Location Prediction Under Data Sparsity
-
Upload
james-mcinerney -
Category
Technology
-
view
323 -
download
1
description
Transcript of Location Prediction Under Data Sparsity
![Page 1: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/1.jpg)
Modelling heterogeneous location habits in human populations for location
prediction under data sparsity
James McInerney1, Jiangchuan Zheng2, Alex Rogers1, Nick Jennings1
1. University of Southampton, United Kingdom2. Hong Kong University of Science and Technology, China
UbiCompZurich, Switzerland
11th September 2013
![Page 2: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/2.jpg)
2
Applications of Mobility Models
![Page 3: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/3.jpg)
3
Applications of GroupMobility Models
Across domains:
– Exploration (e.g. visual summary of many individuals' behaviour; answer “what if?” by modifying parameters)
– Inference (e.g. find out who is similar to whom; step towards semantic labelling)
– Prediction (e.g. alleviate data sparsity in individual prediction; prediction conditioned on other people's locations)
![Page 4: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/4.jpg)
4
Existing Group Mobility Work
Limited by assuming:
– Knowledge of the social network e.g., De Domenico et al. (2012), Sadilek et al. (2012)
– Prior semantic labelling e.g., Eagle & Pentland (2009)
![Page 5: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/5.jpg)
5
Existing Group Mobility Work
Limited by assuming:
– Knowledge of the social network e.g., De Domenico et al. (2012), Sadilek et al. (2012)
– Prior semantic labelling e.g., Eagle & Pentland (2009)
Restrictive when considering large groups of possibly unconnected individuals
![Page 6: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/6.jpg)
6
Existing Group Mobility Work
Limited by assuming:
– Knowledge of the social network e.g., De Domenico et al. (2012), Sadilek et al. (2012)
– Prior semantic labelling e.g., Eagle & Pentland (2009)
Restrictive when considering large groups of possibly unconnected individuals = populations
![Page 7: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/7.jpg)
7
Our Individual Mobility Model
Idea: assign latent location habit to each observation
Location habit :=
Location +
Time of day +
Day of week
![Page 8: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/8.jpg)
8
Our Individual Mobility Model
Idea: assign latent location habit to each observation
Location habit :=
Location +
Time of day +
Day of week
tn » N (µkµkµk)
dn » M(¯k¯k¯k)
x n » N (ÁkÁkÁk)
![Page 9: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/9.jpg)
9
Periodic Location Behaviour
[Krumm & Brush, Pervasive (2011)]
![Page 10: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/10.jpg)
10
How Many Latent Habits?
Use a Dirichlet process
where Ni = ® for new habit
p(hn = ijh1::n¡1) /Ni
N + ®
![Page 11: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/11.jpg)
11
How Many Latent Habits?
Use a Dirichlet process
1 2 3
h1h3
h5
h2h4
Chinese restaurant process (CRP) interpretation:
where Ni = ® for new habit
p(hn = ijh1::n¡1) /Ni
N + ®
![Page 12: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/12.jpg)
12
How Many Latent Habits?
Use a Dirichlet process
(where ® = 1)
1 2 3
h1h3
h5
h2h4
Chinese restaurant process (CRP) interpretation:
where Ni = ® for new habit
p(hn = ijh1::n¡1) /Ni
N + ®
p(h6jh1 = 1; h2 = 2; h3 = 1; h4 = 3; h5 = 1) =
µ1
2
1
6
1
6
1
6
¶
![Page 13: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/13.jpg)
13
Generative ProcessFor each observation n: 1..N of a single individual:
1. Assign observation (e.g., h6 ) to a table (e.g., 3) using CRP
2. Draw spatial and temporal components of the observation
1 2 3
h1h3
h5
h2h4 h6
x 6 » N (Á3)
t6 » N (µ3)
d6 » M(¯3) Wednesday
2:31pm
(47:407877; 8:508089)
![Page 14: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/14.jpg)
14
Modelling Populations Requires Unified Model
– Strength in population exploration, inference, and prediction comes from shared parameters
– But too much sharing leads to inflexibility
![Page 15: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/15.jpg)
15
Modelling Populations Requires Unified Model
– Strength in population exploration, inference, and prediction comes from shared parameters
– But too much sharing leads to inflexibility
Implication: we want global pool of habits but individual mixture proportions and spatial parameters.
