Mining user similarity bas

ed on routine activities
Published in:Journal
Information Sciences: an International Journalarchive
Volume 236, July, 2013
Pages 17-32 Elsevier Science Inc.New York, NY, USA

Report:

Professor:Hsiao-Ping Tsai

kdd912@nchu

Outline
Introduction
System overview

Preliminary
Architecture
Routine activity mining
Reference places extraction
Routine activities mining

User similarity calculation

Reference place similarity calculation
Routine activity similarity calculation

Experiments
Conclusions
kdd912@nchu

Introduction
Mobile user similarity is significant for location-based social network servic

es. With the pervasiveness of location-acquisition technologies, research on

measuring mobile user similarity based on their trajectories has attracted a l
ot of attention.
In this paper, we address the problem of mining users long-term activity si
milarity based on their trajectories.
It propose a two-stage approach. At the rst stage, the notion of routine acti
vity is proposed to capture userslong-term activity regularities. The routine
activities of a user are extracted from his/her daily trajectories. At the secon
d stage, user similarity is calculated hierarchically based on the extracted ro
utine activities.

kdd912@nchu

 The daily routine activities of a student Tom can be summarized in Fig. 1. S

ince routine activity reflects both the temporal and the spatial regularities of
peoples daily lives, we take it as the basis to measure the long-term similari
ty between two users.

kdd912@nchu

System overview
2.1. Preliminary
Our system for user similarity estimation is based on routine activity extracted fr
om raw GPS data. First, we clarify some concepts and their data representation, i
ncluding GPS point, GPS trajectory, visit point, reference place, 1-day activity an
d routine activity.

kdd912@nchu

kdd912@nchu

2.2. Architecture

kdd912@nchu

Routine activity mining

3.1Reference places extraction

kdd912@nchu

kdd912@nchu

3.2Routine activities mining

kdd912@nchu

10

An ideal clustering result should maximize the intra-cluster similarity (i.e.

the average similarity between pairs of 1-day activities in the same cluster)
as well as minimize the inter-cluster similarity (i.e. the average similarity
between pairs of 1-day activities of different clusters), thus we used a
variant of the Dunn index

(3)

kdd912@nchu

11

User similarity calculation

4.1Reference place similarity calculation

(4)

(5)
(6
)
kdd912@nchu

12

4.2 Routine activity similarity calculation

(7)
Two routine activities A 1 and A 2 , and their corresponding reference place sets PS1 and PS2 , their
similarity can be calculated based on Eq. (8), where P 1i PS1, P2k PS2 , OMS is the Optimal
Matching Sequence of A1 and A2 , min(A1ij , A 2 kj ) is the common probability of
reference places i and k within the jth time span, T is the number of time spans.

(8)

kdd912@nchu

13

Experiments

kdd912@nchu

14

kdd912@nchu

15

Conclusions
In this paper, we propose an approach to measure user similarity for LBSNs

based on GPS trajectory mining. The most important novelty of our user si
milarity measure approach is that it can capture the similarity of users longterm activity regularities. To achieve this goal, we propose a framework to e
xtract the routine activities from users daily GPS trajectories,and calculate t
he similarity score between users based on their routine activities.
While our approach exploits the GPS trajectories, it is important to extend t
he routine activity mining framework to make it compatible with other indo
or locating infrastructure, and apply our approach to both indoor and outdo
or environments. Another important issue is to mine user similarity with spa
rse and incomplete trajectory data. We consider these as prom-ising future
works.

kdd912@nchu

16