Efficient Mining of Group Patterns from User Movement Data

1
Efficient Mining of Group Patterns from User
Movement Data

• A/P Lim Ee Peng
• School of Computer EngineeringNanyang
  Technological University

2
Acknowledgement

• PhD students
• Wang Yida just graduated
• Hady Lauw in 2nd year of PhD program
• Collaborator
• A/P San-Yih HwangDepartment of Information
  ManagementNational Sun Yat-Sen
  UniversityKaohsiung, Taiwan

3
Outline

• Introduction
• Algorithms AGP VG-growth
• Spherical Location Summarization
• Performance Evaluation
• Conclusion

4
Introduction

• People are affiliated to different kinds of
  groups family, friends, school/class, company,
  etc..
• Sociologists and management experts have studied
  groups and group dynamics extensively.
• Peer pressure and group conformity can affect the
  behaviors of individuals.
• Traditional applications
• Improvement of communication
• Roles of individuals in a group
• Other applications
• Group-oriented marketing strategy
• Security

5
Introduction

• Existing techniques for deriving groups
• Static attributes age, education, work place,
  income, etc..
• Transactional attributes purchase history
• Problems of existing techniques
• Collection of static and transactional data
• Groups derived are not necessary useful for some
  applications

6
Research Objectives

• Group pattern mining based on user movement data
• Physical proximity Group members are physically
  close to one another
• Temporal proximity Group members stay close for
  some meaningful time duration
• Assumptions
• User movement can be tracked by mobile phones,
  RFIDs and other devices.
• Positioning technology can be as accurate as 1-20
  meters (GPS/AGPS).
• Privacy issues can be addressed separately.

7
Example
8
Group Examples
9
Valid Segment of a User Group

• User movement database D (D1, D2, , DM)
• Di is a time series of locations of user ui, ltt,
  (x, y, z)gt
• uit.p (x, y, z)
• Distance function
• d((x1, y1, z1), (x2, y2, z2))
  ((x1-x2)2(y1-y2)2(z1-z2)2)1/2
• max_dis and min_dur thresholds
• max_dis Two users are close to each other if ?
  max_dist apart
• min_dur The closeness should last ? min_dur to
  become meaningful
• Given max_dis, and min_dur, ta, tb is a valid
  segment for a group of users G if
• ? ui , uj? G, ? t in ta,tb, d(uit.p, ujt.p)
  ? max_dis
• ? ui , uj? G, d(uita-1.p, ujta-1.p) gt
  max_dis
• ? ui , uj? G, d(uitb1.p, ujtb1.p) gt
  max_dis and
• (tb-ta1) ? min_dur

10
User Movement Database
11
Group Pattern

• Group pattern
• Given D, ltG, max_dis, min_durgt forms a group
  pattern if G has a valid segment
• A k-group pattern is a group pattern with k users
• Weight of group pattern
• PltG, max_dis, min_durgt be a group pattern with
  valid segments s1, s2, , sn, and
• N denotes the number of time points in D

12
Valid Group Pattern Mining

• Valid group pattern
• If weight(P) ? min_wei, P is called a valid group
  pattern, and G is called a valid group
• min_wei is a user specified weight threshold
• Valid Group (Pattern) Mining Problem
• Given D, max_dis, min_dur, and min_wei, find all
  the valid group patterns (or valid groups).

13
Valid Group Mining vs Association Rule Mining
(ARM)

• No explicit concept of transaction in a movement
  database
• For each time point, group those users who are
  nearby one another into one transaction gt Large
  number of transactions
• Weight defined for group pattern different from
  support in ARM
• New algorithms are required

14
Apriori-liked Group Pattern (AGP) Mining Algorithm

• Based on Apriori Algorithm
• Sub-group pattern
• PltG, max_dis, min_durgt is a sub-group pattern
  of PltG, max_dis, min_durgt if G ? G
• Apriori property for group patterns
• If a group pattern is valid, all of its sub-group
  patterns are valid as well
• For example u1, u2 are valid if u1, u2 are
  valid.

15
AGP Algorithm
16
AGP Algorithm
17
Example
max_dis 10, min_dur 3, and min_wei 50
s(u1,u2) Valid segments of u1,u2 0,3,
7,9 Weight(u1,u2) 4 3 / 10 0.7
18
Example
Set of all valid groups
19
Design of VG-growth Algorithm

• Weaknesses of AGP
• Huge sets of candidate k-groups
• Multiple scans on D
• FP-tree and FP-growth Han, Pei and Yin 2000
• Challenges
• The concept of transaction does not exist
• weight is more complex than support
• VG-growth algorithm
• Perform mining recursively until the largest
  group containing user ui is found

20
Valid Group graph (VG graph)

• A VG graph is a directed graph (V, E)
• V vertices representing users in the set of
  valid 2-groups
• E directed edges, each representing a valid
  2-group (ui,uj)
• Each edge is assigned valid segments of (ui,uj)
• Construction
• Apply AGP to find valid 2-groups
• Scan D once to derive valid segments for valid
  2-groups.

