Title: Spatio-temporal%20frequent%20pattern%20mining%20for%20public%20safety:%20Concepts%20and%20Techniques
1Spatio-temporal frequent pattern mining for
public safety Concepts and Techniques
- Pradeep Mohan
- Department of Computer Science
- University of Minnesota, Twin-Cities
- Advisor Prof. Shashi Shekhar
- Thesis Committee Prof. F. Harvey, Prof. G.
Karypis, Prof. J. Srivastava
Contact mohan_at_cs.umn.edu
2Biography
- Education
- Ph.D., Student, Department. of Computer Science
and Engineering., University of Minnesota, MN,
2007 Present. - B. E., Department. of Computer Science and
Engineering, Birla Institute of Technology,
Mesra, Ranchi, India. 2003-2007
- Major Projects during PhD
- US DoJ/NIJ- Mapping and analysis for Public
Safety - CrimeStat .NET Libaries 1.0 Modularization of
CrimeStat, a tool for the analysis of crime
incidents. - Performance tuning of Spatial analysis routines
in CrimeStat - CrimeStat 3.2a - 3.3 Addition of new modules for
spatial analysis.
- US DOD/ ERDC/ TEC Cascade models for multi
scale pattern discovery - Designed new interest measures and formulated
pattern mining algorithms for identifying
patterns from large crime report datasets.
1
3Thesis Related Publications
- Cascading spatio-temporal pattern discovery
(Chapter 2) - P. Mohan, S.Shekhar, J.A.Shine, J.P. Rogers.
Cascading spatio-temporal pattern discovery A
summary of results. In Proc. Of 10th SIAM
International Conference on Data Mining 2010 (SDM
2010, Full paper acceptance rate 20) - P. Mohan, S.Shekhar, J.A.Shine, J.P. Rogers.
Cascading spatio-temporal pattern discovery. IEEE
Transactions on Knowledge and Data Engineering
(TKDE). (Accepted Regular Paper, In Press 20
Acceptance Rate)
- Regional co-location pattern discovery (Chapter
3) - P.Mohan, S.Shekhar, J.A. Shine, J.P. Rogers,
Z.Jiang, N.Wayant. A spatial neighborhood graph
based approach to Regional co-location pattern
discovery summary of results. In Proc. Of 19th
ACM SIGSPATIAL International Conference on
Advances in GIS 2011 (ACM SIGSPATIAL 2011, Full
paper acceptance rate 23)
- Crime Pattern Analysis Application (Chapter 4)
- S.Shekhar, P. Mohan, D.Oliver, Z.Jiang, X.Zhou.
Crime pattern analysis A spatial frequent
pattern mining approach. M. Leitner (Ed.), Crime
modeling and mapping using Geospatial
Technologies, Springer (Accepted with Revisions).
2
4Other Publications
- Spatio-temporal data analysis
- X.Zhou, S.Shekhar, P. Mohan, S. Leiss, P. Snyder.
Discovering Interesting sub-paths in
spatiotemporal datasets. In Proc. Of 19th ACM
SIGSPATIAL International Conference on Advances
in GIS 2011 (ACM SIGSPATIAL 2011, Full paper
acceptance rate 23)
- Spatial data analysis
- P. Mohan, R. E. Wilson, S.Shekhar, B.George,
N.Levine, M.Celik Should SDBMS support a join
index? a case study from CrimeStat. In Proc. Of
16th ACM SIGSPATIAL International Conference on
Advances in GIS 2008 (ACM SIGSPATIAL 2008, Full
paper acceptance rate 19) - P. Mohan, X. Zhou, S.Shekhar. Quantifying
resolution sensitivity of spatial
autocorrelation A Resolution Correlogram
approach. In Proc. Of 7th International
Conference on Geographic Information Science,
2012 (GIScience 2012, Full paper) - S.Shekhar, M.R.Evans, J.M.Kang, P. Mohan.
Identifying patterns in spatial information A
survey of methods. (Accepted) WIREs Data Mining
and Knowledge Discovery, Wiley Interdisciplinary
Reviews, John Wiley and Sons, Inc, 2011 (in
press)
3
5Outline
4
6Motivation Public Safety
- Crime generators and attractors
- Identifying events (e.g. Bar closing, football
games) that lead to increased crime.
Question What / Where are the frequent crime
generators ?
- Identifying frequent crime hotspots
- Law enforcement planning
- Courtsey www.startribune.com
Predicting the next location of burglary.
Question Where are the crime hotspots ?
