Combining Linguistic Values and Semantics to Represent User Preferences

Post on 22-Feb-2016

47 views 0 download

description

Combining Linguistic Values and Semantics to Represent User Preferences. Valentin Grouès , Yannick Naudet , Odej Kao. Need for Semantics. Semantic ambiguity: User: u=(Indonesia=0.7;Java=0.9;island=0.2) Items: d1=(Java=0.4;hotel=0.8), d2=(Java=0.4;software=0.8). island. - PowerPoint PPT Presentation

Transcript of Combining Linguistic Values and Semantics to Represent User Preferences

Combining Linguistic Values and Semantics to Represent User Preferences

Valentin Grouès, Yannick Naudet, Odej Kao

Need for Semantics• Semantic ambiguity:

User: u=(Indonesia=0.7;Java=0.9;island=0.2)Items: d1=(Java=0.4;hotel=0.8), d2=(Java=0.4;software=0.8)

pref(d1,u)=pref(d2,u)=0.19

Distinction between the two concepts is essential for not producing undesirable recommendations

island programming language

Iván Cantador. "Exploiting the Conceptual Space in Hybrid Recommender Systems: a Semantic-based Approach". 2008, Madrid

2

• Assumption of terms independance:

User: u=(Indonesia=0.7;Java=0.9;island=0.2)Items: d1=(Java=0.4;hotel=0.8), d2 =(Java=0.4;archipelago=0.8)

• Assumption of terms independance:

User: u=(Indonesia=0.7;Java=0.9;island=0.2)Items: d1=(Java=0.4;hotel=0.8), d2 =(Java=0.4;archipelago=0.8)

Need for Semantics

pref(d1,u)=pref(d2,u)=0.19

Iván Cantador. "Exploiting the Conceptual Space in Hybrid Recommender Systems: a Semantic-based Approach". 2008, Madrid

island island

Semantic relations between concepts have to be considered

3

Friend Of A Friend

• A user model widely adopted by the Semantic Web community

• Personal profiles, activities and relationships• Large websites and software support (Livejournal, TypePad,

Foaf-o-Matic)• Existing datasets (foafPub contains already more than 200

000 triples)

4

eFoaf

• Cover demographic and basic user information

• Context aware (e.g. not only one contact address)

• Simple and complex interests associated with a context of validity

• Open to external RDF datasets

• Skills, abilities and handicaps

5

Weighted Interests Ontology

• URI: http://purl.org/ontology/wi/core#• Authors: Dan Brickley, Libby Miller, Toby Inkster et al• Description: ‘‘The Weighted Interests Vocabulary specification

provides basic concepts and properties for describing describing preferences (interests) within contexts, their temporal dynamics and their origin on/ for the Semantic Web’’

ex:JohnDoe a foaf:Person ; foaf:name "John Doe" ; wi:preference [ a wi:WeightedInterest ; wi:topic dbpedia:The_Terminator ; wo:weight [ a wo:Weight ; wo:weight_value 0.5 ; wo:scale ex:aScale ; ]; wi:interest_dynamics ex:atHome ];

6

Fuzzy Sets

• To represent imprecise information inherent to the human way of thinking

• Humans have a tendency to use imprecise concepts for claiming tastes: “cheap restaurant”, “long movie”, “young actor”, etc.

• Limitations of crisp systems:• For a user willing to find a restaurant with a cost up to 20€ the

system will equally discard a restaurant costing 21€ as a restaurant costing 300€.

a user would prefer having an answer proportional to the distance between his ideal preference and the recommended content

7

Common membership functions

• Trapezoidal (e.g. “moderate temperatures”)• Triangular (e.g. “close to”)• Left shoulder (e.g. “cheap”)• Right Shoulder (e.g. “expensive”)

8

(x) 1

kernelsupport

(x) 0

Integrating fuzzy sets within ontologies

• FuSOR: A model for representing fuzzy sets and linguistic values within ontologies (Y. Naudet, V. Grouès, M. Foulonneau, Introduction to Fuzzy-Ontological Context-Aware Recommendations in Mobile Environments, APRESW 2010)

9

FuSor: Characteristics of the approach

• Can be used as an extension of an ontology without requiring any modifications, OWL DL compliant

• Allows using fuzzy sets and their membership functions for any datatype property

• Supports context and domain dependency

10

Yannick Naudet, Valentin Groues, Muriel Foulonneau, Introduction to Fuzzy-Ontological Context-Aware Recommendations in Mobile Environments, APRESW 2010, Heraklion, Greece

Membership functions can be used to define the way a user interest deviates from an “ideal” value.

Ex: “I am looking for a restaurant with prices up to 20€ but I could accept up to 25€ even if I would be less satisfied”.

11

Ex: Describing interest boundaries

Combining eFoaf with Fuzzy Sets

ex:JohnDoe a foaf:Person ;foaf:name "John Doe" ; wi:preference [ a wi:WeightedInterest ; wi:topic [

a ex:Restaurant ; ex:fuzzyCost ex:john_Cheap; ];

];

12

Combining eFoaf with Fuzzy Sets

ex:Cost fusor:hasFuzzyVersion ex:fuzzyCost; ; ex:john_Cheap a fusor:LinguisticValue [

fusor:hasSupport [ a fusor:Range; fusor:hasLowBoundary –INF;

fusor:hasHighBoundary 25; ];

fusor:hasKernel [ a fusor:Range; fusor:hasLowBoundary –INF; fusor:hasHighBoundary 20;

];];

13

Application to knowledge-based recommender systems

14

• : aggregation function to compute the recommendation score of an item regarding the user preferences

• : an item having characteristics • : the set of fuzzy sets representing the preferences of the user

for each respective characteristic of the items• : the membership degree of the characteristic of an item to the

fuzzy set

Application to knowledge-based recommender systems

15

• Intuitive heuristics for :

1. 2. (

If an item has a higher membership degree than an other item for each of their characteristics then should get a higher recommendation score

If there are no characteristics of the item having a membership value higher than the corresponding one of and at least one characteristic of having a membership value higher than the corresponding one of then should get a higher recommendation score

If two items and have the same average of their characteristics membership values, then the item having the highest minimum membership value should get a higher recommendation

If the average of the membership values of an item is much higher than the average of an other item, the first one should get a higher recommendation score

Example

16

• A user looking for a restaurant with moderate prices and close to his position

Conclusions and perspectives

17

• Propositions:• eFoaf: representation of weighted interests, user relationships, abilities,

etc.• A method to use linguistic values to describe user interests• A list of intuitive heuristics to determine an aggregation method

• Future work:• Evaluations of the added value of using linguistic values to describe user

interests, empirical comparison of different aggregation functions• Integration with semantic similarity measures• Semantic implicit profiling

Thank you for your attention

18

Any questions ?