Ongoing work on context inference, place detection. Presented at the International Conference and Exhibition of Ubiquitous Positioning, Indoor Positioning and Location Based Services. Corpus Christi, Texas. Nov 2014.
1. Semantic Labeling of Places
based on Phone Usage Features
using Supervised Learning
A. Rivero-Rodriguez, H. Leppäkoski ,R. Piché
1
19.11.2014
Tampere University of Technology
Tampere, Finland
www.tut.fi/posgroup
November 21, 2014
Corpus Christi, Texas, USA
UPIN-LBS
Context inference and awareness
2. This talk describes the design of the algorithms
for a smartphone to learn your significant
places
Training data Features Classifiers
2
19.11.2014
3. MDC dataset
Idiap and NRC-Lausanne
Lausanne Data Collection Campaign (2009-2011)
Records of 200 users over 18 months
Captures all types of information
Users provide extra information (labels!)
Anonymisation
46 GB of data!
3
19.11.2014
Training Data
4. Active Phone Usage
calls, messages
calendar, contacts
application usage
Pasive Phone Usage
network information
system Information
location & movement
4
19.11.2014
Features Available
Training Data
5. 5
19.11.2014
The places were identified by
clustering, then labeled by the user Training Data
200 m
Friend’s Home
Restaurant
Work
Home
6. 6
19.11.2014
We selected 14 features that could be
used by a place-labelling application
Call logs
callsTimeRatio
callsPerHour
Accelerometer
idleStillRatio
walkRatio
vehicleRatio
sportRatio
Features
System
duration
startHour
endHour
nightStay
batteryAvg
chargingTimeRatio
sysActiveRatio
sysActStartsPerHour
8. 8
19.11.2014
visits_20min.csv
places.csv
Definitions
for DB queries Make queries
system
call logs
accel activity
start times,
end times,
used ids,
place labels
Accumulate times & counts,
weight averages
feature vectors
for places
for each
user & place
Compute times,
counts, averages
for each
visit
Compute ratios Compute ratios
feature vectors
for visits
Features
We preprocessed the data to obtain
the features for both approaches
9. 9
19.11.2014
We applied five popular classification
methods to the data Classifiers
ܲ X | ܣ, ܤ =
ܲ ܣ| ܺ ܲ B|ܺ ܲ ܺ )
ܲ ܣ ܲ(ܤ)
Naïve Bayes (NB)
Decision Tree (DT)
K-nearest neighbors (K-NN)
Bagged Tree (DT)
Neural Networks (NN)
10. 10
19.11.2014
10000
9000
8000
7000
6000
5000
4000
3000
2000
1000
0
70%
O
W
H
28%
O
H
W
65%
W
H
O
82%
O
W
H
80%
O
H
W
12%
W
H
O
80%
O
W
H
89%
O
H
W
7%
W
H
O
96%
W O
H
29%
O
H
W
2%
W
H
O
93%
W O
H
25%
O
H
W
7%
W
H
O
Number of cases(visits)
Well Classified Misclassified
NB
53%
DT
75%
BT
77%
NN
61%
KNN
58%
H: Home
W: Work
O: Others
Results - Visits approach
Classifiers
11. 11
19.11.2014
40
35
30
25
20
15
10
5
0
97%
O
H
88%
O
H
W
69%
W
H
O
86%
O
H
91%
O
H
W
67%
W
H
O
97%
O
H
91%
O
H
W
69%
W
H
O
93%
O
H
85%
O
H
W
69%
W
H
O
86%
O
H
79%
O
H
W
53%
W
H
O
Number of cases(visits)
Well Classified Misclassified
NN
71%
DT
81%
NB
84%
BT
85%
KNN
71%
H: Home
W: Work
O: Others
Results - Places approach
Classifiers
12. Naive Bayes and Bagged Decision Tree with Places data-representation
are best
NN and K-NN underperform and are computationally demanding
Most 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 Rivero
alejandro.rivero@tut.fi