016). San Diego, CA, USA, 12–17 June 2016, pp. 260–270.
23. Bojanowski P., Grave E., Joulin A., Mikolov T. (2017) Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics, no. 5, pp. 135–146.
24. Kingma D.P., Ba J.L. (2017) Adam: A method for stochastic optimization. arXiv:1412.6980v9 [cs. LG]. Available at: https://arxiv.org/pdf/1412.6980.pdf (accessed 10 October 2018).
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
17
1 This study was supported by the Russian Foundation for Basic Research, project No. 18-37-00035 “On the dependence of the dynamics of the development of forest fires on the influence of environmental factors, the nature of forest plantations and the type of fire under conditions of nonstationarity and uncertainty”
INFORMATION SYSTEMS AND TECHNOLOGIES IN BUSINESS
The use of convolutional neural networks to forecast the dynamics of spreading forest fires in real time1
Tatiana S. StankevichAssistant Professor, Department of Technosphere SafetyKaliningrad State Technical UniversityAddress: 1, Sovietsky Prospect, Kaliningrad, 236022, RussiaE-mail: [email protected]
Аbstract
This work focuses on the relevant task of increasing the effi ciency of forecasting the dynamics of forest fi res spreading in real time. To address the problem, it was proposed to develop a method for operational forecasting the forest fi re spread dynamics in the context of unsteadiness and uncertainty based on some advanced information technologies, i.e. artifi cial intelligence and deep machine learning (the convolutional neural network). As part of the research, both domestic and foreign models for the spread of forest fi res were evaluated, and the key limitations of using models in real fi re conditions were identifi ed (high degree of dynamism and uncertainty of input parameters, the need to ensure minimum collection time and input parameters, as well as minimum response time of the model). Based on the data obtained, the need to use artifi cial neural network tools to solve the problem of predicting the forest fi re’s spread dynamics was substantiated. A general logic diagram of the method for forecasting the forest fi re dynamics in real time has been developed, the main feature of which is the construction of a tree of convolutional neural networks. To enhance the quality of learning convolutional neural networks that implement the function of predicting the spread of forest fi res, we propose to create a database of forest fi re dynamics.
Key words: forest fire; database; visual data; artificial intelligence; deep machine learning; convolutionalneural network; big data; real-time forecasting.
Citation: Stankevich T.S. (2018) The use of convolutional neural networks to forecast the dynamics of spreading forest fires in real time. Business Informatics, no. 4 (46), pp. 17–27. DOI: 10.17323/1998-0663.2018.4.17.27
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
18
Introduction
The significant negative effect from forest fires for the Russian Federation is clearly confirmed by statistics from the Federal Forestry Agency (Rosleskhoz), as presented in the Unified Interdepartmen-tal Statistical Information System(UISIS)1. According to the statistics from 2009 to 2017, there was a 27.92% increase in the forest land covered by forest fires. From 2013 to 2017, the growth of expenses related to the protection and restoration of forest area of the Russian Federation was 22.32%. At the same time, the aforementioned trend data comes against the background of a decrease in the total number of forest fires in the country.
According to the European forest fire sta-tistics presented by the European Forest Fire Information System (EFFIS) [1], from 2009 to 2016 there was a 39.8% decrease in the total number of forest fires in the five southern member states of the European Union (Portu-gal, Spain, France, Italy and Greece). There was also a 4.6% decrease in the total forest area covered by forest fires in those countries. How-ever, the total forest area covered by forest fires in 2016 was 316,866 hectares, which is higher than in previous years (from 2013 to 2015), and the number of forest fires was 31,751, which is lower than the long-term average values and is slightly lower than that in the previous year of 2015 (38,171 fires) but higher than in 2014 (23,425 fires).
According to the forest fire statistics in the United States of America, presented by the National Centers for Environmental Informa-tion of the National Oceanic and Atmospheric Administration (NCEI NOAA)2, the num-ber of fires decreased by 14.5% from 2009 to 2017, and the forest area covered by forest fires increased by 65.4%.
Although statistics vary considerably from year to year (which clearly shows how forest fires depend on seasonal meteorological con-ditions), the global forest fire statistics show similar dynamics, i.e. a decrease in the num-ber of forest fires, an increase in the number of areas affected by forest fires, and an increase in the material expenses associated with forest fires.
Thus, it is extremely important for both the Russian Federation and other states to prevent, localise and eliminate forest fires.
One of the most important elements in addressing the problem is forecasting the for-est fire spread in a real-time mode. Currently, it is not easy to use existing models for pre-dicting the forest fire dynamics in difficult real fire conditions due to the limited func-tionality of models in the unstable and uncer-tain contexts.
