1. Implementation of ChiMerge (https://www.aaai.org/Papers/AAAI/1992/AAAI92-019.pdf)

Works sklearn way

2. Supervised discretization using target, Chi2 statistics & test
3. Can be configured to multiprocess (n_jobs)

from discretization.chi_merge import *
chi_merge = ChiMerge(con_features=X.columns, significance_level=0.1, n_jobs=-3)
chi_merge.fit_transform(X)

It follows below rules.

• If continuous feature is discretized,
  1. Within interval, class frequency is stable.
  2. Two adjacent intervals should have no similar class frequencies.
• This is tested by Chi2 test

• What if considering k-adjacent, not 2-adjacent ?

  Should be normalized and reflected to formula in paper

• ChiMerge: Discretization of Numeric Attributes (https://www.aaai.org/Papers/AAAI/1992/AAAI92-019.pdf (https://www.aaai.org/Papers/AAAI/1992/AAAI92-019.pdf))
• Discretization: An Enabling Technique (https://cs.nju.edu.cn/zhouzh/zhouzh.files/course/dm/reading/reading03/liu_dmkd02.pdf)

TODO

• dataset to s3