Mining for Spatial Patterns

1
Mining for Spatial Patterns

• Shashi Shekhar
• Department of Computer Science
• University of Minnesota http//www.cs.umn.edu/s
  hekhar
• Collaborators
• U. of Minnesota V. Kumar, G. Karypis, C.T. Lu,
  W. Wu, Y. Huang, V. Raju, P. Zhang, P. Tan, M.
  Steinbach
• NASA Ames Research Center C. Potter
• California State University, Monterey Bay S.
  Klooster
• This work was partially funded by NASA and Army
  High Performance Computing Center

2
Background

• NSF workshop on GIS and DM (3/99)
• Spatial data - traffic, bird habitats, global
  climate, logistics, ...
• For spatial patterns - outliers, location
  prediction, associations, sequential
  associations, clustering, trends,

3
Framework

• Problem statement capture special needs
• Data exploration maps, new methods
• Try reusing classical methods
• from data mining, spatial statistics
• If reuse is not possible, invent new methods
• Validation, Performance tuning

4
Research Goals

• Research Goals
• modeling of ecological data
• event modeling
• zone modeling.
• finding spatio-temporal patterns
• associations
• predictive models.

A key interest is finding connections between the
ocean and the land.
5
Sources of Earth Science Data

• Before 1950, very sparse, unreliable data.
• Since 1950, reliable global data.
• Ocean temperature and pressure are based on data
  from ships.
• Most land data, (solar, precipitation,
  temperature and pressure) comes from weather
  stations.
• Since 1981, data has been available from Earth
  orbiting satellites.
• FPAR, a measure related to plant
• Since 1999 TERRA, the flagship of the NASA Earth
  Observing System, is providing much more detailed
  data.

6
Example Pattern Teleconnections

• Teleconnections are the simultaneous variation in
  climate and related processes over widely
  separated points on the Earth.
• For example, El Nino is the anomalous warming of
  the eastern tropical region of the Pacific, and
  has been linked to various climate phenomena.
• Droughts in Australia and Southern Africa
• Heavy rainfall along the western coast of South
  America
• Milder winters in the Midwest

7
Net Primary Production (NPP)

• Net Primary Production (NPP) is the net
  assimilation of atmospheric carbon dioxide (CO2)
  into organic matter by plants.
• NPP is driven by solar radiation and can be
  constrained by precipitation and temperature.
• NPP is a key variable for understanding the
  global carbon cycle and ecological dynamics of
  the Earth.
• Keeping track of NPP is important because it
  includes the food source of humans and all other
  organisms.
• Sudden changes in the NPP of a region can have a
  direct impact on the regional ecology.
• An ecosystem model for predicting NPP, CASA (the
  Carnegie Ames Stanford Approach) provides a
  detailed view of terrestrial productivity.

8
Benefits of Data Mining

• Data mining provides earth scientist with tools
  that allow them to spend more time choosing and
  exploring interesting families of hypotheses.
• However, statistics is needed to provide methods
  for determining the statistical significance of
  results.
• By applying the proposed data mining techniques,
  some of the steps of hypothesis generation and
  evaluation will be automated, facilitated and
  improved.
• Association rules provide a new framework for
  detecting relationships between events.

9
Approaches
10
Clustering

• Interested in relationships between regions, not
  points.
• For land, clustering based on NPP or other
  variables, e.g., precipitation, temperature.
• For ocean, clustering based on SST (Sea Surface
  Temperature).
• When raw NPP and SST are used, clustering can
  find seasonal patterns.
• Anomalous regions have plant growth patterns
  which reversed from those typically observed in
  the hemisphere in which they reside, and are easy
  to spot.

11
Clustering
El Nino Regions
SNN clusters of SST that are highly correlated
with El Nino indices.
12
Spatial Association Rule

• Citation Symp. On Spatial Databases 2001
• Problem Given a set of boolean spatial features
• find subsets of co-located features, e.g. (fire,
  drought, vegetation)
• Data - continuous space, partition not natural,
  no reference feature
• Classical data mining approach association rules
• But, Look Ma! No Transactions!!! No support
  measure!
• Approach Work with continuous data without
  transactionizing it!
• confidence Pr.fire at s drought in N(s) and
  vegetation in N(s)
• support cardinality of spatial join of instances
  of fire, drought, dry veg.
• participation min. fraction of instances of a
  features in join result
• new algorithm using spatial joins and apriori_gen
  filters

13
Event Definition

• Convert the time series into sequence of events
  at each spatial location.

