a location predictor based on dependencies between ...clu/lbsn2010/slides/nishino_lbsn2010.pdf ·...
TRANSCRIPT
A Location Predictor based on Dependencies Between Multiple Lifelog Data
Masaaki NishinoTakashi Yagi
Graduate School of Natural Science and Technology, Okayama University
Yukihiro Nakamura Shinyo Muto
NTT Cyber Solutions Laboratories, NTT Corporation
Masanobu Abe
1
ACM LBSN’10, November 2, 2010, San Jose, CA, USA
Copyright(c) 2010 NTT, All rights reserved.
Location Prediction
• Time limited sale infoat the supermarketsnear the predictedlocations
• Weather infoof future locations
•To-Do Tasks relatedto future locations
2
By predicting future locations, we can provide useful information to a user.
To-do tasksrelated to the
place
Time limitedSales
Copyright(c) 2010 NTT, All rights reserved.
Related Works
Extract and use regularities in movement
from location log・Markov Model[Ashbrook, 2003]・Dynamic Bayesian Network[Liao, 2007]・Sequential Pattern Mining[Monreale, 2009]
accumulatedlocation logs
predictionmodels
3
Home
Station
Office
0.6
0.4
1.0
1.0
Copyright(c) 2010 NTT, All rights reserved.
Problems
Since previous research uses regularities,
they cannot predict irregular movements.
Regulargo to school on weekdays
go to the gym every Monday
etc...
Irregularirregular meeting
business trips
etc...
4Copyright(c) 2010 NTT, All rights reserved.
Our proposal
Integrate different kinds of lifelogs to predict
both regular and irregular movements
Regulargo to school on weekdays
go to the gym every Monday.
etc...
Irregularirregular meeting
business trips
can predict with location log.
can predict with integrating different kinds of logs.
5Copyright(c) 2010 NTT, All rights reserved.
Calendar data
• People enter info about irregular events into it
• Widely used.
We use calendar data as a source for
making predictions of irregular movements
6Copyright(c) 2010 NTT, All rights reserved.
Dynamic Bayesian Networks modelfor integrating different data.
GPS Data
Calendar Data
DBN model can make reasonable predictions when prediction with only calendar/GPS data is difficult
7
Integrated DBN Model
Office OfficeBldg
A
“Lunch” “Bob”
Location
Stay Duration
Calendar
Time of Day Evening Evening Evening
10 min 0 min 15 min
Copyright(c) 2010 NTT, All rights reserved.
Preprocesses
8
Stayed places
(Office)
(Station A)
ClusteringGPS Data Stayed Places
Calendar Data Extracting Keywords Keywords
(37.53,-122.08)(37.50,-122.00)
(Office)(Station A)
“Meeting with Bob”“See the Dentist”
“Meeting” , “Bob”“Dentist”
“Meeting” “Bob”
“Dentist”
Copyright(c) 2010 NTT, All rights reserved.
Concepts of DBN Model
Place-place relationship
Home OfficeBuilding
A
Markov Model
Simple model for predicting locations.Can predict regular movements.
9
Location
Copyright(c) 2010 NTT, All rights reserved.
Concepts of DBN Model
Place-Keyword relationship
Building A
“Meeting” “Bob”
・ Can predict irregular movements
Infer a user’s own relationshipbetween place and keywordfrom co-occurrences.
10
Location
Calendar
Copyright(c) 2010 NTT, All rights reserved.
DBN model (basic)
Home OfficeBuilding
A
“Lunch” “Bob”
Integrate place-place relationship
and place-keyword relationship
11
Location
Calendar
Both regular and irregular movementscan be predicted.
Copyright(c) 2010 NTT, All rights reserved.
DBN model (actual model)
Office OfficeBldg
A
“Lunch” “Bob”
Make some extensions to basic model.
- add node that represents time of day, stay duration
Location
Stay Duration
Calendar
Time of Day Evening Evening Evening
10 min 0 min 15 min
12Copyright(c) 2010 NTT, All rights reserved.
Learning
• Estimating the parameters of the probability distributions of DBN from data.
• Using maximum a posteriori (MAP) estimation.
13Copyright(c) 2010 NTT, All rights reserved.
Inference
14
Use the Viterbi algorithm to infer a state sequencethat maximizes probability.
InitialState
Calendar
Time of Day
10:00 10:10 10:20 10:30 10:40 10:50
Location
Stay Duration
Copyright(c) 2010 NTT, All rights reserved.
Experimental Settings
Whether prediction accuracy is improved or not by
using calendar data.
Baseline:
DBN model without Calendar data
Dataset
GPS and calendar data of two subject
(about 50 days)
15Copyright(c) 2010 NTT, All rights reserved.
Evaluation metrics
• Evaluate the accuracy of prediction by changing the time difference between the subject of prediction and the time to start prediction
16
Time(ago)
Building A
Station A
Station B
12:00(0)
11:00(60)
10:00(120)
Subject ofthe prediction
Time to StartPrediction
Home
Copyright(c) 2010 NTT, All rights reserved.
Prediction ResultsRegular Movements
There are no much differences.17
0.3
0.4
0.5
0.6
0.7
050100150200
GPSGPS+Calendar
Blue lineRed line
0.5
0.55
0.6
0.65
0.7
0.75
0.8
050100150200
Subject A Subject B
Copyright(c) 2010 NTT, All rights reserved.
Prediction Results
18
0.1
0.2
0.3
0.4
0.5
0.6
050100150200
Irregular Movements
Subject A Subject B
The accuracy was improvedfor irregular movements
GPSGPS+Calendar
Blue lineRed line
0.1
0.15
0.2
0.25
0.3
0.35
050100150200
Copyright(c) 2010 NTT, All rights reserved.
Example of Predictions
The prediction failed since the wrong place is estimated from the calendar entry
19Copyright(c) 2010 NTT, All rights reserved.
Example of Predictions
The result was modified because the time needed for movement was considered
20Copyright(c) 2010 NTT, All rights reserved.
Conclusion
• We show a DBN model for making prediction for both regular and irregular movements by using GPS and calendar data.
– The accuracy of predictions for irregular movementwas improved.
– Wrong prediction due to wrong schedule can be modified by using GPS data.
• Future works
– Use other kinds of logs.
21Copyright(c) 2010 NTT, All rights reserved.
Thank you.
22Copyright(c) 2010 NTT, All rights reserved.