MINING OBJECT MOVEMENT PATTERNS
FROM TRAJECTORY DATA
Phan Nhat Hai
4th October, 2013
Supervisors
Dr. Dino Ienco, Pr. Pascal Poncelet, Dr. Maguelonne Teisseire
BACKGROUND AND MOTIVATION
 Nowadays, many electronic devices are used for real
world applications
 GPS, sensor networks, mobile phone, …
 « interesting » patterns for:
 movement pattern analysis, animal behavior, route
planning and vehicle control, location prediction, …
2
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
SOME EXAMPLES
3
Route Planning
Animal migration analysis
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
The world’s largest traffic jam
in history (China)
SPATIO-TEMPORAL DATA (ST)
 Represented as a list of points, located in space and
time
 T=(x1,y1, t1), …, (xn, yn, tn)  position in space at time ti
was (xi, yi)
4
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
MINING SPATIO-TEMPORAL PATTERNS FROM TRAJECTORY
DATA (1)
 Frequent Patterns:
 Frequent followed paths:
 Group pattern [6], Tralus [7], …
5
Region (Cluster)
[6] Y. Wang et. al. Data Knowl.
Eng., June 2006.
[7] J. G. Lee et. al. In ACM
SIGMOD ’07.
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
o1
o2
o3 o4
[1] Z.Li et. al. PVLDB 10.
[2] P. Kalnis et. al. SSTD’05.
[3] J. Gudmundsson et. al. ACM GIS’06.
[4] H. Jeung et. al. VLDB 08.
[5] F. Verhein. SDM’09.
MINING SPATIO-TEMPORAL PATTERNS FROM TRAJECTORY
DATA (2)
 Clustering:
 Group together similar trajectories
 For each group produce a summary
 Flock [3], convoy [4], moving cluster [2], swarm & closed
swarm [1], k-Star [5]
6
Region (Cluster)
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
o1
o2
o3
o4
SWARM – CLOSED SWARM [1]
 Swarm - groups of objects (O, T ):
 At least objects move together
 timestamps
 Closed Swarm
 Swarm which cannot be enlarged
 Algorithm
 ObjectGrowth
7
[1] Z.Li et. al. Swarm: mining
relaxed temporal moving object
clusters. PVLDB 2010.
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
CONVOY [4]
 Convoy - groups of objects (O, T ):
 At least objects move together
 consecutive timestamps
 Algorithm
 CuTS*
8
[4] H. Jeung et. al. Discovery of
convoys in trajectory
databases. PVLDB 2008.
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
MOTIVATIONS (1)
 Motivations:
 Complexity?
 Are they enough?
 Informative patterns?
9
data
Informative
patterns
extract
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
MOTIVATIONS (2)
 Proposed solution
data
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
Unifying
10
OUTLINE
 Background and Motivations
 Unifying Framework
 Gradual Trajectory
 Mining Representative Movement Patterns
 Conclusions and Perspectives
11
CLUSTER MATRIX
 Objects: transactions
 Clusters: items
12
diaper
beer
diaperbeer
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
FREQUENT CLOSED ITEMSET FROM CLUSTER
MATRIX
13
Frequent Itemset
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
THE MAIN INTUITION (FOLLOWING…)
 We are now able to extract itemsets corresponding to
a set of clusters occurring over time
 Not movement patterns yet!
 What about properties on Itemsets?
14
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
SWARM
15
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
PROPERTIES
 In the same way it is possible to define properties for:
Swarm, Closed Swarm, Convoy, Moving Cluster,
Periodic Pattern, …
 We are now able to extract different movement
patterns!
