| Description |
Anomaly detection in large spatial data sets is difficult. Anomaly detection in large spatial data sets with multiple correlated features, becomes even more difficult. Moran's I is a useful function for auto-correlating spatial observations and detecting anomalous observations. Unfortunately, Moran's I has only been developed for single scalar feature comparison. We propose instead to us a vector of features. Now a much more comprehensive data set with feature correlation can be utilized to find outliers based on weighted neighbor values, instead of arbitrary or administrative aggregation. The new enhancements proposed here allow for richer and nuanced data analysis. With the use of Principal Component Analysis and high dimension feature vectors, regions of interest are less ambiguous to detect. We describe new techniques that reduce feature noise as well as computational and operational complexity. Our techniques are also able to replace other dimensional reduction techniques that introduce distortion or skewing to the feature set. |