- Predicting crime events
- Predictive policing (e.g. Predict next location
of offense, forecast crime levels around
conventions etc.)
Question What are the crime levels 1 hour after
a football game within a radius of 1 mile ?
- Courtsey https//www.llnl.gov/str/September02/Ha
ll.html
Other Applications Epidemiology
5
7Scientific Domain Environmental Criminology
Routine activity theory and Crime Triangle
Crime pattern theory
Courtsey http//www.popcenter.org/learning/60step
s/index.cfm?stepnum8
Courtsey http//www.popcenter.org/learning/60step
s/index.cfm?stepNum16
- Crime Event Motivated offender, vulnerable
victim (available at an appropriate location and
time), absence of a capable guardian.
- Crime Generators offenders and targets come
together in time place, large gatherings (e.g.
Bars, Football games)
- Crime Attractors places offering many
criminal opportunities and offenders may relocate
to these areas (e.g. drug areas)
6
8Outline
- Problem Statement
- Spatio-temporal frequent pattern mining problem
- Challenges
7
9Spatio-temporal frequent pattern mining problem
- Given
- Spatial / Spatio-temporal framework.
- Crime Reports with type, location and / or time.
- Spatial Features of interest (e.g. Bars).
- Interest measure threshold (P?)
- Spatial / Spatio-temporal neighbor relation.
- Find
- Frequent patterns with interestingness gt P?
- Objective
- Minimize computation costs.
- Constraints
- Correctness and Completeness.
- Statistical Interpretation (i.e. account for
autocorrelation or heterogeneity)
8
10Illustration Output
Regional Co-location patterns (Inputs Spatial
Neighborhood 1 mile, Threshold- 0.25)
9
11Challenges
Time partitioning misses relationships
-
- Spatio-temporal Semantics
- Continuity of space / time
- Partial order
- Conflicting Requirements
- Statistical Interpretation
- Computational Scalability
- Computational Cost
- Exponential set of Candidate patterns
Space partitioning misses relationships
Patterns Exponential ( event types)
10
12Our Contributions
- New Spatio-temporal frequent pattern families.
- Ex Cascading ST Patterns and Regional
Co-location patterns.
- Novel interest measures guarantee statistical
interpretation and computable in polynomial time.
- Scalable algorithms based on properties of
spatio-temporal data and interest measures.
- Experimental evaluation using synthetic and real
crime datasets.
11
13Outline
- Our Approach
- Big Picture
- Cascading Spatio-temporal pattern discovery
- Other Frequent Pattern Families
12
14Spatio-temporal frequent pattern mining (SFPM)
Process of discovering interesting, useful and
non-trivial patterns from spatiotemporal data.
Taxonomy of Spatio-temporal frequent patterns
Input Data Input Data
Spatial Spatio-temporal (ST)
Pattern Semantics Unordered Co-location Patterns ST Co-occurrences
Pattern Semantics Totally Ordered X ST Sequences
Pattern Semantics Partially Ordered X Cascading ST Patterns
Statistical Foundation Autocorrelation Co-location Patterns Cascading ST Patterns
Statistical Foundation Heterogeneity Regional Co-location Patterns X
X Unexplored
Todays Focus
13
15Cascading ST pattern (CSTP)
- Input Crime reports with location and time.
- Output CSTP
- Partially ordered subsets of ST event types.
- Located together in space.
- Occur in stages over time.
14
16Related Pattern Semantics ST Data mining
- ST Co-occurrence Celik et al. 2008, Cao et al.
2006 - Designed for moving object datasets by treating
trajectories as location time series - Performs partitioning over space and time.
- ST Sequence Huang et al. 2008, Cao et al. 2005
- Totally ordered patterns modeled as a chain.
- Does not account for multiply connected
patterns(e.g. nonlinear) - Misses non-linear semantics.
- No ST statistical interpretation.
16
15
17Limitations of Related ST Pattern Semantics
- ST Sequence
- Total order
- Ex. (B?A,A?C)
- No ST statistical interpretation.
- Limitations
- Absence of Partial Order
- Ex. (B?A, B?C, A?C)
16
18Interpretation Model Directed Neighbor Graph
(DNG)
- Nodes Individual Events
- Directed Edge (N1 ? N2) iff
- Neighbor( N1, N2)
- and
- After(N2, N1)
17
19Statistical Foundation Interest Measures
- Instances of CSTP P1 (B?A, B?C, A?C) are
- (B1?A1, B1?C1, A1?C1)
- (B1?A3, B1?C2, A3?C2)
- ? ?(B1?A1 A1? C2 B1 ? C2)
- Cascade Participation Ratio CPR (CSTP, M)
- Conditional Probability of an instance of CSTP
in neighborhood, given an instance of event-type
M - Examples
- Cascade Participation Index CPI(CSTP)
- Min ( CPR(CSTP, M) ) over all M in CSTP
- Example
18
20Analytical Evaluation Statistical Interpretation
Spatial Statistics ST K-Function (Diggle et al.