16
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
THE MAIN PROCESS (GET_MOVE)
17
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
INCREMENTAL GET_MOVE
18
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
THE MAIN PROCESS
19
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
CMC CuTS* ObjectGrowth Vg-Growth Incremental
GeT_Move
Convoys X X X
Closed Swarms X X
Group Patterns X X
Moving Cluster X
EXPERIMENTAL RESULTS
 Datasets:
 Competitive algorithms:
#objects #timestamps
Swainsoni 43 4,425
Buffalo 165 3,000
Synthetic* 500 10,000
Synthetic 2 50,000 10,000
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
* http://iapg.jade-hs.de/personen/brinkhoff/generator/
20
SWAINSONI
21
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
UNIFYING FRAMEWORK – CONCLUSIONS
 GeT_Move: a unifying movement pattern mining
approach
 Properties adapted to specific movement patterns
 Proofs of properties
 Theorem providing that all the patterns are found
 Incremental GeT_Move
 A new approach for identifying the size of blocks
 Fully nested block partition
22
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
OUTLINE
 Background and Motivations
 Unifying Framework
 Gradual Trajectory
 Mining Representative Movement Patterns
 Conclusions and Perspectives
23
ONE OF CLOSED SWARMS …
24
o1
o2
o3
o4
o6
o5
c1 c2
c3
c4
c5
t1 t2 t3 t4 t5 t6
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
…GRADUAL TRAJECTORIES
25
o1
o2
o3
o4
o6
o5
c1 c2
c3
c4
c5
t1 t2 t3 t4 t5 t6
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
A CONCRETE EXAMPLE
26
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
PATTERN DEFINITION
 The objects still remain in the next cluster
 The number of objects is equal-increasing (resp. equal-
decreasing)
 At least a number of certain timestamps
 non-consecutive
27
- Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
27
TIME RELAXED GRADUAL TRAJECTORIES
 Timestamps can be:
 non-consecutive
 within a sliding time window
28
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
o1
o2
o3
o4
c1
c2
t1 t2 t3 ………… t999 t1000
A
F
Sliding
window
Too far away
EXPERIMENTAL RESULTS
 Synthetic data: 500 objects - 10,000 timestamps
 Reasonable scalability
 Low complexity
29
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
GRADUAL TRAJECTORY - CONCLUSIONS
 New kinds of trajectories: gradual trajectory
 ClusterGrowth: an efficient algorithm to extract all
gradual trajectories
 Fuzzy closed swarm
 Too many extracted patterns:
 DiCompoGP algorithm to directly extract the top-k gradual
trajectories 30
Convergent Divergent
OUTLINE
 Background and Motivations
 Unifying Framework
 Gradual Trajectory
 Mining Representative Movement Patterns
 Conclusions and Perspectives
31
 Methodology
 A set of movement patterns (closed swarms, convoys,
gradual trajectories, etc.)
 Employ MDL (Minimum Description Length) schema to
select the most informative and less redundant pattern
set
 Compo Algorithm
 Rank and select the most representative patterns
 Allow different types of pattern in the final results
 Characterize data by the selected patterns
32
CONTRIBUTIONS
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
MOTIVATIONS
33
data
Patterns
1) One kind of patterns is not enough
to describe the data!
2) Overlapping!
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
PROBLEM STATEMENT
 Given a spatio-temporal DB Odb and a set of patterns F
(extracted from Odb)
 Discover the optimal dictionary P (subset of F)
 compresses the data best w.r.t. the given encoding schema
 L(p): number of bits to encode the pattern p + extra bit
to encode the type of pattern
 L(Odb|P): number of bits to encode the dataset Odb
given P
34
MDL approach: LP(Odb) = L(P) + L(Odb|P)
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
ENCODING EXAMPLE (I)
35
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
ENCODING EXAMPLE (II)
36
L(ODB|P) = 4 + 6 + 2 + 1 + 1 = 14
L(P) = 4 + 5 + 3 + 4 = 16
LP(ODB) = 30
L(ODB|P) = 4 + 5 + 2 + 1 + 1 = 13
