1. Mining User Similarity from Semantic Trajectories
Josh Jia-Ching Ying, Eric Hsueh-Chan Lu, Wang-Chien Lee, Tz-Chiao Weng, and Vincent S. Tseng
Presenter: Josh Jia-Ching Ying

2. Outline Introduction Semantic Trajectory Based Friend Recommendation
Experiment
Conclusion and future work

3. Introduction
With the rapid growth and fierce competition in the market of social networking services, many service providers have deployed various recommendation services
friend recommendation
Most of the friend recommendation engines use profiles or on- line behavior of users to make recommendations instead of capturing the ‘’real’’ characteristics in user behavior
In recent years, a new breed of social networking services, called location-based social networks (LBSNs), have emerged

4. Introduction
Obviously user similarity plays a crucial role in these friend recommendation services
Most studies of measuring mobile users’ similarity focus only on analyzing geographic features of user trajectories
The notion of semantic trajectory has been proposed by Alvares et al. in 2007
sequence of locations with semantic tags to capture the landmarks passed by
eg. School  Park  Restaurant

5. Introduction

6. Introduction

7. Semantic Trajectory Based Friend Recommendation
Framework
Trajectory logs
Geographic
Information
Semantic Trajectory
Semantic Trajectory Transformation
Semantic Trajectory Pattern Mining
Pattern Sets
User Similarity Measurement
User Similarity
Matrix
Potential Friends Recommender
Smart-phones or PDAs
Laptops or PCs 1 3 4
input 2

8. Semantic Trajectory Transformation
GPS trajectory
Cell trajectory (Eagle et al. )
- 8 -

9. Semantic Trajectory Transformation
For GPS trajectory (basically follow Alvares et al’s approach)
US Post Office  Seniore’s Pizza  Fremont Park
Post Office  Restaurant  Park

10. Semantic Trajectory Transformation
For Cell trajectory
If the stay time > time threshold, the cell is called a stay cell.
stay time = leave time – arrive time

11. Semantic Trajectory Transformation
<Stay Cell0, Stay Cell1, Stay Cell2, Stay Cell3>
<{Unknown}, {School, Park }, {Park}, {Hospital}>

12. Framework Trajectory logs Geographic Information Semantic Trajectory
Semantic Trajectory Transformation
Semantic Trajectory Pattern Mining
Pattern Sets
User Similarity Measurement
User Similarity
Matrix
Potential Friends Recommender
Smart-phones or PDAs
Laptops or PCs 1 3 4
input 2

13. User Similarity Measurement
Maximal Semantic Trajectory Pattern Similarity (MSTP-Similarity)
Similarity between two Semantic Trajectory Pattern Sets
<A,{BC}> …
<A,{BC},E> …
<D,{AC},E> …
<D,{AC},E> …
- 13 -

14. MSTP-Similarity Common part
Given two Maximal Semantic Trajectory Patterns, we argue that they are more similar when they have more common parts
the longest common sequence (LCS) of the two patterns
- 14 -

15. MSTP-Similarity
The participation ratio of the common part to a pattern
- 15 -

16. MSTP-Similarity Pattern similarity Equal Average Weighted Average
- 16 -

17. Similarity between two Users
To measure how similar two pattern sets are:
Equal weight
Weighting by support
Weighting by TFIDF
user V
user U P1 … Pm
P1’ …
Pn’
There are m×n Maximal Semantic Trajectory Pattern Similarity
- 17 -

18. Weighting by support A pattern with a high support is more important
Geometric mean
Arithmetic average
user V
user U P1 … Pm
P1’ …
Pn’
- 18 -

19. TFIDF TFIDF=TF*log(IDF) Term frequency Inverse document frequency
<A,{BC}>: 6
<A,D> : 2
User a
TFIDF=TF*log(IDF)
TF: term frequency
IDF: inverse document frequency
Term frequency
User a
<A,{BC}>: 6/(6+2) = 3/4
<A,D> :2/(6+2) = 1/4
User b
<A,{BC}>: 3/(3+6) = 1/3
<B,E> :6/(3+6) = 2/3
User c
<A,{BC}>: 3/(3+2) = 3/5
<A,D> :2/(3+2) = 2/5
Inverse document frequency
<A,{BC}>: 3/(1+1+1) = 1
<A,D> : 3/(1+0+1) = 3/2
<B,E> : 3/(0+1+0) = 3
<A,{BC}>: 3
<B,E> : 6
User b
<A,{BC}>: 3
<A,D> : 2
User c
TFIDF in User a:
<A,{BC}>: (3/4)*log1 = 0
<A,D>: (1/4)*log1.5 = 0.04
- 19 -

20. Using TFIDF as the weight
Geometric mean
Arithmetic average
user V
user U P1 … Pm
P1’ …
Pn’
- 20 -

21. Experiment — dataset MIT reality mining dataset
The Reality Mining project was conducted from at the MIT Media Laboratory
Cell trajectory
Cell annotation

22. Experiment — ground truth
MIT Media Laboratory has conducted an online survey, which was completed by 94 mobile users
The survey data present the summarized behavior of a mobile user
Among the 94 mobile users,
7 users who do not have cell trajectory logs,
10 users who do not have cell annotation logs.
remaining 77 mobile users are used in our experiments

23. Experiment
Baseline
We directly perform a maximal sequential pattern mining algorithm on the stay cell sequence set for each mobile user

24. Conclusion and future work
We propose a novel framework to support friend recommendation services
the semantic trajectories of mobile users
MSTP-Similarity
for measuring the similarity between two semantic trajectory patterns
Through a series of experiments, our proposed friend recommendation framework has excellent performance under various conditions
Future work
Consider stay time (Duration) for the recommender
Consider geographic features for the recommender

25. Thank you for your attention
Question?