The study is aimed at developing a method of forecasting of the forest fire dynamics in real time and in complex environments (with uncertainty and instability) by using artificial intelligence and deep machine learning. To achieve this goal it is necessary:
to justify the need to use artificial neural network tools to predict the forest fire spread dynamics;
to develop a common logic for the method of forecasting the forest fire dynamics in real time;
to create a visual database on the forest fire dynamics.
This work is part of a research project to identify the fundamental dependencies of the influence of environmental factors, the nature of forest plantations and the type of fire on the forest fire dynamics.
2 https://www.fedstat.ru/organizations/3 https://www.ncdc.noaa.gov/sotc/
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
19
4 http://firegrowthmodel.ca5 https://www.firelab.org/project/flammap6 https://www.firelab.org/project/farsite
1. Convolutional neural network as an advanced tool
for the forest fires spread forecasting in real time
According to [2], it is quite a challenge to simulate a fire in a forest due to two main rea-sons: the extreme complexity of the physical phenomenon (fire) due to heterogeneous fuel and many influencing environmental factors (wind, relative humidity etc.), and the diffi-culty of carrying out real experiments to vali-date the models developed.
Currently, both domestic and foreign researchers representing various fields of sci-ence have developed an extensive set of mod-els based on various methods for predict-ing the fire’s behaviour in order to minimise the destructive consequences of this natural emergency [3–5].
In studying the models for forecasting the forest fire’s spread, models are distinguished in terms of modelling into real-time, tactical and strategic models [6]. Since each model-ling level is characterised by a specific goal and an appropriate level of management deci-sions, each level corresponds to its model type; e.g. real-time models are designed for the real-time level, tactical ones for the tac-tical level and strategic ones for the strategic level.
It is common to single out the following key areas of forest fire modelling. [3–6]:
empirical and quasi-empirical models based on the statistical findings of experimen-tally obtained data to determine the statisti-cal dependencies between the input and out-put parameters;
physical and quasi-physical models based on the fundamental chemistry and/or physics methods for describing the processes occur-ring during a forest fire;
mathematical models (including simu-lation and wave models) that use formulas to describe the fire dynamics, in some cases, with statistical data.
In addition, according to the method for displaying the forest fire modelling results, existing models can be divided into spatial and non-spatial ones. Also, depending on the availability or unavailability of random varia-bles among the model parameters, they dis-tinguish between deterministic and stochastic prediction models. Since forest fires are char-acterised by complex conditions (uncertainty and instability), stochastic models are most promising.
The works [3–5] present the results of ana-lysing the key types of forest fire distribution models (empirical and quasi-empirical, physi-cal and quasi-physical, mathematical and imi-tation) developed from 1990 to 2007. In the work [7] in-depth research was undertaken into a 3% model, the quasi-empirical Rothermel model, Balbi model and Balbi non-stationary model. The works [8, 9] focus on various mod-els for forest fire forecasting, i.e. a mathemat-ical model for the escalation of surface forest fire and/or scrub fire, in which complete trees burn, and a discrete forest fire model on the upper half-plane.
Some of the models considered are inte-grated into computer systems and are com-monly used in practice. For example, in the Prometheus4 and FlamMap5 forest fire fore-casting systems, wave fire models are applied, where the combustion process is described with the Huygens principle [6, 10], and the fire propagation rate is calculated with experi-mental data. The use of the Van Wagner model and the quasi-empirical Rothermel model is based on a fire forecasting system such as FARSITE6 [2].
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
20
However, despite the wide variety of models for predicting forest fire escalation dynamics, during analysis of the reference sources, which consider the features and functionality of all types of models [2–15], we see limitations, which significantly affect the forecast accuracy, as follows:
a high degree of dynamic input parameters (the parameters dynamically changing in time are considered as constant);
a significant degree of uncertainty of the input parameters (inability to obtain a series of data by direct measurement).
In addition, the time used to collect and enter the input data as well as the response time of the model have a significant impact on the ability to use models in a real fire environment. When minimised, these time characteristics pose a critical problem in developing and using models in practice.
The recent breakthrough in the field of infor-mation technology, which has promoted the emergence and active improvement of prom-ising technologies – artificial intelligence, big data processing systems, and deep machine learning – have created unprecedented oppor-tunities to improve the fire safety of forests.
Nowadays, both the models for forecast-ing the forest fire breaking-out and models for forecasting the fire spread dynamics based on neural network technologies already are available (for example, the works [16, 17]). Although models which use artificial neural networks can eliminate a number of draw-backs inherent in traditional models, the con-struction and practical application of mod-els based on neural network technologies can be associated with some challenges. First of all, it should be noted that it is a challenge to collect a sufficient number of training exam-ples in preparing test and training data sets. In addition, the network architecture construc-tion can be characterised by complexity and labour intensity, and the network training pro-cedure is time-consuming.