21
VG Graph Example
22
VG-growth

• Prefix-neighbor
• If (ugv ) is a directed edge in a VG-graph, u is
  called the prefix-neighbor of v
• VG-growth algorithm
• Traversing the VG-graph
• Examine all the prefix-neighbors for each vertex
• Recursively mining the conditional VG-graph
• VG-growth is much more efficient than AGP with
  respect to mining valid k-groups (kgt2)

23
VG-growth
24
Mining of u5
Output valid groups u1,u5, u2,u5, u4,u5
5,9
3,9
0,3,7,9
3,9
0,9
25
Visualization of Group Patterns
26
Experimental Results

• IBM City Simulator generated 3-D user movement
  over city layout
• 1000 users, 1000 time points (10 mins per point),
  1000 x 1500 x 100 m3 gt 12 MB
• Performance metrics
• Total execution time (T)
• Time for mining valid 2-groups (T2)
• Time for mining all other valid groups (Tk)
• Parameter settings
• Change min_wei from 1 to 10
• max_dist 30, min_dur 4

27
Results of AGP and VG-growth
28
Results of AGP and VG-growth
29
Results
30
Results
31
Common problem of AGP VG-growth

• A common step of AGP and VG-growth mining valid
  2-groups
• T2 time used for mining valid 2-groups
• Tk time used for mining valid k-groups (kgt2)
• T whole execution time (T T2Tk)
• Bottleneck
• Assume M users and N time points in D
• T2 dominates T, especially when M and N are large

32
Summarization Method for Mining Group Patterns

• Idea
• Number of candidate 2-groups
• Reduce the number of candidate 2-groups
• Spherical location summarization method
• Partition each users movement data into segments
  with equal duration
• Summarize the locations within each segment by a
  sphere
• D Summarized user movement database

33
Spherical Location Summarization

• Spherical location summarization
• Divide movement database of a user into time
  segments with fixed length, w.
• Summarize location points within each segment by
  a sphere.

Sphere(pc, r)
34
Spherical Location Summarization
35
Preprocessing the Sphere Data

• Observation
• If the minimum distance between two spheres S1
  and S2 gt dist, then all pairs of points from S1
  and S2 gt dist.
• Preprocessing step
• upper bound of max_dis
• Compute the upper bounds of weight count and
  valid segment length (longest close sphere
  segment length) for each candidate 2-group

max_dis
36
Upper Bound Weight Count of a User Pair
Set of CSSs of ui,uj
37
Upper Bound Weight Count of a User Pair
38
Preprocessing the Sphere Data
39
Spherical Location Summarization based Algorithm
for Mining Valid 2-Groups (SLSV2G)

• Prune unpromising candidate 2-groups using
  pre-computed upper bound information
• Check the closeness relationship for each
  candidate 2-group based on the summarized
  database
• Check the original database only when the
  closeness relationship cannot be determined

40
SLSV2G Algorithm
41
SLSV2G Algorithm
42
Closeness between Spheres

• Case 1 All points in 2 spheres ? max_dis
• Case 2 All points in 2 spheres gt max_dis

d(c1,c2) - (r1r2)
S2 (c2, r2)
S1 (c1, r1)
S1
S2
max_dis
d(c1,c2) - (r1r2) gt max_dis
43
Closeness between Spheres

• Case 3 Only some points in 2 spheres ? max_dis
  (need to check the original database)

44
Performance Evaluation

• Synthetic data generator
• IBM City Simulator
• Three datasets

45
Performance Evaluation

• Series-I SLSV2G versus VG-growth
• Focus on finding valid 2-groups
• max_dist 30, min_dur 4
• With location summarization vs. without
  summarization
• Compare T2 and C2

D is loaded in main memory
VG-growth
Without location summarization
D is on hard disk D is loaded in main memory
SLSV2G
With location summarization
46
Time for finding valid 2-groups (T2)

• T2 of SLSV2G is only 5 - 8 of T2 of VG-growth
• DBIII is too large to be mined without location
  summarization

47
Number of candidate 2-groups (C2)

• Without location summarization, C2 is a
  constant
• With location summarization, C2 is much smaller
  (1 7)

48
Performance Evaluation

• Series-II Scalability of SLSV2G
• Time window size, w
• Number of users, M
• Number of time points, N

49
Performance Evaluation

• Does not scale linearly with time window size, w
• There exists an optimal w for a given dataset
• Overhead of scanning summarized database vs
  granularity of summarization

50
Performance Evaluation

• SLSV2G is much more scalable than VG-growth wrt M
• The gap becomes larger when M increases

51
Performance Evaluation

• Both are linear with N
• SLSV2G is much more scalable

52
Conclusion

• Methods for valid group pattern mining AGP,
  VG-growth
• Location summarization method to efficiently
  mining valid 2-groups.
• Experiments show that summarization method reduce
  the processing time significantly.
• Other related work
• Other kinds of location summarization models
• Extension of valid group pattern mining to handle
  dynamic user movement database
• Other practical issues missing movement data,
  location accuracies

53
End of TalkQA?

• Ee-Peng Lim
• aseplim_at_ntu.edu.sg

54
Publications

• Jeng-Kuen Chiu, San-Yih Hwang, Ying-Han Liu,
  Ee-Peng Lim, Mining Mobile Group Patterns A
  Trajectory-based Approach, 9th Pacific-Asia
  Conference on Knowledge Discovery and Data Mining
  (PAKDD2005), Hanoi, Vietnam, May 2005.
• Hady Wirawan Lauw, Ee-Peng Lim, Mining Social
  Network from Spatio-Temporal Events, Workshop on
  Link Analysis, Counterterrorism and Security, in
  conjunction with SIAM Data Mining Conference,
  Newport Beach, April 2005.
• Yida Wang, Ee-Peng Lim, San-Yih Hwang, Efficient
  mining of group patterns from user movement data.
  To appear at Data and Knowledge Engineering,
  2005.
• Yida Wang, Ee-Peng Lim, San-Yih Hwang, Efficient
  mining of group patterns using data
  Summarization, International Conference on
  Database Systems for Advanced Applications
  (DASFAA2004), Jeju Island, 2004.
• Yida Wang, Ee-Peng Lim, San-Yih Hwang, On mining
  group patterns of mobile users. International
  Conference on Database and Expert Systems
  Applications (DEXA2003), Prague, 2003