14
Interesting Association Patterns

• Use domain knowledge to eliminate uninteresting
  patterns.
• A pattern is less interesting if it occurs at
  random locations.
• Approach
• Partition the land area into distinct groups
  (e.g., based on land-cover type).
• For each pattern, find the regions for which the
  pattern can be applied.
• If the pattern occurs mostly in a certain group
  of land areas, then it is potentially
  interesting.
• If the pattern occurs frequently in all groups of
  land areas, then it is less interesting.

15
Association Rules

• Intra-zone non-sequential Patterns

• Region corresponds to semi-arid grasslands, a
  type of vegetation, which is able to quickly take
  advantage of high precipitation than forests.
• Hypothesis FPAR-Hi events could be related to
  unusual precipitation conditions.

16
Co-location
Can you find co-location patterns from the
following sample dataset?
Answers and
17
Co-location
Can you find co-location patterns from the
following sample dataset?
18
Co-location
Spatial Co-location A set of features
frequently co-located Given A set T of K
boolean spatial feature types Tf1,f2, ,
fk A set P of N locations Pp1, , pN in
a spatial frame work S, pi? P is of some spatial
feature in T A neighbor relation R over
locations in S Find Tc ?subsets of T
frequently co-located Objective Correctness
Completeness Efficiency Constraints R
is symmetric and reflexive Monotonic
prevalence measure
Reference Feature Centric
Window Centric
Event Centric
19
Co-location
Comparison with association rules
Participation index Participation ratio pr(fi, c)
of feature fi in co-location c f1, f2, , fk
fraction of instances of fi with feature f1, ,
fi-1, fi1, , fk nearby 2.Participation index
minpr(fi, c) Algorithm Hybrid Co-location
Miner
20
Spatial Co-location Patterns

• Dataset

• Spatial feature A,B,C and their instances
• Possible associations are (A, B), (B, C), etc.
• Neighbor relationship includes following pairs
• A1, B1
• A2, B1
• A2, B2
• B1, C1
• B2, C2

21
Spatial Co-location Patterns

• Partition approachYasuhiko, KDD 2001
• Support not well defined,i.e. not independent of
  execution trace
• Has a fast heuristic which is hard to analyze for
  correctness/completeness

• Dataset

Spatial feature A,B, C, and their instances
Support A,B1 B,C2
Support A,B 2 B,C2
22
Spatial Co-location Patterns

• Dataset

• Reference feature approach Han SSD 95
• C as reference feature to get transactions
• Transactions (B1) (B2)
• Support (A,B) ? from Apriori algorithm
• Note Neighbor relationship includes following
  pairs
• A1, B1
• A2, B1
• A2, B2
• B1, C1
• B2, C2

Spatial feature A,B, C, and their instances
23
Spatial Co-location Patterns

• Our approach (Event Centric)
• Neighborhood instead of transactions
• Spatial join on neighbor relationship
• Support ? Prevalence
• Participation index min. p_ratio
• P_ratio(A, (A,B)) fraction of instance of A
  participating in join(A,B, neighbor)
• Examples
• Support(A,B)min(2/2,3/3)1
• Support(B,C)min(2/2,2/2)1

• Dataset

Spatial feature A,B, C, and their instances
24
Spatial Co-location Patterns

• Partition approach

• Our approach

• Dataset

Support(A,B)min(2/2,3/3)1
Spatial feature A,B, C, and their instances
Support(B,C)min(2/2,2/2)1
Support A,B 2 B,C2

• Reference feature approach

C as reference feature Transactions (B1)
(B2) Support (A,B) ?
Support A,B1 B,C2
25
Spatial Outliers

• Spatial Outlier A data point that is extreme
  relative to it neighbors
• Case Study traffic stations different from
  neighbors SIGKDD 2001
• Data - space-time plot, distr. Of f(x), S(x)
• Distribution of base attribute
• spatially smooth
• frequency distribution over value domain normal
• Classical test - Pr.item in population is low
• Q? distribution of diff.f(x), neighborhood
  aggf(x)
• Insight this statistic is distributed normally!
• Test (z-score on the statistics) gt 2
• Performance - spatial join, clustering methods