Taking into account the above disadvan-tages, we proposed to develop a method for forecasting the forest fire spread dynamics in real time in case of non-stationarity and uncertainty with a convolutional neural net-work (CNN). The convolutional neural net-work, being a multi-layered neural network, is part of the deep learning technology and addresses the problem of pattern recognition from visual data [18, 19]. The features of the construction and operation of convolutional neural networks are described in detail in the work [18, 19].
The choice of a convolutional neural network is due to the advantages of the type of networks revealed as a result of analysis of both domestic and foreign sources [18, 19]. They are highly accurate, resistant to changes and input data distortions, real-time, capable of performing self-tuning, allow for paralleling high perfor-mance computing etc. In addition, although convolutional neural networks are commonly used to solve recognition and classification problems (for the classification of images, automatic speech recognition etc.), they can also be used for forecasting due to their indis-putable advantages.
The use of a convolutional neural network for forecasting the forest fire spread dynam-ics in real time makes it possible to generate a forecast in complex environments (with uncer-tainty and non-stationarity) and minimise time input due to paralleling high-performance computing. Thus, a convolutional neural net-work is an effective tool for obtaining a forest fire spread forecast in real time in the case of application in real environments.
2. Forecasting the forest fire dynamics in real time under non-stationary
and uncertain conditions with a convolutional neural network
During the research, a method for forecast-ing the forest fire dynamics in real time under the conditions of non-stationarity and uncer-
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
21
7 https://jointmission.gsfc.nasa.gov/viirs.html8 https://firms.modaps.eosdis.nasa.gov/9 https://www.ventusky.com/10 http://maps.elie.ucl.ac.be/CCI/viewer/
tainty was developed. A feature of the proposed method is to identify the dependencies of the influence of environmental factors, the nature of forest plantations and the type of forest fire on the forest fire dynamics with a convolutional neural network.
According to the works [18, 19], the idea of a CNN network is to implement a sequence of transitions from specific features of the visual input data to more abstract ones. The CNN network architecture can be characterised by alternating convolution layers and subsam-pling layers. The main purpose of the convo-lutional network layers is to implement the convolution operation with the subsequent creation of a feature map. Subsampling lay-ers of the network can reduce the dimension-ality of the previously created feature maps by selecting the maximum neuron from a num-ber of neighbouring neurons of the map and replacing the given neuron with the entire considered set of neurons. Fully connected layers are used as the output layer of neurons in the CNN network, where a fully connected neural network is created.
To develop a convolutional neural network in order to generate a forest fire forecast in real time, CPython software is proposed to use. At the same time, the fire spread informa-tion obtained in real time for three hours from a satellite with a moderate-resolution spec-troradiometer (36-channel spectroradiome-ter MODIS, Terra and Aqua satellites) and a visible infrared X-ray diffraction set (VIIRS) were used as the input visual data.
From 1999 to the present, MODIS has been one of the most widely used satellite tools for conducting global and regional research [20]. MODIS can be used to view the entire sur-face of the Earth every one or two days in 36
spectral bands at moderate resolutions rang-ing from 0.25 km to 1 km to obtain a data set (land and ocean surface temperatures, vegeta-tion indices, land cover data, forest fires, vol-canoes, clouds, aerosols etc.) [20]. VIIRS7 is a 22-channel radiometer that collects images in the visible, infrared and ultraviolet ranges (0.45–12 m) and performs radiometry of the land, atmosphere, cryosphere and oceans. The spatial resolution of the VIIRS data is in the range from 0.38–0.75 km (at nadir) to 0.8–1.6 km (at the edge of the zone) in the 3.000 km wide survey strip. The visual data on the fire spread is available in the NASA Fire Information for Resource Management System(FIRMS8).
The method developed for predicting the forest fire dynamics also includes using data on environmental factors (air temperature, air humidity and wind speed), data on the nature of forest plantations (type of forest stands) and data on the type of fire. The visual data on environmental factors are obtained with Ventusky InMeteo9; the data on the nature of forest plantations are obtained with the Land Cover Map10 of the Institute for Climate Change and the European Space Agency.
The existing NASA Earth Observation Sys-tem, as well as other global systems, provide sufficiently accurate information of various kinds in real time on the state of land, water and the planet’s atmosphere [21]. The infor-mation is publicly available and, by enriching the global information space, is widely used to improve the accuracy of meteorological fore-casts, environmental monitoring, pollution control etc.
Despite the advantages of the NASA Aer-ospace Earth Observation System and other global systems, countries are interested in
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
22
creating national satellite monitoring sys-tems. For example, in Mexico in 1999, a hot spot detection system (as an indicator of pos-sible forest fires) was introduced with day and night sensor images on NOAA satellites, and the possibility of creating a national system is under consideration [21].
