Location%20Prediction%20and%20Spatial%20Data%20Mining%20(S.%20Shekhar)

1
Location Prediction and Spatial Data Mining (S.
Shekhar)

• Specific Project in 2001-2002
• Evaluation of location prediction techniques
• Towards high performance parallel implementation
• AHPCRC Relevance Projectile Target Interaction
  Portfolio
• Increase lethality of weapons such as guided
  missiles
• Location prediction for map matching
• to check correctness of missile trajectory
• To identify unanticipated obstacle
• Towards possible rerouting
• Army Relevance in general
• Predicting global hot spots (FORMID)
• Army land management endangered species vs.
  training and war games
• Search for local trends in massive simulation
  data
• Critical infra-structure defense (threat
  assessment)
• Inferring enemy tactics (e.g. flank attack) from
  blobology
• Locating enemy (e.g. sniper in a haystack, sensor
  networks)
• Locating friends to avoid friendly fire

2
Location Prediction

• Problem Definition
• Given 1. Spatial Framework
• 2. Explanatory functions
• 3. A dependent function
• 4. A family of function mappings
• Find A function
• Objective maximize classification accuracy
• Constraints Spatial Autocorrelation in
  dependent function
• Past Approaches
• Non-spatial logistic regression, decision
  trees, Bayesian
• Assume independent distribution for learning
  samples
• Auto-correlation gt poor prediction performance
• Spatial Spatial auto-regression (SAR), Markov
  random field Bayesian classifier (MRF)
• No literature comparing the two!
• Learning algorithms for SAR are slow (took 3
  hours for 5000 data points)!

3
Accomplishments

• Formal Results
• SAR - parametric statistics, provides confidence
  measures in model
• MRF from non-parametric statistics
• SAR MRF-BC linear regression Bayesian
  Classifier
• Rewrite SAR as y (QX) ? Q?, where Q (I-
  ?W)-1
• SAR has linear class boundaries in transformed
  space (QX, y)
• MRF-BC can represent non-linear class boundaries
• Experimental results
• MRF-BC can provide better classification
  accuracies than SAR
• But solution procedure is very slow
• Details in Recent paper in IEEE Transactions on
  Multimedia