L(P) = 4 + 5 = 9
LP(ODB) = 22
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
NAÏVE COMPO VS SMART COMPO
We design two different approaches:
 Naive Compo (baseline)
 Work in a greedy way
 Given the actual P, for each candidate p’ recompress the data
with P U p’
 Select the p’ that obtain the best performance
 Smart Compo
 Compute the gain incrementally
 Avoid to recompress the whole data
 Directly compute Gain(p’,P) = L(Odb|P) - L(Odb|P U p’) without
compute L(Odb|P U p’)
37
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
EXPERIMENTAL RESULTS
38
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
REPRESENTATIVE PATTERN - CONCLUSIONS
 Propose an encoding scheme allowing multi-
overlapping movement patterns
 Propose two algorithms
 Naïve Compo (greedy approach)
 Smart Compo (compute gain incrementally)
 Experimental results show that the top-k
representative patterns are well adapted to the data
39
OUTLINE
 Background and Motivations
 Unifying Framework
 Gradual Trajectory
 Mining Representative Movement Patterns
 Conclusions and Perspectives
40
OVERALL CONCLUSIONS (1)
 Three step framework
 GeT_Move: a unifying movement pattern mining approach
 Discovering novel patterns: Gradual trajectory + Fuzzy
closed swarm
 Mining representative movement patterns
41
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
data
OVERALL CONCLUSIONS (2)
DEMONSTRATION SYSTEM
 Link: http://www.lirmm.fr/~phan/multimove.jsp
42
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
OVERALL CONCLUSIONS (3) – OTHER APPLICATIONS
 Mining trajectories on genes
 Mining trajectories on tweets
43
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
PERSPECTIVES (1)
 Streaming GeT_Move
 Mining representative movement patterns from streaming
trajectory data
44
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
PERSPECTIVES (2)
 Trajectory mining on remote sensing, spatial
information on satellite image processing
45
EXTRA WORK
 Mining multi-relational gradual patterns (with Prof. Donato
Malerba)
 Kendal’s tau
 Gradual support
 Communication graph summarization (with Dr. Francesco
Bonchi)
46
-Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
PUBLICATIONS
[1] P. N. Hai, D. Ienco, P. Poncelet, M. Teisseire. "Mining Representative Movement Patterns
through Compression". PAKDD 2013.
[2] A.Z.E. Aabidine, A. Sallaberry, S. Bringay, M. Fabregue, C. Lecellier, P. N. Hai, P. Poncelet.
“Co2Vis: A Visual Analytics Tool for Mining Co-expressed And Co-regulated Genes Implied
in HIV Infections”. IEEE BioVis 2013.
[3] P. N. Hai, D. Ienco, P. Poncelet, M. Teisseire. "Mining Fuzzy Moving Object Clusters".
ADMA 2012.
[4] P. N. Hai, D. Ienco, P. Poncelet, M. Teisseire. "Mining Time Relaxed Gradual Moving Object
Clusters". ACM GIS 2012.
[5] F. Bouillot, P. N. Hai, N. Béchet, S. Bringay, D. Ienco, S. Matwin, P. Poncelet, M. Roche,
and M. Teisseire. "How to Extract Relevant Knowledge from Tweets?". ISIP 2012.
[6] P. N. Hai, P. Poncelet, M. Teisseire. "GET_MOVE: An Efficient and Unifying Spatio-
Temporal Pattern Mining Algorithm for Moving Objects". IDA 2012.
[7] P. N. Hai, P. Poncelet, M. Teisseire. "An Efficient Spatio-Temporal Mining Approach to
Really Know Who Travels with Whom!". BDA 2012. (selected as Best papers)
[8] P. N. Hai, D. Ienco, P. Poncelet, M. Teisseire. "Extracting Trajectories through an Efficient
and Unifying Spatio-Temporal Patten Mining System". ECML-PKDD 2012.
[9] P. N. Hai, P. Poncelet, M. Teisseire. "MovingObjects: Combining Gradual Rules and Spatio-
Temporal Patterns". IEEE ICSDM 2011.
[10] P. N. Hai, P. Poncelet, M. Teisseire. "An Efficient Spatio-Temporal Mining Approach to
Really Know Who Travels with Whom!". ISI special issue, selected papers from BDA’12,
2013, to appear. 47

Mining Object Movement Patterns from Trajectory Data

  • 1.
    MINING OBJECT MOVEMENTPATTERNS FROM TRAJECTORY DATA Phan Nhat Hai 4th October, 2013 Supervisors Dr. Dino Ienco, Pr. Pascal Poncelet, Dr. Maguelonne Teisseire
  • 2.