This problem is also typical for the Russian Federation. Currently, Russia is planning to cre-ate a national aerospace earth observation sys-tem, the Multi-Purpose Aerospace Prediction Monitoring System11. However, at the moment, there is no real alternative to using sources other than those used in the research.
As an output, the proposed method provides for preparing a prediction of the fire escalation dynamics in real time in the form of a visual image, a map with a selected area with the coor-dinates of the fire’s spread area over time.
The general logical pattern of the method developed for predicting the forest fire dynam-ics in real time with non-stationarity and uncertainty based on a convolutional neural network is shown in Figure 1.
The method for forecasting the forest fire dynamics in real time with non-stationarity and uncertainty based on a convolutional neu-ral network includes the following steps:
Stage 1 (data input) – visual data input;
Stage 2 (preprocessing) – preprocessing of the input visual data to eliminate distorted ele-ments of the input image;
Stage 3 (building and setting up a convolu-tional neural network) – building a network with subsequent training with the backpropa-gation method;
Stage 4 (a forest fire forecast in real time) – identification of dependencies of the influ-ence of environmental factors, the nature of forest plantations and type of fire on the forest fire dynamics and (with the identified depen-
dencies) creation of an operational forest fire dynamics forecast in real time.
The main feature of the proposed method is the construction of a tree of convolutional neu-ral networks as a directed acyclic graph for ana-lysing a significant amount of visual data. This graph includes one root node, a CNN, which performs the last stage of forecasting, and three intermediate nodes, CNNs, where depen-dences of the influence of environmental fac-tors, the nature of forest plantations and the type of fire on the forest fire dynamics are cre-ated.
Thus, we developed a method for forecasting the forest fire dynamics in real time in the con-text of non-stationarity and uncertainty based on advanced information technologies, arti-ficial intelligence and deep machine learning (convolutional neural network). This type of network allows analysis of visual data, determi-nation of key dependencies of forest fire spread on environmental factors, the nature of for-est plantations and the type of forest fires, and drafting a fire escalation forecast in real time. The main feature of the proposed method is the construction of a tree of convolutional neu-ral networks.
3. Building a visual forest fire dynamics database
Since the quality of a convolutional neural network depends on the data set used to build the network and learn, it is necessary to cre-ate an appropriate database. To this end, an analysis of (hierarchical, network, relational, post-relational, object-oriented, multidimen-sional and object-relational) database models was performed with an expert ranking method and a modified hierarchy analysis method. On the basis of the database requirements created (large amount of data, visual data, the ability to quickly build/modify a database with the mini-
11 http://russianspacesystems.ru/bussines/bezopasnost/maksm/
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
23
Ñ1
S1
...
Ñn
Sn
FC
Ñ1
S1
...
Ñn
Sn
FC
Ñ1
S1
...
Ñn
Sn
FC
Ñ1
S1
...
Ñn
Sn
FC
Construction of a convolutional
neural network
Fig. 1. General logical pattern of the method for forecasting forest fire dynamics (C1, Cn – convolutional layers; S1, Sn – subsampling layers; FS – fully connected layer)
mal time and computational costs, the mini-mal time and computational costs when work-ing with a database), we proposed to develop a relational database model on the forest fire dynamics.
A visual database on the forest fire dynamics was built. Its elements are the Forest Fire, Envi-ronmental Factors and Nature of Forest Planta-tions tables. The Forest Fire table is intended for storing and displaying in a user-friendly form
Stage 1
Stage 2
Stage 3
Stage 4
Preprocessing
The dependences of the influence
of environmental factors, the nature of forest plantations
and the type of fire on the forest fire dynamics
Network training
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
24
a)
c)
b)
d)
Fig. 2. Space images of the area covered by forest fire (near the village of Tokma, Irkutsk Region, Russia, 58°1536 N 105°5224 E):
a) 16:00 GMT, 11.05.18; VIRS 375 m
b) 12:30 GMT, 12.05.18; VIRS 375 m
c) 13:48 GMT, 13.05.18; VIRS 375 m
d) 14:42 GMT, 21.05.18; VIRS 375 m
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
25
the visual data on the fire spread over time. The Environmental Factors table is for the vis-ual data on environmental factors. Finally, the Nature of Forest Plantations table is for the vis-ual data on the nature of the forest vegetation. The database created covers the period from 2018, and geographically all countries of the world. In addition, this database is constantly updated. To optimise the user experience with the database, the corresponding forms and requests were created.
An example of the fire history visual data is presented in Figure 2, which shows a vis-ual change in the fire history over four days as space images of an area (2a, II; 2b, II; 2c, II; 2d, II) with the Blue Marble map (the image resolution is 1,000 m). For illustration pur-poses, space images of the area (2a, II; 2b, II; 2c, II; 2d, II) are shown with the Topographic map (the image resolution is 10,000 m) and a conventional symbol of space orientation is added.