![Page 16: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/16.jpg)
16
Hierarchical Dirichlet Process Expresses Topic Heterogeneity
– Globally shared set of topics (e.g., [car, drive, wheel, road] and [film, movie, actor, star, blockbuster]) but each document expresses topics to different extents
Article 1 Article 2
![Page 17: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/17.jpg)
17
Hierarchical Dirichlet Process Expresses Topic Heterogeneity
– Globally shared set of topics (e.g., [car, drive, wheel, road] and [film, movie, actor, star, blockbuster]) but each document expresses topics to different extents
Article 1 Article 2
![Page 18: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/18.jpg)
18
Extension to Mobility Analysis
– Observing continuous locations and times instead of discrete words
– Shared locations is overly restrictive assumption
→ keep spatial parameters local to individuals
![Page 19: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/19.jpg)
19
LocHDP
Shaded nodes = known values
Square nodes = hyperparameters
Round nodes = random variables
![Page 20: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/20.jpg)
20
Simplifying Assumption of Habits
![Page 21: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/21.jpg)
21
Predictive Density
p(x¤jt¤; d¤;x1::Nx1::Nx1::N ; t1::N ; d1::Nd1::Nd1::N)
/Zp(x1:Nx1:Nx1:N ; t1:N ; d1:Nd1:Nd1:N jh1::Nh1::Nh1::N ; ´́́)p(´́́)dh1::Nh1::Nh1::Nd´́́
(Hyperparameters omitted)
![Page 22: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/22.jpg)
22
Predictive Density
p(x¤jt¤; d¤;x1::Nx1::Nx1::N ; t1::N ; d1::Nd1::Nd1::N)
/Zp(x1:Nx1:Nx1:N ; t1:N ; d1:Nd1:Nd1:N jh1::Nh1::Nh1::N ; ´́́)p(´́́)dh1::Nh1::Nh1::Nd´́́
(Hyperparameters omitted)
Intractable, so use Gibbs sampling on hierarchical version of CRP, the Chinese restaurant franchise [Teh et al. 2006]
![Page 23: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/23.jpg)
23
Nokia Dataset
– Nokia Lausanne dataset (38 individuals, 1 year)
![Page 24: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/24.jpg)
24
Applications
– Exploration (e.g. visual summary of many individuals' behaviour; answer “what if?” by modifying parameters)
– Inference (e.g. find out who is similar to whom; step towards semantic labelling)
– Prediction (e.g. alleviate data sparsity in individual prediction; prediction conditioned on other people's locations)
![Page 25: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/25.jpg)
25
Experiment 1: Can LocHDP Help Overcome Data Sparsity?
Methodology:
– Simulate arrival of new user by truncating their observation history (real data)
– Gradually introduce longer history (independent variable)
– Examine predictive performance (dependent variable)
![Page 26: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/26.jpg)
26
Experiment 1: Can LocHDP Help Overcome Data Sparsity?
(Error bars indicate 95% confidence range)
![Page 27: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/27.jpg)
27
Exploration Using LocHDPGPS (normalised) Day of Week Time of Day
Latitude
Longitude
Pr Pr
![Page 28: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/28.jpg)
28
Experiment 1: Can LocHDP Help Overcome Data Sparsity?
![Page 29: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/29.jpg)
29
Experiment 2: How Many Auxiliary Users Needed?
(Error bars indicate 95% confidence range)
![Page 30: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/30.jpg)
30
Applications
– Exploration (e.g. visual summary of many individuals' behaviour; answer “what if?” by modifying parameters)
– Inference (e.g. find out who is similar to whom; step towards semantic labelling)
– Prediction (e.g. alleviate data sparsity in individual prediction; prediction conditioned on other people's locations)
![Page 31: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/31.jpg)
31
Exploration Using LocHDP
![Page 32: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/32.jpg)
32
Conclusions
– LocHDP unified model for location behaviour in populations that retains individual predictive/descriptive power
– Applied to help overcome data sparsity for individual prediction in ubiquitous systems
– Factor 2.4 benefit with < 2 weeks data for new users in Nokia dataset
– Maximum benefit achieved with pool of 10 established users
![Page 33: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/33.jpg)
33
Future Work
– Test on many more users: Openpaths dataset (GPS), Orange Ivory Coast dataset (cell tower)
– Faster variational Bayes derivation for parameter inference
– Fuller exposition of capabilities of unified model, i.e., show what can be done further in exploration and inference
![Page 34: Location Prediction Under Data Sparsity](https://reader034.fdocuments.us/reader034/viewer/2022051817/5482f30c5806b5d9048b4735/html5/thumbnails/34.jpg)
34
Future Work
– Test on many more users: Openpaths dataset (GPS), Orange Ivory Coast dataset (cell tower)
– Faster variational Bayes derivation for parameter inference
– Fuller exposition of capabilities of unified model, i.e., show what can be done further in exploration and inference
Thank you – Poster P42 on Thursday at 12:40pm