    BACKGROUND AND MOTIVATION Nowadays, many electronic devices are used for real world applications  GPS, sensor networks, mobile phone, …  « interesting » patterns for:  movement pattern analysis, animal behavior, route planning and vehicle control, location prediction, … 2 - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 3.
    SOME EXAMPLES 3 Route Planning Animalmigration analysis -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives The world’s largest traffic jam in history (China)
  • 4.
    SPATIO-TEMPORAL DATA (ST) Represented as a list of points, located in space and time  T=(x1,y1, t1), …, (xn, yn, tn)  position in space at time ti was (xi, yi) 4 - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 5.
    MINING SPATIO-TEMPORAL PATTERNSFROM TRAJECTORY DATA (1)  Frequent Patterns:  Frequent followed paths:  Group pattern [6], Tralus [7], … 5 Region (Cluster) [6] Y. Wang et. al. Data Knowl. Eng., June 2006. [7] J. G. Lee et. al. In ACM SIGMOD ’07. - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives o1 o2 o3 o4
  • 6.
    [1] Z.Li et.al. PVLDB 10. [2] P. Kalnis et. al. SSTD’05. [3] J. Gudmundsson et. al. ACM GIS’06. [4] H. Jeung et. al. VLDB 08. [5] F. Verhein. SDM’09. MINING SPATIO-TEMPORAL PATTERNS FROM TRAJECTORY DATA (2)  Clustering:  Group together similar trajectories  For each group produce a summary  Flock [3], convoy [4], moving cluster [2], swarm & closed swarm [1], k-Star [5] 6 Region (Cluster) - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives o1 o2 o3 o4
  • 7.
    SWARM – CLOSEDSWARM [1]  Swarm - groups of objects (O, T ):  At least objects move together  timestamps  Closed Swarm  Swarm which cannot be enlarged  Algorithm  ObjectGrowth 7 [1] Z.Li et. al. Swarm: mining relaxed temporal moving object clusters. PVLDB 2010. - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 8.
    CONVOY [4]  Convoy- groups of objects (O, T ):  At least objects move together  consecutive timestamps  Algorithm  CuTS* 8 [4] H. Jeung et. al. Discovery of convoys in trajectory databases. PVLDB 2008. - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 9.
    MOTIVATIONS (1)  Motivations: Complexity?  Are they enough?  Informative patterns? 9 data Informative patterns extract - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 10.
    MOTIVATIONS (2)  Proposedsolution data - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives Unifying 10
  • 11.
    OUTLINE  Background andMotivations  Unifying Framework  Gradual Trajectory  Mining Representative Movement Patterns  Conclusions and Perspectives 11
  • 12.
    CLUSTER MATRIX  Objects:transactions  Clusters: items 12 diaper beer diaperbeer - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 13.
    FREQUENT CLOSED ITEMSETFROM CLUSTER MATRIX 13 Frequent Itemset - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 14.
    THE MAIN INTUITION(FOLLOWING…)  We are now able to extract itemsets corresponding to a set of clusters occurring over time  Not movement patterns yet!  What about properties on Itemsets? 14 - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 15.
    SWARM 15 - Background andMotivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 16.
    PROPERTIES  In thesame way it is possible to define properties for: Swarm, Closed Swarm, Convoy, Moving Cluster, Periodic Pattern, …  We are now able to extract different movement patterns! 16 - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 17.
    THE MAIN PROCESS(GET_MOVE) 17 - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 18.
    INCREMENTAL GET_MOVE 18 - Backgroundand Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 19.
    THE MAIN PROCESS 19 -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 20.
    CMC CuTS* ObjectGrowthVg-Growth Incremental GeT_Move Convoys X X X Closed Swarms X X Group Patterns X X Moving Cluster X EXPERIMENTAL RESULTS  Datasets:  Competitive algorithms: #objects #timestamps Swainsoni 43 4,425 Buffalo 165 3,000 Synthetic* 500 10,000 Synthetic 2 50,000 10,000 - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives * http://iapg.jade-hs.de/personen/brinkhoff/generator/ 20
  • 21.