In addition, to facilitate the user data han-dling experience, we intend to create a web application. Currently, the plan is to protect IP assets by filing an application for state reg-istration of the database.
Conclusion
Our research provides insight into the exist-ing domestic and foreign models for predict-ing the forest fire escalation. Based on the findings, the main constraints of the applied models under real fire conditions were iden-tified, e.g. the highly dynamic and uncertain input parameters, the need to minimise the
time of input parameter collection and entry as well as minimise the response time of the model. There are grounds for using artificial neural network (convolutional neural net-work) tools to make it possible to forecast the forest fire spread dynamics, i.e. the possibility of generating a forecast in complex environ-ments of a real fire as well as the possibility of minimising time costs due to paralleling high-performance computing.
We developed a method for forecasting the forest fire dynamics in real time and in the conditions of non-stationarity and uncer-tainty using a convolutional neural network. The general logic pattern of the developed method is described (Figure 1). The main feature of the proposed method is the con-struction of a tree of convolutional neural networks as a directed acyclic graph for ana-lysing a significant amount of visual data. This graph includes one root node, a CNN, which performs the last stage of forecasting, and three intermediate nodes, the CNNs, where dependences of the influence of environmen-tal factors, the nature of forest plantations and the type of fire on the fire escalation dynam-ics are created.
An assessment of the existing database mod-els was carried out and the preferred version of the forest fire behaviour database model was selected, a relational database. A visual for-est fire behaviour database was built. Its ele-ments are the tables Forest Fire, Environmen-tal Factors and Nature of Forest Plantations. To optimise the user experience in handling the base, corresponding forms and requests were implemented.
References
1. European Commission (2017) Forest fires in Europe, Middle East and North Africa 2016. JRC Science for Policy Report. Available at: http://effis.jrc.ec.europa.eu/media/cms_page_media/40/Forest_fires_in_Europe_Middle_east_and_North_Africa_2016_final_pdf_JZU7HeL.pdf (accessed 03 August 2018).
2. Silva F.R., Guijarro M., Madrigal J., Jimenez E., Molina J.R., Hernando C., Velez R., Vega J.A. (2017) Assessment of crown fire initiation and spread models in Mediterranean conifer forests by using data from field and laboratory experiments. Forest Systems, vol. 26, no. 2. Available at: http://revistas.inia.es/index.php/fs/article/view/10652 (accessed 18 October 2018).
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
26
3. Sullivan A.L. (2009) Wildland surface fire spread modelling, 1990–2007. 1: Physical and quasi-physical models. International Journal of Wildland Fire, no. 18, pp. 349–368.
4. Sullivan A.L. (2009) Wildland surface fire spread modelling, 1990–2007. 2: Empirical and quasi-empirical models. International Journal of Wildland Fire, no. 18, pp. 369–386.
5. Sullivan A.L. (2009) Wildland surface fire spread modelling, 1990–2007. 3: Simulation and mathematical analogue models. International Journal of Wildland Fire, no. 18, pp. 387–403.
6. Khodakov V.E., Zharikova M.V. (2011) Lesnye pozhary: metody issledovaniya [Forest fires: research methods]. Kherson: Grin’ D.S. (in Russian).
7. Filippi J.B., Mallet V., Nader B. (2014) Evaluation of forest fire models on a large observation database. Natural Hazards and Earth System Sciences, vol. 14. no. 11, pp. 3077–3091. Available at: https://www.nat-hazards-earth-syst-sci.net/14/3077/2014/ (accessed 18 October 2018).
8. Perminov V., Goudov A. (2017) Mathematical modeling of forest fires initiation, spread and impact on environment. International Journal of GEOMATE, vol. 13, no. 35, pp. 93–99. Available at: http://www.geomatejournal.com/sites/default/files/articles/93-99-6704-Valeriy-July-2017-35-a1.pdf (accessed 18 October 2018).
9. Graf R. (2014) A forest-fire model on the upper half-plane. Electronic Journal of Probability, no. 19, pp. 1–27. Available at: https://projecteuclid.org/download/pdf_1/euclid.ejp/1465065650 (accessed 18 October 2018).
10. Lawson B.D., Armitage O.B., Hoskins W.D. (1996) Diurnal variation in the Fine Fuel Moisture Code: Tables and computer source code. FRDA Report 245. Victoria, B.C.: Canadian Forest Service, Pacific Forestry Center. Available at: https://www.for.gov.bc.ca/hfd/pubs/Docs/Frr/FRR245.pdf (accessed 18 October 2018).
11. Grishin A.M. (1992) Matematicheskoe modelirovanie lesnyh pozharov i novye sposoby bor’by s nimi [Mathematical modeling of forest fires and new ways of fighting them]. Novosibirsk: Nauka, Siberian Branch (in Russian).