26
Spatial Outlier Detection
Given A spatial graph GV,E A neighbor
relationship (K neighbors) An attribute
function V -gt R An aggregation function
R k -gt R A comparison function
Confidence level threshold ? Statistic test
function ST R -gtT, F Find O vi vi ?V,
vi is a spatial outlier Objective
Correctness The attribute values of vi is
extreme, compared with its neighbors
Computational efficiency Constraints
and ST are algebraic aggregate functions of
and Computation cost dominated by I/O op.
27
Spatial Outlier Detection
Spatial Outlier Detection Test 1. Choice of
Spatial Statistic S(x) f(x)E y?
N(x)(f(y)) Theorem S(x) is normally
distributed if f(x) is
normally distributed 2. Test for Outlier
Detection (S(x) - ?s) / ?s gt ?
Hypothesis I/O cost determined by clustering
efficiency
f(x)
S(x)
28
Spatial Outlier Detection
Results 1. CCAM achieves higher clustering
efficiency (CE) 2. CCAM has lower I/O cost
3. High CE gt low I/O cost 4. Big Page gt high
CE
I/O cost
CE value
Z-order
CCAM
Cell-Tree
29
A Unified Approach Spatial Outliers

• Tests quantitative, graphical
• Results
• Computation spatial self-join
• Tests algebraic functions of join
• Join predicate neighbor relations
• I/O-cost f(clustering efficiency)
• Our algorithm is I/O-efficient for
• Algebric tests

Scatter Plot
Original Data
Our Approach
30
Graphical Spatial Tests
Moran Scatter Plot
Original Data
Variogram Cloud
31
Location Prediction

• Citations IEEE Tran. on Multimedia 2002, SIAM DM
  Conf. 2001, SIGKDD DMKD 2000
• Problem predict nesting site in marshes
• given vegetation, water depth, distance to edge,
  etc.
• Data - maps of nests and attributes
• spatially clustered nests, spatially smooth
  attributes
• Classical method logistic regression, decision
  trees, bayesian classifier
• but, independence assumption is violated ! Misses
  auto-correlation !
• Spatial auto-regression (SAR), Markov random
  field bayesian classifier
• Open issues spatial accuracy vs. classification
  accurary
• Open issue performance - SAR learning is slow!

32
Location Prediction
Given 1. Spatial Framework 2. Explanatory
functions 3. A dependent class 4. A family
of function mappings Find Classification
model Objectivemaximize classification_accurac
y Constraints Spatial Autocorrelation exists

Nest locations
Distance to open water
Water depth
Vegetation durability
33
Motivation and Framework
34
Solution Procedures

• Spatial Autoregression Model (SAR)
• y ?Wy X? ?
• W models neighborhood relationships
• ? models strength of spatial dependencies
• ? error vector
• Solutions
• ? and ? - can be estimated using ML or Bayesian
  stat.
• e.g., spatial econometrics package uses Bayesian
  approach using sampling-based Markov Chain Monte
  Carlo (MCMC) method.
• Likelihood-based estimation requires O(n3) ops.
• Other alternatives divide and conquer, sparse
  matrix, LU decomposition, etc.

35
Evaluation

• Linear Regression
• Spatial Regression
• Spatial model is better

36
Solution Procedures

• Markov Random Field based Bayesian Classifiers
• Pr(li X, Li) Pr(Xli, Li) Pr(li Li) / Pr
  (X)
• Pr(li Li) can be estimated from training data
• Li denotes set of labels in the neighborhood of
  si excluding labels at si
• Pr(Xli, Li) can be estimated using kernel
  functions
• Solutions
• stochastic relaxation Geman
• Iterated conditional modes Besag
• Graph cut Boykov

37
Comparison

• SAR can be rewritten as y (QX) ? Q?
• where Q (I- ?W)-1 which can be viewed as a
  spatial smoothing operation.
• This transformation shows that SAR is similar to
  linear logistic model, and thus suffers with same
  limitations i.e., SAR model assumes linear
  separability of classes in transformed feature
  space
• SAR model also make more restrictive assumptions
  about the distribution of features and class
  shapes than MRF
• The relationship between SAR and MRF are
  analogous to the relationship between logistic
  regression and Bayesian classifiers.
• Our experimental results shows that MRF model
  yields better spatial and classification
  accuracies than SAR predictions.