    SWAINSONI 21 - Background andMotivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 22.
    UNIFYING FRAMEWORK –CONCLUSIONS  GeT_Move: a unifying movement pattern mining approach  Properties adapted to specific movement patterns  Proofs of properties  Theorem providing that all the patterns are found  Incremental GeT_Move  A new approach for identifying the size of blocks  Fully nested block partition 22 - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 23.
    OUTLINE  Background andMotivations  Unifying Framework  Gradual Trajectory  Mining Representative Movement Patterns  Conclusions and Perspectives 23
  • 24.
    ONE OF CLOSEDSWARMS … 24 o1 o2 o3 o4 o6 o5 c1 c2 c3 c4 c5 t1 t2 t3 t4 t5 t6 - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 25.
    …GRADUAL TRAJECTORIES 25 o1 o2 o3 o4 o6 o5 c1 c2 c3 c4 c5 t1t2 t3 t4 t5 t6 -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 26.
    A CONCRETE EXAMPLE 26 -Backgroundand Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 27.
    PATTERN DEFINITION  Theobjects still remain in the next cluster  The number of objects is equal-increasing (resp. equal- decreasing)  At least a number of certain timestamps  non-consecutive 27 - Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives 27
  • 28.
    TIME RELAXED GRADUALTRAJECTORIES  Timestamps can be:  non-consecutive  within a sliding time window 28 -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives o1 o2 o3 o4 c1 c2 t1 t2 t3 ………… t999 t1000 A F Sliding window Too far away
  • 29.
    EXPERIMENTAL RESULTS  Syntheticdata: 500 objects - 10,000 timestamps  Reasonable scalability  Low complexity 29 -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 30.
    GRADUAL TRAJECTORY -CONCLUSIONS  New kinds of trajectories: gradual trajectory  ClusterGrowth: an efficient algorithm to extract all gradual trajectories  Fuzzy closed swarm  Too many extracted patterns:  DiCompoGP algorithm to directly extract the top-k gradual trajectories 30 Convergent Divergent
  • 31.
    OUTLINE  Background andMotivations  Unifying Framework  Gradual Trajectory  Mining Representative Movement Patterns  Conclusions and Perspectives 31
  • 32.
     Methodology  Aset of movement patterns (closed swarms, convoys, gradual trajectories, etc.)  Employ MDL (Minimum Description Length) schema to select the most informative and less redundant pattern set  Compo Algorithm  Rank and select the most representative patterns  Allow different types of pattern in the final results  Characterize data by the selected patterns 32 CONTRIBUTIONS -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 33.
    MOTIVATIONS 33 data Patterns 1) One kindof patterns is not enough to describe the data! 2) Overlapping! -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 34.
    PROBLEM STATEMENT  Givena spatio-temporal DB Odb and a set of patterns F (extracted from Odb)  Discover the optimal dictionary P (subset of F)  compresses the data best w.r.t. the given encoding schema  L(p): number of bits to encode the pattern p + extra bit to encode the type of pattern  L(Odb|P): number of bits to encode the dataset Odb given P 34 MDL approach: LP(Odb) = L(P) + L(Odb|P) -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 35.
    ENCODING EXAMPLE (I) 35 -Backgroundand Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 36.
    ENCODING EXAMPLE (II) 36 L(ODB|P)= 4 + 6 + 2 + 1 + 1 = 14 L(P) = 4 + 5 + 3 + 4 = 16 LP(ODB) = 30 L(ODB|P) = 4 + 5 + 2 + 1 + 1 = 13 L(P) = 4 + 5 = 9 LP(ODB) = 22 -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 37.
    NAÏVE COMPO VSSMART COMPO We design two different approaches:  Naive Compo (baseline)  Work in a greedy way  Given the actual P, for each candidate p’ recompress the data with P U p’  Select the p’ that obtain the best performance  Smart Compo  Compute the gain incrementally  Avoid to recompress the whole data  Directly compute Gain(p’,P) = L(Odb|P) - L(Odb|P U p’) without compute L(Odb|P U p’) 37 -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 38.