12. Komorovsky V.S., Dorrer G.A. (2010) Metodika rascheta parametrov lesnyh pozharov kak dinamicheskih protsessov na poverhnosti zemli s ispol’zovaniem dannyh kosmicheskogo monitoringa [Method of calculating parameters of forest fires as dynamic processes on the earth’s surface using space monitoring data]. Vestnik SibGAU, no. 3 (29), pp. 47–50 (in Russian).
13. Rylkova O.I., Kataeva L.Yu., Maslennikov D.A., Romanova N.A., Rylkov I.V., Loshchilov A.A. (2013) Chislennoe modelirovanie lesnogo pozhara v lesah Vysokoborskogo lesnichestva Borskogo rayona Nizhegorodskoy oblasti [Numerical modeling of forest fire in the forests of Vysokoborsky forestry in the Bor District of the Nizhny Novgorod Region]. Modern Problems of Science and Education, no. 6. Available at: http://www.science-education.ru/ru/article/view?id=11671 (accessed 30 May 2018).
14. Maslennikov D.A., Kataeva L.Yu. (2011) Modelirovanie lesnyh pozharov v trekhmernoy sisteme koordinat s uchetom rel’efa [Modeling of forest fires in a three-dimensional coordinate system taking into account the landform]. Vestnik of Lobachevsky University of Nizhni Novgorod, no. 4 (5), pp. 2338–2340 (in Russian).
15. Perminov V.A. (2015) Matematicheskoe modelirovanie vozniknoveniya i rasprostraneniya verhovyh lesnyh pozharov v osrednennoy postanovke [Mathematical modeling of forest fires emergence and spread in the averaged setting]. Journal of Technical Physics, vol. 85, no. 2, pp. 24–30 (in Russian).
16. Yasinsky F.N., Potemkina O.V., Sidorov S.G., Evseeva A.V. (2011) Prognozirovanie veroyatnosti vozniknoveniya lesnyh pozharov s pomoshch’yu neyrosetevogo algoritma na mnogoprotsessornoy vychislitel’noy tekhnike [Predicting the probability of forest fires using neural network algorithm and multiprocessor computers]. Vestnik IGEU, no. 2. Available at: http://ispu.ru/files/str.82-84_0.pdf (accessed 31 March 2018) (in Russian).
17. Vahidnia M.H., Alesheikh A.A., Behzadi S., Salehi S. (2013) Modeling the spread of spatio-temporal phenomena through the incorporation of ANFIS and genetically controlled cellular automata: a case study on forest fire. International Journal of Digital Earth, vol. 6, no. 1, pp. 51–75.
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
27
18. Krizhevsky A., Sutskever I., Hinton G. (2012) ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems 25: 26th Annual Conference on Neural Information Processing Systems 2012 (NIPS 2012). Lake Tahoe, US. 3–8 December 2012. P. 1097–1105. Available at: https://www.cs.toronto.edu/~fritz/absps/imagenet.pdf (accessed 03 August 2018).
19. Hamed H.A., Elnaz J.H. (2017) Guide to convolutional neural networks. A practical application to traffic-sign detection and classification. Springer International Publishing.
20. Babu S., Roy A., Prasad R.C. (2016) Forest fire risk modeling in Uttarakhand Himalaya using TERRA satellite datasets. European Journal of Remote Sensing, no. 49, pp. 381–395. Available at: https://doi.org/10.5721/EuJRS20164921 (accessed 18 October 2018).
21. Zuniga-Vasquez J.M., Cisneros-Gonzalez D., Pompa-Garcia M., Rodriguez-Trejo D.A., Perez-Verdin G. (2017) Spatial modeling of forest fires in Mexico: An integration of two data sources. BOSQUE, vol. 38, no. 3, pp. 563–574. Available at: https://scielo.conicyt.cl/pdf/bosque/v38n3/art14.pdf (accessed 18 October 2018).