1995)
- Cascade Participation Index (CPI) is an upper
bound to the ST K-Function per unit volume.
Example
ST -K (B ? A) 2/6 0.33 3/6 0.5 6/6 1
CPI (B ? A) 2/3 0.66 1 1
20
21Comparison with Related Interest Measures
Measure Key Property
Frequency Double counting of pattern instances
Maximum Independent Set (MIS) Size Kuramochi and Karypis, 2004 NP Complete
Scoring Criterion for Bayesian Networks Neopolitan, 2003 Chickering, 1996 NP Complete Learning requires Prior specification
Lower bound on vertex label frequency Frequency based interpretation.
Measure Value
Frequency 3 / (What is the of transactions ?)
MIS 2
Lower Bound on Frequency min1,2,2 1
19
22Computational Structure CSTP Miner Algorithm
- for k in (1,23..K-1) and prevalent CSTP found
do
- Generate size k candidates.
- Compute CSTP instances / Materialize part of DNG
- Calculate interest measure and select prevalent
CSTPs.
- Item sets in Association rule mining
- Chemical compounds/sub graphs in graph mining.
- Directed acyclic graph in CSTP mining
Not part of a conventional apriori setting
21
23CSTP Miner Algorithm Illustration
CPI Threshold 0.33
Null
C.2
0
0.4
0.8
0.75
0.2
0
A.1
B.1
C.3
A.3
0.4
0.4
0.8
C.4
C.1
A.5
B.2
A.2
0.4
A.4
Spatio-temporal join
22
24Computational Structure CSTP Miner Algorithm
Fixed Parameters Spatial neighborhood 0.62
miles and temporal neighborhood 1hr, CPI
threshold 0.0055
- Interest measure evaluation
- Exponential pattern space
- Reduce irrelevant interest measure evaluation
- Compute interest measure efficiently
- Time Ordered Nested Loop Strategy
- Space-Time Partition Join Strategy
23
25CSTP Miner Algorithm Interest Measure Evaluation
- ST Join Strategies Perform each interest
measure computation efficiently - Time Ordered Nested Loop (TONL) Strategy
- Space-Time Partitioning (STP) Strategy
volume of ST neighborhood
C.2
A.1
B.1
C.3
A.3
ST join by plane sweep
Space
C.4
C.1
A.5
A.2
B.2
A.4
Time
Edges 13
24
26CSTP Miner Algorithm Alternative Ideas
- Can neighborhood graph be pre-computed ?
- Trade off Storage versus Online computation
- Cost of Storage
- Pre-computed Graph O(EdgesNodes)
- Example 24
- On-the-fly O(Nodes)
- Example 11
- Cost of computation
- Pre-computed graph O(EdgesNodes)
- Example 24
- On-the-fly O(Nodes Log(Nodes))
- Example 38
- Other factors
- Dense vs Sparse data
- Positive ST autocorrelation
25
27CSTP Miner Algorithm Filtering Strategies
- Key Rationale Enhance Savings filter non
prevalent candidates early
- Key Idea
- CPI has anti-monotone upper bound.
- Multi-resolution ST(MST) filter
- Key Idea
- There exists a low dimensional embedding in
space and time. - Over estimate CPI by coarsening ST dataset.
- If Overestimate (CPI) lt Threshold Pruned
26
28CSTP Miner Algorithm Filtering Strategies
- Multi resolution ST Filter
Summarizing on a coarser neighborhood yields
compression in most cases.
CPI Threshold 0.33
B?A B?C A?C C?A
B.1 A.1 B.1 C.2 A.1 C.2 C.1 A.5
B.1 A.3 B.1 C.3 A.3 C.3
B.2 A.2 B.2 C.1 A.1 C.3
B.2 A.4 A.3 C.4
0.8 0.75 0.4 0.2
B?A B?C A?C C?A
(0,0) (1,0) (0,2) (1,2) (1,2)(1,2) (1,1)(2,0)
(0,2) (1,2) (0,0)(1,1) (1,0)(1,1) (2,1)(2,0)
(1,2)(2,1)
(1,0)(2,1)
0.8 0.75 0.8 0.2
27
29Experimental Evaluation Experiment Setup
Goals 1. Compare different design decisions of
the CSTPM Algorithm - Performance
Run-time 2. Test effect of parameters on
performance - Number of event types,
Dataset Size, Clumpiness Degree Experiment
Platform CPU 3.2GHz, RAM 32GB, OS Linux,
Matlab 7.9
28
30Experimental Evaluation Datasets
Lincoln, NE Dataset
Real Data
- Data size 5 datasets
- Drawn by increments of 2 months
- 5000- 33000 instances
- Event types
- Drawn by increments of 5 event types
- 5 25 event types.