38
MRF vs. SAR
Confusion Matrix
Spatial Confusion Matrix
39
Experiment Design
40
Conclusion and Future Directions

• Spatial domains may not satisfy assumptions of
  classical methods
• data auto-correlation, continuous geographic
  space
• patterns global vs. local, e.g. spatial outliers
  vs. outliers
• data exploration maps and albums
• Open Issues
• patterns hot-spots, blobology (shape), spatial
  trends,
• metrics spatial accuracy(predicted locations),
  spatial contiguity(clusters)
• spatio-temporal dataset
• scale and resolutions sentivity of patterns
• geo-statistical confidence measure for mined
  patterns

41
Reference

• S. Shekhar, S. Chawla, S. Ravada, A. Fetterer, X.
  Liu and C.T. Liu, Spatial Databases
  Accomplishments and Research Needs, IEEE
  Transactions on Knowledge and Data Engineering,
  Jan.-Feb. 1999.
• S. Shekhar and Y. Huang, Discovering Spatial
  Co-location Patterns a Summary of Results, In
  Proc. of 7th International Symposium on Spatial
  and Temporal Databases (SSTD01), July 2001.
• S. Shekhar, C.T. Lu, P. Zhang, "Detecting
  Graph-based Spatial Outliers Algorithms and
  Applications, the Seventh ACM SIGKDD
  International Conference on Knowledge Discovery
  and Data Mining, 2001.
• S. Shekhar, C.T. Lu, P. Zhang, Detecting
  Graph-based Saptial Outlier, Intelligent Data
  Analysis, To appear in Vol. 6(3), 2002
• S. Shekhar, S. Chawla, the book Spatial
  Database Concepts, Implementation and Trends,
  Prentice Hall, 2002
• S. Chawla, S. Shekhar, W. Wu and U. Ozesmi,
  Extending Data Mining for Spatial Applications
  A Case Study in Predicting Nest Locations, Proc.
  Int. Confi. on 2000 ACM SIGMOD Workshop on
  Research Issues in Data Mining and Knowledge
  Discovery (DMKD 2000), Dallas, TX, May 14, 2000.
• S. Chawla, S. Shekhar, W. Wu and U. Ozesmi,
  Modeling Spatial Dependencies for Mining
  Geospatial Data, First SIAM International
  Conference on Data Mining, 2001.
• S. Shekhar, P.R. Schrater, R. R. Vatsavai, W. Wu,
  and S. Chawla, Spatial Contextual Classification
  and Prediction Models for Mining Geospatial
  Data,To Appear in IEEE Transactions on
  Multimedia, 2002.
• S. Shekhar, V. Kumar, P. Tan. M. Steinbach, Y.
  Huang, P. Zhang, C. Potter, S. Klooster, Mining
  Patterns in Earth Science Data, IEEE Computing
  in Science and Engineering (Submitted)

42
Reference

• S. Shekhar, C.T. Lu, P. Zhang, A Unified
  Approach to Spatial Outliers Detection, IEEE
  Transactions on Knowledge and Data Engineering
  (Submitted)
• S. Shekhar, C.T. Lu, X. Tan, S. Chawla, Map Cube
  A Visualization Tool for Spatial Data Warehouses,
  as Chapter of Geographic Data Mining and
  Knowledge Discovery. Harvey J. Miller and Jiawei
  Han (eds.), Taylor and Francis, 2001, ISBN
  0-415-23369-0.
• S. Shekhar, Y. Huang, W. Wu, C.T. Lu, What's
  Spatial about Spatial Data Mining Three Case
  Studies , as Chapter of Book Data Mining for
  Scientific and Engineering Applications. V.
  Kumar, R. Grossman, C. Kamath, R. Namburu (eds.),
  Kluwer Academic Pub., 2001, ISBN 1-4020-0033-2
• Shashi Shekhar and Yan Huang , Multi-resolution
  Co-location Miner a New Algorithm to Find
  Co-location Patterns in Spatial Datasets, Fifth
  Workshop on Mining Scientific Datasets (SIAM 2nd
  Data Mining Conference), April 2002