DATA ANALYSIS AND INTELLIGENCE SYSTEMS
-
BUSINESS INFORMATICS No. 4(46) – 2018
28
INFORMATION SYSTEMS AND TECHNOLOGIES IN BUSINESS
1 The article is based on the results of studies carried out using budget funds and the 2018 Financial University State Mission “Improvement of information support for personnel management system based on the competence approach and individual career tracking of civil servants”, state registra-tion number AAA-A18-118052490063-1
Digital competence development of state civil servants in the Russian Federation1
Elena V. Vasilieva Professor, Department of Business Informatics Financial University under the Government of the Russian FederationAddress: 38, Scherbakovskaya Street, Moscow, 105187, Russia E-mail: [email protected]
Valentina N. PulyaevaAssociate Professor, Department of Human Resources Management and Psychology Financial University under the Government of the Russian FederationAddress: 15, V. Maslovka Street, Moscow, 127083, RussiaE-mail: [email protected]
Vera A. Yudina Associate Professor, Department of Management, Information Science and Humanitarian ScienceFinancial University under the Government of the Russian Federation, Penza BranchAddress: 33В, Kalinina Street, Penza, 440052, Russia E-mail: [email protected]
Аbstract
In the international fi eld of public services, the competence approach is used as a basis for developing productivity, innovation and responsibility among employees. In Russia, the competence approach is central to legislative and regulatory documents but has not yet become a working tool. Russia’s transition to the digital economy in accordance with the Federal Program necessitates the transformation of professional qualities and qualifi cation requirements for positions of the state civil service. The development of a single information space of the state civil service and the widespread introduction of e-government technologies impose increased requirements on public servants’ competencies in the fi eld of information and communication technologies. However, studies have shown that until now, Russian civil servants consider as a primary priority only those competencies that focus on results, discipline, time and stress management skills, and to a lesser degree adaptability, willingness to change, creativity, initiative, and adopting new ideas and innovations. Management by competences requires an individual approach, taking into account the characteristics of each employee, as well as the development and implementation of competence models, in which all aspects of work in the digital world should be refl ected.
-
BUSINESS INFORMATICS No. 4(46) – 2018
29
INFORMATION SYSTEMS AND TECHNOLOGIES IN BUSINESS
Introduction
The current stage of economic and social development can be character-ized as the “new economy”, “innova-tion economy”, and “knowledge economy” [1]. All these terms imply a level of development in social and economic life such that the fol-lowing fundamental changes occur: networked coordination of economic entities becomes a priority, an orientation towards innovation takes place, human capital and information become primary competitive factors, and the tertiary sector prevails in the national economy [2]. In 2017, the Federal Program “Digital Economy” was developed in the Russian Fed-eration, representing one of the most promis-ing areas in the field of public administration [3]. Thus, it has become necessary to study the process of expanding the competence model of civil servants to ensure their suitability for the challenges of the digital environment, as well as the degree of readiness of civil servants for digital transformation. It is important to revise
The aim of the study is to develop guidelines for improving offi cial regulations of state civil servants in terms of qualifi cation requirements for competencies in the fi eld of information and communication technologies (ICT). The use of comparative analysis methods in the study of the content of offi cial regulations of state civil service in various subjects of the Russian Federation, as well as an expert survey on the content and current level of development of ICT competencies of civil servants allowed the authors to identify “basic,” “advanced” and “special” components in the structure of competencies. We also propose methodological recommendations for the transformation of ICT competences into digital components that provide an expanded set of knowledge and skills required for the digitalization of the civil service. These changes will allow the HR services of public authorities of the Russian Federation to provide a unifi ed approach to the formation of requirements for the maturity level of digital competencies of applicants seeking positions in the state civil service. It also will help to implement a targeted approach in the formation of programs for the development of personnel potential, taking into account the requirements of digital literacy.
Key words: digital competencies; digital economy; civil servant; public administration; position regulations; meta-competence; soft skills.
Citation: Vasilieva E.V., Pulyaeva V.N., Yudina V.A. (2018) Digital competence development of state civil servants in the Russian Federation. Business Informatics, no. 4 (46), pp. 28–42. DOI: 10.17323/1998-0663.2018.4.28.42
the qualification requirements for potential and current holders of civil service positions in the Russian Federation in terms of infor-mation and communication (ICT) skills and knowledge, highlighting new requirements for employee knowledge and the skills necessary for working in a digital environment.
The purpose of this work is to identify the digital components of ICT competencies of civil servants for inclusion in the official work regulations and Handbook of Qualification Requirements for applicants to civil service positions. To achieve this goal, the following tasks were completed:
reviewed existing job descriptions and methodological tools;
identified in regulatory documents the pre-requisites for expanding the competence model of civil servants to ensure the effective func-tioning of government agencies in the innova-tion environment;
conducted a survey of specialists and man-agers of the civil service in order to identify
-
BUSINESS INFORMATICS No. 4(46) – 2018
30
INFORMATION SYSTEMS AND TECHNOLOGIES IN BUSINESS
current requirements for specialists of various position groupings;
developed guidelines for the inclusion of digital components in the provisions of the official regulations of civil servants.
1. Background on changes in the competence model
of state civil servants
There are many definitions of the concept of “competences”. Many experts and special-ists in personnel management [4–10] offer their own interpretations, but two approaches to understanding competences are consid-ered basic: the American and the European. According to the American approach, compe-tences are defined as a description of employee behavior such that an employee who possesses appropriate competences demonstrates the correct behavior and achieves the desired work results [11]. The European approach focuses on the ability of an employee to solve certain tasks in order to achieve results in accord-ance with organization-defined requirements and standards [12–14]. Thus, the American approach involves the use of behavioral indi-cators to assess employees and generally relies on the postulates of behaviorism, while the European approach is more functional, since it focuses on the solving of specific professional problems.
