The Evolution of Spatial Outlier Detection Algorithms An Analysis of Design

1
The Evolution of Spatial Outlier Detection
Algorithms- An Analysis of Design

• CSci 8715 Spatial Databases
• Ryan Stello Kriti Mehra

2
Outline

• Background of the project
• Problem Statement
• Review of issues in traditional outlier detection
• Spatial Outlier Detection
• Outlier Detection in Spatio-Temporal Data
• Issues with modeling spatio-temporal data
• Solutions proposed
• Techniques to model spatio-temporal data
• Issues with detecting outliers in these models
• Summary
• Suggestions for Future Work

3
Background of the project

• - Survey of papers
• Focus is on Spatio-Temporal Outlier Detection
• Classification is done with a view to understand
  outlier detection in the Spatio-Temporal domain

4
Problem Statement

• Given Techniques for outlier detection in
  traditional, spatial and spatio-temporal domain
• Find To provide a classification of spatial
  outlier detection algorithms and highlight
  shortcomings as problem complexity increases to
  the spatio-temporal domain
• Objective To inform the reader of the complexity
  of spatial outlier detection and motivate further
  efforts
• Constraints Non-exhaustive

5
Review of issues in traditional outlier techniques

• Low Dimensional spatial outlier detection
• Restricted so that imposing a grid is easy
• Distributive
• Normalizes the whole data and pulls out an
  outlier
• Average of values is a typical metric
• Median could be used
• Iterative
• Apply to each neighborhood
• High- Dimensional Spatial Outlier Detection
• Statistical Applied to the entire data and hence
  fails
• Space Reduction

6
Spatial Outlier Detection

• Spatial Data
• Define a region
• Define proximity relationship
• Only when region and proximity relationship are
  defined can the concept of spatial outlier be
  defined
• Various methods have been devised to define
  region and proximity relationship.
• Which one should be applied to our application?

7
Spatio-Temporal Data

• Increased Complexity due to
• Issues with spatial outlier detection already
  exist
• A new attribute has to be considered - Time
• High dimensionality
• Scenarios
• Car making a sharp turn
• Movement of a cluster of stars
• Pollution of a lake due to industrial dump
• Global warming
• Should the definition of region be the same?
• Should the definition of proximity relationship
  be the same?
• Would the data model used in these scenarios be
  the same?

8
Modeling Spatio-Temporal Data
Issues and Solutions
9
Issues

• Based on the snapshot phenomena GIS
  applications take photographs of a region
  periodically.
• Difficult to determine whether two mobile systems
  interacted between snapshots
• Drawback Data-oriented, data can be recorded
  only at fixed intervals of time.

F1, F2 Initial and final position of Flock of
sheep I1, I2 Initial and final position of rain
clouds Did the flock of sheep get wet?
10
Solution Proposed

• Transition from Data based modeling to
  Representation based modeling
• Representation of the data is required to
  incorporate spatial and temporal aspect

11
Spatio-Temporal Representations

• - Neighborhood-Based
• - Time Series Matching

12
Neighborhood-Based

• Determine the neighborhood of object
• Merge neighborhood sharing edges based on common
  concept
• Process
• 1. Create Micro Neighborhood based on immediate
  spatial neighborhood to obtain a Voronoi Polygon
• Voronoi Diagram of a set of objects O is the
  subdivision of the
• plane into n polygons, with the property that a
  point q lies in the
• polygon corresponding to an object oi iff dist
  (q,oi)ltdist (q,oj) for
• each oj belonging to O and jltgti
• 2. Create Macro Neighborhood by merging micro
  neighborhoods that share an edge.
• 3. Detect outliers based on how different the
  value is from threshold.

13
Extension to the concept

• Is the temporal aspect embedded in the semantic
  process?
• More spatial than spatio temporal
  representation
• Extension E.g. Cars neighborhood overlap

X
14
Outlier Detection in Mobile Objects

• Data for mobile objects contains large number of
  outliers
• Metric-based outlier detection is not effective
• Non-metric distance based functions Similarity
  Based
• Time Series compared against a known/expected
  time series
• This method has complexity due to difficulty in
• Determining the expected time series
• What is the acceptable tolerance for imprecise
  matches?
• How much noise is acceptable?

15
Example of Spatio-temporal analysis of mobile
object

• Normal behavior
• Representation of normal behavior of a car would
  require defining possibilities ( variations
  caused by taking an exit
• and lane changing)
• Precision
• frequency matching
• deviance from norm

16
Summary
17
Future Work

• Past helps to analyze cause of events
• Food for thought
• Using spatio-temporal outlier detection to
  predict the future is more relevant than using it
  to analyze the past

18

• Questions?