View
47
Download
1
Category
Preview:
Citation preview
Semantic Labeling of Placesbased on Phone Usage Features
using Supervised LearningA. Rivero-Rodriguez, H. Leppäkoski ,R. Piché
1
19.11.2014
Tampere University of TechnologyTampere, Finlandwww.tut.fi/posgroup
November 21, 2014Corpus Christi, Texas, USAUPIN-LBSContext inference and awareness
This talk describes the design of the algorithmsfor a smartphone to learn your significantplaces
Training data Features Classifiers
2
19.11.2014
MDC dataset
Idiap and NRC-LausanneLausanne Data Collection Campaign (2009-2011)Records of 200 users over 18 monthsCaptures all types of informationUsers provide extra information (labels!)Anonymisation46 GB of data!
3
19.11.2014
Training Data
Active Phone Usagecalls, messagescalendar, contactsapplication usage
Pasive Phone Usagenetwork informationsystem Informationlocation & movement
4
19.11.2014
Features AvailableTraining Data
5
19.11.2014
The places were identified byclustering, then labeled by the userTraining Data
200 m
Friend’s Home
Restaurant
Work
Home
6
19.11.2014
Call logscallsTimeRatiocallsPerHour
AccelerometeridleStillRatiowalkRatiovehicleRatiosportRatio
SystemdurationstartHourendHournightStaybatteryAvgchargingTimeRatiosysActiveRatiosysActStartsPerHour
Features
We selected 14 features that could beused by a place-labelling application
7
19.11.2014
Features
We considered two different datarepresentations
8
19.11.2014
visits_20min.csvplaces.csv
Definitionsfor DB queries Make queries
system
call logs
accel activity
start times,end times,
used ids,place labels
feature vectorsfor places
Accumulate times & counts,weight averages
for eachuser & place
Compute times,counts, averages
for eachvisit
Compute ratios Compute ratios
feature vectorsfor visits
Features
We preprocessed the data to obtainthe features for both approaches
9
19.11.2014
X | , = | B| )
( )
We applied five popular classificationmethods to the dataClassifiers
Naïve Bayes (NB)
Decision Tree (DT)
Bagged Tree (DT)
Neural Networks (NN)
K-nearest neighbors (K-NN)
10
19.11.2014
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
H
W
O
70%
W
H
O
28%
O
HW
65%
H
W
O
82%
W
H
O80%
O
H
W12%
H
W
O80%
W
HO
89%
O
H
W
7%
H
WO
96%
W
H
O29%
O
H
W2%
H
WO
93%
W
H
O25%
O
H
W7%
Num
bero
fcas
es(v
isits
)Well Classified Misclassified
NB53%
DT75%
BT77%
NN61%
KNN58%
H: HomeW: WorkO: Others
Results - Visits approachClassifiers
11
19.11.2014
0
5
10
15
20
25
30
35
40
H
O97%
W
HO
88%
O
H
W
69%
H
O
86%
W
HO
91%
O
H
W
67%
H
O97%
W
HO
91%
O
H
W
69%
H
O
93%
W
H
O85%
O
H
W
69%
H
O
86%
W
H
O
79%
O
H
W
53%
Num
bero
fcas
es(v
isits
)Well Classified Misclassified
NN71%
DT81%
NB84%
BT85%
KNN71%
H: HomeW: WorkO: Others
Results - Places approachClassifiers
Naive Bayes and Bagged Decision Tree with Places data-representation are bestNN and K-NN underperform and are computationally demandingMost relevant features are: night stay, stay duration, start time,battery status, idle time
Other classifiers (logistic regresion, support vector machine)Combine Places and Visits data-representations
12
19.11.2014
Classifiers
Results & Future Work
Alejandro Riveroalejandro.rivero@tut.fi
Recommended