Until recently, Russian academics dealt with issues of competence primarily from a ped-agogical science standpoint that considered the acquisition of knowledge and skills. Thus, according to I.A. Zimnyaya [15], social-pro-fessional competence is the cumulative inte-gral personal characteristic of a person who has received qualification and is characterized by a certain level of professionalism. Competence, as a professional characteristic of an individ-ual, is based on his or her personal qualities, intelligence and experience. Social-profes-sional competence is broken down into four blocks, two of which are basic: the intellectual
abilities and personal qualities that exist in a person before entering vocational training that serve as the foundation for the development of the personal competences inherent to a given profession [15].
Recently, domestic practice has seen an inte-gration of existing approaches to the under-standing of professional competences due to the activity of a large number of transnational companies using foreign management tech-niques, as well as the transition of vocational education away from the assessment of knowl-edge, abilities and skills and towards the assess-ment of competences [16].
The need to apply a competence approach to the state civil service system was first mentioned in Presidential Decree No. 601 of 7 May 2012, “On the main paths for improving the public administration system”. This decree called for the creation of a list of qualification require-ments for filling civil service posts based on the competence approach and taking into account the specific duties and functions, as well as the requirements of certain professional groups [17]. Subsequently, the Presidential Decree No. 403 of 11 August 2016 “On the main direc-tions of development of the civil service of the Russian Federation for 2016–2018” outlined the main pathways for the development of the civil service, including the creation of a single information and communication space in the civil service through a unified information sys-tem for the management of civil service per-sonnel, electronic personnel workflow, and the creation of a single, specialized information resource for the continued professional devel-opment of civil servants.
In accordance with the above legal/regula-tory actions, the Ministry of Labor and Social Affairs of the Russian Federation has devel-oped the following advisory documents:
Methodology for a comprehensive assess-ment of the professional performance of a pub-lic civil servant [18] (hereinafter referred to as the Methodology);
-
BUSINESS INFORMATICS No. 4(46) – 2018
31
INFORMATION SYSTEMS AND TECHNOLOGIES IN BUSINESS
Methodological toolkit for establishing qualification requirements for filling civil ser-vice positions (version 3.2) [19] (hereinafter referred to as Methodological Toolkit);
According to the Methodological Toolkit, “competence is a complex of professional and personal qualities manifested in the behavior of a civil servant, indicating the presence of the knowledge and skills necessary for the effec-tive and efficient fulfillment of official duties” [19]. A professional quality is understood as a characteristic demonstrated in the behavior of a civil servant that reflects the unity of the aspi-rations, abilities, knowledge, skills and per-sonal qualities necessary for the effective and efficient execution of official duties.
1.1. Professional qualities of state civil servants: Russian practice
Today there are three groups of professional qualities of a civil servant: general, applied and managerial (Table 1).
At the same time, as can be seen from the table, the professional qualities (competences) of civil servants identified by the Ministry of Labor do not take into account modern trends in the digitization of the economy and society (the only applied professional quality in this area is a more general “collection and analy-sis of information”). This gap in the legisla-tion, as well as other factors described above, made it necessary to undertake a study to form a model of digital competences of civil servants and identify the degree of readiness of govern-ment bodies to function in the innovation envi-ronment.
1.2. Official regulations for state civil servants: Russian practice
In accordance with Clause 7, Article 12 of the Federal Law “On the public civil service of the Russian Federation”, the knowledge and skill qualification requirements necessary for the performance of official duties depend on the
General professional qualities
Applied professional qualities
Managerial professional qualities
Results oriented Collection and analysis of information Planning activities and resources
Strengthening the authority of civil servants High-quality preparation of documents Task setting and organization of activities
Interpersonal understanding and communication style
Focus on ensuring the protection of legitimate interests of citizens Monitoring and evaluation of performance
Creative approach and innovation Motivating and developing subordinates
Persuasive communication* Making managerial decisions
Working on a team* Strategic vision
Professional self-development (hereinafter self-development) Managing changes
Transferring knowledge and experience* Public speaking and external communications*
Table 1. Professional qualities of state civil servants
* soft skills [20]
-
BUSINESS INFORMATICS No. 4(46) – 2018
32
INFORMATION SYSTEMS AND TECHNOLOGIES IN BUSINESS
field and nature of service performed by a civil servant. The official civil servant regulations (hereinafter referred to as the official regula-tions) may also provide qualification require-ments for a given specialization and a training regime for filling a civil service position.
Article 47 of this law states that the profes-sional service activity