Synthetic Data
- Data size 5 datasets
- 5000- 26000 instances
- Event types
- 5 25 event types.
- Clumpiness Degree
- 5- 25 instances per event type per cell.
29
31Experimental Evaluation Join strategy performance
Question What is the effect of dataset size on
performance of join strategies?
Fixed Parameters Real Data (CPI 0.15, 0.31
Miles, 10 Days) Synthetic data(0.5,25,25)
Trends ST Partitioning improves performance by a
factor of 5-10 on synthetic data and by a factor
of 3 on real data.
30
32Experimental Evaluation Join strategy performance
Experiment 2 What is the effect of of event
types on performance of join strategies?
Fixed Parameters Real Data (CPI 0.15, 0.31
Miles, 10 Days) Synthetic data(0.5,25,25)
Trends ST Partitioning improves performance by a
factor 10 on synthetic data and by a factor of
2.5 on real data.
31
33Experimental Evaluation Filtering strategy
performance
Experiment 3 What is the effect of dataset size
on performance of filtering strategies?
Fixed Parameters Real Data (CPI 0.15, 0.435
Miles, 10 Days) Synthetic data(0.65,70,70)
Trends Filtering improves performance by a
factor 5 on synthetic data and by a factor of 1.5
on real data.
32
34Experimental Evaluation Filtering strategy
performance
Question What is the effect of of event types
on performance of filtering strategies?
Fixed Parameters Real Data (CPI 0.15, 0.435
Miles, 10 Days) Synthetic data(0.65,70,70)
Trends Filtering improves performance by a
factor 2.5 on synthetic data and by a factor of
1.3 on real data.
33
35Experimental Evaluation Filtering strategy
performance
Question What is the effect of clumpiness
degree on different design decisions?
Fixed Parameters CPI 0.5, 15.53 Miles, 1.04
Days
- Trends
- Filtering improves performance by a factor 40
- ST Partitioning improves performance by a factor
of 10.
34
36Lincoln, NE crime dataset Case study
- Is bar closing a generator for crime related
CSTP ?
Bar locations in Lincoln, NE
Questions
- Observation Crime peaks around bar-closing!
- Is bar closing a crime generator ?
- Are there other generators (e.g. Saturday Nights
)?
K.S Test Saturday night significantly different
than normal day bar closing (P-value 1.249x10-7
, K 0.41)
35
37Lincoln, NE crime dataset Case study
36
38Lincoln, NE crime dataset Case study
Pop I Pop II KS P-Val. a 0.05 a 0.2
Sat Night All Year 0.4187 1.249x10-7 Yes Yes
Football Night All Year 0.3400 0.1067 NO Yes
Sat Night Football Night 0.1987 0.7899 NO No
37
39Outline
- Our Approach
- Big Picture
- Cascading Spatio-temporal pattern discovery
- Other Frequent Pattern Families
38
40Regional co-location patterns (RCP)
- Input Spatial Features, Crime Reports.
- Output RCP (e.g. lt (Bar, Assaults), Downtown gt)
- Subsets of spatial features.
- Frequently located in certain regions of a study
area.
39
41Statistical Foundation Accounting for
Heterogenity
- Conditional probability of observing a pattern
instance within a locality given an instance of a
feature within that locality.
Regional Participation Ratio
Example
Regional Participation index
Example
Quantifies the local fraction participating in a
relationship.
40
42Performance Tuning Key Ideas
Key Idea
- Interest Measure shows special pruning
properties in certain subsets of the spatial
framework.
Maximal Locality
Key Properties
- Collection of connected instances.
- Maximal localities are mutually disjoint.
- Contains several RCPs.
Key Observations
- RPI shows anti-monotonicity property within
Maximal Localities - Pruning a co-location, AB, prunes all its
super sets (e.g. ABC, ABCDetc.).
- RPI within a Maximal locality is an upper bound
to RPI of constituent Prevalence localities.
