Anda di halaman 1dari 37

Outline

Introduction
Practical Issues
Proposed Solution
Results

Meaningful Locations from Cell Tower Information


Data

Werner Creixell1 Tomas Arredondo1 Sebastian Contreras1


1 Departamento de Electronica

Universidad Tecnica Federico Santa Mara

ALIO INFORMS
Buenos Aires, June 2010

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Outline

1 Introduction

2 Practical Issues

3 Proposed Solution

4 Results

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Introduction

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Introduction

Today more than 60 % of world population are carrying


cellphones.
Modern smartphones have the computing power of a 1995
desktop computer.
They are equipped with an increasing set of sensors and
communication means: GPS, accelerometers, Wi-Fi,
Bluetooth, etc.
The mobile phone has become the ubiquitous computer.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Introduction

Today more than 60 % of world population are carrying


cellphones.
Modern smartphones have the computing power of a 1995
desktop computer.
They are equipped with an increasing set of sensors and
communication means: GPS, accelerometers, Wi-Fi,
Bluetooth, etc.
The mobile phone has become the ubiquitous computer.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Introduction

Today more than 60 % of world population are carrying


cellphones.
Modern smartphones have the computing power of a 1995
desktop computer.
They are equipped with an increasing set of sensors and
communication means: GPS, accelerometers, Wi-Fi,
Bluetooth, etc.
The mobile phone has become the ubiquitous computer.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Introduction

Today more than 60 % of world population are carrying


cellphones.
Modern smartphones have the computing power of a 1995
desktop computer.
They are equipped with an increasing set of sensors and
communication means: GPS, accelerometers, Wi-Fi,
Bluetooth, etc.
The mobile phone has become the ubiquitous computer.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Introduction

In this work we explore the problem of inferring meaningful


locations using cellphone context data.
Meaningful locations such us home and work.
The context is date, time, cell towers id and Bluetooth
encounters.
Real data is used to test our approach.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Introduction

In this work we explore the problem of inferring meaningful


locations using cellphone context data.
Meaningful locations such us home and work.
The context is date, time, cell towers id and Bluetooth
encounters.
Real data is used to test our approach.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Introduction

In this work we explore the problem of inferring meaningful


locations using cellphone context data.
Meaningful locations such us home and work.
The context is date, time, cell towers id and Bluetooth
encounters.
Real data is used to test our approach.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Introduction

In this work we explore the problem of inferring meaningful


locations using cellphone context data.
Meaningful locations such us home and work.
The context is date, time, cell towers id and Bluetooth
encounters.
Real data is used to test our approach.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Data Source

MITs Reality Mining1 Project


N. Eagle and A. Pentland (2006), Reality Mining: Sensing
Complex Social Systems, Personal and Ubiquitous
Computing, Vol 10, #4, 255-268.
The cellphone usage of a hundred users (students and faculty
members) was recorded from July 2004 to May 2005.
The records include: Bluetooth encounters, cell towers and
phone calls.

1
http://reality.media.mit.edu/
Werner Creixell et al., UTFSM ALIO INFORMS
Outline
Introduction
Practical Issues
Proposed Solution
Results

Data Source

MITs Reality Mining1 Project


N. Eagle and A. Pentland (2006), Reality Mining: Sensing
Complex Social Systems, Personal and Ubiquitous
Computing, Vol 10, #4, 255-268.
The cellphone usage of a hundred users (students and faculty
members) was recorded from July 2004 to May 2005.
The records include: Bluetooth encounters, cell towers and
phone calls.

1
http://reality.media.mit.edu/
Werner Creixell et al., UTFSM ALIO INFORMS
Outline
Introduction
Practical Issues
Proposed Solution
Results

Data Source

MITs Reality Mining1 Project


N. Eagle and A. Pentland (2006), Reality Mining: Sensing
Complex Social Systems, Personal and Ubiquitous
Computing, Vol 10, #4, 255-268.
The cellphone usage of a hundred users (students and faculty
members) was recorded from July 2004 to May 2005.
The records include: Bluetooth encounters, cell towers and
phone calls.

1
http://reality.media.mit.edu/
Werner Creixell et al., UTFSM ALIO INFORMS
Outline
Introduction
Practical Issues
Proposed Solution
Results

Practical Issues

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Practical Issues

In the Reality Mining Project, the users register the cell tower
antenna with a label for work, home and Else.
But cell towers are dynamically assigned to the cell phone
users. They can change if the signal is weak or due to
congestion in the connected tower.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Practical Issues

In the Reality Mining Project, the users register the cell tower
antenna with a label for work, home and Else.
But cell towers are dynamically assigned to the cell phone
users. They can change if the signal is weak or due to
congestion in the connected tower.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Practical Issues

One location can have a set


of possible antennas.
A change in the connected
antenna does not necessarily
means a change in the user
position.
In practice the antenna
records oscillate between
a pair of antennas in a short
period of time.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Practical Issues

One location can have a set


of possible antennas.
A change in the connected
antenna does not necessarily
means a change in the user
position.
In practice the antenna
records oscillate between
a pair of antennas in a short
period of time.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Practical Issues