    EXPERIMENTAL RESULTS 38 -Background andMotivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 39.
    REPRESENTATIVE PATTERN -CONCLUSIONS  Propose an encoding scheme allowing multi- overlapping movement patterns  Propose two algorithms  Naïve Compo (greedy approach)  Smart Compo (compute gain incrementally)  Experimental results show that the top-k representative patterns are well adapted to the data 39
  • 40.
    OUTLINE  Background andMotivations  Unifying Framework  Gradual Trajectory  Mining Representative Movement Patterns  Conclusions and Perspectives 40
  • 41.
    OVERALL CONCLUSIONS (1) Three step framework  GeT_Move: a unifying movement pattern mining approach  Discovering novel patterns: Gradual trajectory + Fuzzy closed swarm  Mining representative movement patterns 41 -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives data
  • 42.
    OVERALL CONCLUSIONS (2) DEMONSTRATIONSYSTEM  Link: http://www.lirmm.fr/~phan/multimove.jsp 42 -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 43.
    OVERALL CONCLUSIONS (3)– OTHER APPLICATIONS  Mining trajectories on genes  Mining trajectories on tweets 43 -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 44.
    PERSPECTIVES (1)  StreamingGeT_Move  Mining representative movement patterns from streaming trajectory data 44 -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 45.
    PERSPECTIVES (2)  Trajectorymining on remote sensing, spatial information on satellite image processing 45
  • 46.
    EXTRA WORK  Miningmulti-relational gradual patterns (with Prof. Donato Malerba)  Kendal’s tau  Gradual support  Communication graph summarization (with Dr. Francesco Bonchi) 46 -Background and Motivations - Unifying Framework - Gradual Trajectory - Mining Representative Movement Patterns - Conclusions & Perspectives
  • 47.
    PUBLICATIONS [1] P. N.Hai, D. Ienco, P. Poncelet, M. Teisseire. "Mining Representative Movement Patterns through Compression". PAKDD 2013. [2] A.Z.E. Aabidine, A. Sallaberry, S. Bringay, M. Fabregue, C. Lecellier, P. N. Hai, P. Poncelet. “Co2Vis: A Visual Analytics Tool for Mining Co-expressed And Co-regulated Genes Implied in HIV Infections”. IEEE BioVis 2013. [3] P. N. Hai, D. Ienco, P. Poncelet, M. Teisseire. "Mining Fuzzy Moving Object Clusters". ADMA 2012. [4] P. N. Hai, D. Ienco, P. Poncelet, M. Teisseire. "Mining Time Relaxed Gradual Moving Object Clusters". ACM GIS 2012. [5] F. Bouillot, P. N. Hai, N. Béchet, S. Bringay, D. Ienco, S. Matwin, P. Poncelet, M. Roche, and M. Teisseire. "How to Extract Relevant Knowledge from Tweets?". ISIP 2012. [6] P. N. Hai, P. Poncelet, M. Teisseire. "GET_MOVE: An Efficient and Unifying Spatio- Temporal Pattern Mining Algorithm for Moving Objects". IDA 2012. [7] P. N. Hai, P. Poncelet, M. Teisseire. "An Efficient Spatio-Temporal Mining Approach to Really Know Who Travels with Whom!". BDA 2012. (selected as Best papers) [8] P. N. Hai, D. Ienco, P. Poncelet, M. Teisseire. "Extracting Trajectories through an Efficient and Unifying Spatio-Temporal Patten Mining System". ECML-PKDD 2012. [9] P. N. Hai, P. Poncelet, M. Teisseire. "MovingObjects: Combining Gradual Rules and Spatio- Temporal Patterns". IEEE ICSDM 2011. [10] P. N. Hai, P. Poncelet, M. Teisseire. "An Efficient Spatio-Temporal Mining Approach to Really Know Who Travels with Whom!". ISI special issue, selected papers from BDA’12, 2013, to appear. 47