43Performance Tuning
Prevalence Threshold 0.25
Null
C
B
A
ML1
ML2
ML3
AB,0.167
BC,0.167
AC,0.25
AB,0.25
BC,0.33
AC,0.25
?
?
No RCP
No RCP
?
ltBC,PL3(BC)gt,0.167
ltAC,PL1(AC)gt,0.25
?
ltBC,PL4(BC)gt,0.167
Completeness
ABC Pruned Automatically
- Pruning a pattern within a maximal locality does
not prune any valid RCPs.
Compute Maximal Locality
Correctness
Due to upper bound property of RPI
- Accepting a pattern involves additional checks
so that only prevalent RCPs are reported.
Due to anti-monotonicity of RPI
44Experimental Evaluation Spatial Neighborhood Size
- What is the effect of spatial neighborhood size
on performance of different algorithms ? - Fixed Parameters Dataset Size 7498 instances
Features 5 Prevalence Threshold 0.07
of RCPs
Run Time
Trends
- Run Time ML Pruning out performs PS Enumeration
by a factor of 1.5 - 5 - of RCPs examined ML Pruning out performs PS
Enumeration by a factor of 4.13 - 19
45Experimental Evaluation Feature Types
- What is the effect of number of feature types
on performance of different algorithms ? - Fixed Parameters Dataset Size 7498 instances
Spatial neighborhood size 800 feet Prevalence
Threshold 0.07
of RCPs
Run Time
Trends
- Run Time ML Pruning out performs PS Enumeration
by a factor of 1.2 - of RCPs examined ML Pruning out performs PS
Enumeration by a factor of 1.6 3.5
46RCPs from Lincoln Crime Data
This result shows the interaction between Alcohol
and Vandalism apart from highlighting outbreaks
41
47Conclusions
-
- Proposed SFPM techniques (e.g., Cascading ST
Patterns and Regional Co-location patterns) honor
ST Semantics (e.g., Partial order, Continuity). - Interest measures achieve a balance between
statistical interpretation and computational
scalability. - Algorithmic strategies exploiting properties of
ST data (e.g., multiresolution filter) and
properties of interest measures enhance
computational savings.
42
48Future Work Short and Medium Term
X Unexplored
Input Data Input Data
Spatial Spatio-temporal (ST)
Pattern Semantics Unordered ? ?
Pattern Semantics Totally Ordered X ?
Pattern Semantics Partially Ordered X CSTP discovery
Statistical Foundation Autocorrelation ? CSTP discovery
Statistical Foundation Heterogeneity RCP Discovery X
Underlying Framework Euclidean RCP Discovery CSTP discovery
Underlying Framework Non-Euclidean (Networks) X X
Neighbor Relation User specified RCP Discovery CSTP discovery
Neighbor Relation Algorithm Determined X X
Interestingness Criterion Interest measure threshold RCP Discovery CSTP discovery
Interestingness Criterion Threshold free X X
Type of data Boolean / Categorical RCP Discovery CSTP discovery
Type of data Quantitative data (e.g., Climate) X X
43
49Future Work Long Term
-
- Exploring interpretation of discovered patterns
by law enforcement. - ST Predictive analytics, Predictive models based
on SFPM and Predictive policing. - Towards Geo-social analytics for policing (e.g.
Criminal Flash mobs, gangs, groups of offenders
committing crimes) - New ST frequent pattern mining algorithms based
on depth first graph enumeration. - ST frequent pattern mining techniques that
account for patron demographic levels. - Explore evaluation of choloropeth maps via ST
frequent pattern mining. -
43
50 Acknowledgment
- Members of the Spatial Database and Data Mining
Research Group University of Minnesota,
Twin-Cities. - This Work was supported by Grants from U.S.ARMY,
NGA and U.S. DOJ. - Advisor Prof. Shashi Shekhar, Computer Science,
University of Minnesota. - Thesis committee.
- U.S. DOJ National Institute of Justice Mr.
Ronald E. Wilson (Program Manager, Mapping and
Analysis for Public Safety) , Dr. Ned Levine (Ned
Levine and Associates, CrimeStat Program) - U.S. Army Topographic Engineering Center Dr.
J.A.Shine (Mathematician and Statistician,
Geospatial Research and Engineering Division )
and Dr. J.P. Rogers (Additional Director,
Topographic Engineering Center) - Mr. Tom Casady, Public Safety Director (Formerly
Lincoln Police Chief), Lincoln, NE, USA
Thank You for your Questions, Comments and
Attention!
44