One location can have a set


of possible antennas.
A change in the connected
antenna does not necessarily
means a change in the user
position.
In practice the antenna
records oscillate between
a pair of antennas in a short
period of time.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Practical Issues

One location can have a set


of possible antennas.
A change in the connected
antenna does not necessarily
means a change in the user
position.
In practice the antenna
records oscillate between
a pair of antennas in a short
period of time.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Proposed Solution

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Labeling Cell Towers Antennas

We identify the oscillations between a pair of antenna by


defining a time threshold.
We use the oscillations to associate unlabeled antennas to the
labeled ones.
Use bluetooth data to associate unlabeled antennas to some
location.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Labeling Cell Towers Antennas

We identify the oscillations between a pair of antenna by


defining a time threshold.
We use the oscillations to associate unlabeled antennas to the
labeled ones.
Use bluetooth data to associate unlabeled antennas to some
location.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Labeling Cell Towers Antennas

We identify the oscillations between a pair of antenna by


defining a time threshold.
We use the oscillations to associate unlabeled antennas to the
labeled ones.
Use bluetooth data to associate unlabeled antennas to some
location.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Classification

After oscillations filtering and labeling we have a set of


labeled events.
We use this set to train and test a binary Support Vector
Machine(SVM) classifier.
The input vector to the SVM classifier was defined as follows
(Wday , H0 , . . . , H23 , T1 , . . . , TJ , B1 , . . . , BK )
Where Wday , Hi , Tj , Bk {0, 1} 1 j J and 1 k K .

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Classification

After oscillations filtering and labeling we have a set of


labeled events.
We use this set to train and test a binary Support Vector
Machine(SVM) classifier.
The input vector to the SVM classifier was defined as follows
(Wday , H0 , . . . , H23 , T1 , . . . , TJ , B1 , . . . , BK )
Where Wday , Hi , Tj , Bk {0, 1} 1 j J and 1 k K .

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Classification

After oscillations filtering and labeling we have a set of


labeled events.
We use this set to train and test a binary Support Vector
Machine(SVM) classifier.
The input vector to the SVM classifier was defined as follows
(Wday , H0 , . . . , H23 , T1 , . . . , TJ , B1 , . . . , BK )
Where Wday , Hi , Tj , Bk {0, 1} 1 j J and 1 k K .

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Results

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Subjects to Show Results

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Results for the Home Classifier

Subject 7 Subject 95
Month Hits Total Accuracy Hits Total Accuracy
Jul 04 0 0 0 33 34 97.05 %
Ago 04 81 113 71.68 % 0 0 0
Sep 04 283 394 71.82 % 330 330 100 %
Oct 04 285 402 70.89 % 251 251 100 %
Nov 04 301 309 97.41 % 254 254 100 %
Dec 04 317 318 99.68 % 252 258 97.67 %
Jan 05 Training Training
Feb 05 244 248 98.38 % 142 143 99.30 %
Mar 05 328 335 97.38 % 306 308 99.35 %
Apr 05 288 291 98.96 % 177 178 99.43 %
May 05 0 0 0 52 53 98.11 %
Jul 04 - May 05 2127 2410 88.25 % 1797 1809 99.33 %

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Results for the Work Classifier

Usuario 7 Usuario 95
Month Hits Total Accuracy Hits Total Accuracy
Jul 04 0 0 0 33 34 97.05 %
Ago 04 112 113 99.11 % 0 0 0
Sep 04 391 394 99.23 % 326 330 98.78 %
Oct 04 391 402 97.26 % 243 251 96.81 %
Nov 04 303 309 98.05 % 250 254 98.42 %
Dec 04 314 318 98.74 % 253 258 98.06 %
Jan 05 Training Training
Feb 05 247 248 99.59 % 143 143 100 %
Mar 05 330 335 98.05 % 301 308 97.72 %
Apr 05 286 291 98.28 % 173 178 97.19 %
May 05 0 0 0 51 53 96.22 %
Jul 04 - May 05 2374 2410 98.50 % 1773 1809 98.01 %

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Conclusions

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Conclusions

In this work we have showed a method to infer the user


location, using context phone data.
The SVM classifier showed high accuracy in the examples
showed.
The worst performance was for subject 7 on the summer
months for the home location.
Given the appropriate context information more sophisticated
inferencing can be done.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Conclusions

In this work we have showed a method to infer the user


location, using context phone data.
The SVM classifier showed high accuracy in the examples
showed.
The worst performance was for subject 7 on the summer
months for the home location.
Given the appropriate context information more sophisticated
inferencing can be done.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Conclusions

In this work we have showed a method to infer the user


location, using context phone data.
The SVM classifier showed high accuracy in the examples
showed.
The worst performance was for subject 7 on the summer
months for the home location.
Given the appropriate context information more sophisticated
inferencing can be done.

Werner Creixell et al., UTFSM ALIO INFORMS


Outline
Introduction
Practical Issues
Proposed Solution
Results

Conclusions

In this work we have showed a method to infer the user


location, using context phone data.
The SVM classifier showed high accuracy in the examples
showed.
The worst performance was for subject 7 on the summer
months for the home location.
Given the appropriate context information more sophisticated
inferencing can be done.

Werner Creixell et al., UTFSM ALIO INFORMS

Anda mungkin juga menyukai