The probability that a person visits a location. ~ means that two locations are equivalent. Grids -> close to the border of two different.

PGT: Measuring Mobility

Relationship using Personal, Global

and Temporal Factors

Shahab Helmi

BD Lab Seminar Series Spring 2015

Authors:

H. Wang

Z. Li

W. C. Lee

Conference: ICDM

Year: 2014

Paper info

IntroductionGoal:

Measuring the mobility relationship between two mobile users based on their interactions in the real world.

Problems?

Measuring the relationship strength.

Applications

Urban Planning

Recommendations

Advertisement targeting

Privacy protection

Anomaly detection

Measuring the meeting frequency is no enough! Meetings should not be treated equally!

i. Modeling:

personal background for each user.

global background for each location.

ii. Mobility relationship Mining:

Determining the weights of the meeting events.

Then the strength of the mobility relationship .

Approach

Personal Factor:

The probability that a person visits a location.

If two persons’ offices are located in the Times Square , they don not necessarily know each other, although they are co-located most of the times. On the other hand, if two persons have only visited Times Squares 3 times but at the same time, it is more likely that they know each other.

Global Factor:

Popularity of a location (can be estimated among all users in the dataset).

If two persons are co-located in a stadium, it is less likely that they know each other comparing when they co-locate in a private house.

Temporal Factor:

The time gap between consecutive meeting events.

Co-locating in 2 different days vs. 2 times in a same day!

Approach (cont’d)

Approach (cont’d)

Meeting frequency.

More recently: considering the global factor of a location using the entropy to measure its popularity.

What’s new?

Personal factor and temporal factor have been considered in few works (and usually not together).

Related Works

Problem Definition

Given a location dataset of n users, the location history of a user i is represented as a sequence of location and time stamps:

Given any two users i and j, the goal is to calculate a relationship measure Fij based on their mobility data -> meeting events.

Meeting Events

A meeting event is formed when two users being spatially close at the same time.

Problem Definition

The probability that a person visits a location.

~ means that two locations are equivalent.

Grids -> close to the border of two different cells.

Threshold -> 31m and 30m!

To solve above problems, a density function could be used:

Cd is a parameter that determines the impact of the distance.

d<1km walking distance

5k>d driving distance

Modeling Personal Mobility Background

Background Modeling Relationship Mining

Popularity of a location.

The set of location records of user i visiting location lock:

The ratio that user i’s visit o the entire population:

Shannon entropy of a location can be estimated using the probability vector of all users visiting this location:

Lower entropy means that this place is visited by few users.

Modeling Global Mobility Background


We aim to determine the significance of a meeting event ek between users i and j at location lock.

Personal factor weight (more weights for the places that users rarely visit):

Strength of a meeting (average weight X numbers):

60% of friends and 5% on non-friends have weights higher than 18.

**CDF: cumulative distribution function

Personal Factor


The problem of location factor is that it should be calculated from location history of all users.

Higher entropy -> lower weight

Combining global and local factors:

Global Factor


If two events are temporally closer -> lower weight for the meeting!

Temporal Factor


Finally!


?