JavaRRCF: Java implementation of a Robust Random Cut Forest

About

• Inspired by https://github.com/kLabUM/rrcf
• Original paper: http://proceedings.mlr.press/v48/guha16.pdf
• An unsupervised framework for detecting and classifying anomalies in streaming data
• Features:
  • Is well-suited to streaming data
  • Handles high-dimensional data
  • Avoids influence from irrelevant dimensions
  • Handles duplicates which would mask outliers

Usage

General package:

• Includes generalized versions of Random Cut Trees and Random Cut Forests
• Essentially the same as kLabUM/rrcf
• Supports multidimensional data

Memory package (deprecated):

• Theoretically decreases minimum memory usage by about 40%, unclear how this actually affects the numbers in practice though
• Attempted memory-optimized version of RRCF (storing only delta bounding boxes and using a shared buffer for shingled points)
• Less time efficient
• Intended for use with shingled single-dimensional data

Sample run command (with Numenta anomaly benchmark taxi data):

mvn package -DskipTests && curl https://raw.githubusercontent.com/numenta/NAB/master/data/realKnownCause/nyc_taxi.csv | tail -n +2 | awk -F',' '{print $2}' | time bash -c "java -cp target/rrcf-1.0.jar rrcf.ShingleCsv false false 48 200 1000 1234 > ~/Downloads/output.csv"