Skip to content

Latest commit

 

History

History
52 lines (39 loc) · 2.2 KB

README.md

File metadata and controls

52 lines (39 loc) · 2.2 KB
Raw

Quantitative Input Influence tool

This code was originally developed for Datta-Sen-Zick Oakland'16 but has evolved since. Presently the tool requires python2 in some cases so if something is going wrong, make sure you run it with python2 instead of python3.

  • To install required python modules:

    make requirements

  • To try on the adult dataset run:

    python2 qii.py adult --show-plot

  • To see additional options, run:

    python2 qii.py -h

Optional arguments

  • -h, --help Show this help message and exit
  • -m {average-unary-individual,unary-individual,discrim,banzhaf,shapley} --measure {...} Quantity of interest
  • -s SENSITIVE, --sensitive SENSITIVE Sensitive field
  • -t TARGET, --target TARGET Target field
  • -e, --erase-sensitive Erase sensitive field from dataset
  • -p, --show-plot Output plot as pdf
  • -o, --output-pdf Output plot as pdf
  • -c {logistic,svm,decision-tree,decision-forest}, --classifier {...} Classifier(s) to use
  • --max_depth MAX_DEPTH Max depth for decision trees and forests
  • --n_estimators N_ESTIMATORS Number of trees for decision forests
  • --seed SEED Random seed, auto seeded if not specified
  • -a ACTIVE_ITERATIONS, --active-iterations ACTIVE_ITERATIONS Active Learning Iterations
  • -r, --record-counterfactuals Store counterfactual pairs for causal analysis
  • -i INDIVIDUAL, --individual INDIVIDUAL Index for Individualized Transparency Report
  • --batch_mode BATCH_MODE Run in batch mode
  • --batch_mode_samples BATCH_MODE_SAMPLES Number of samples to compute.
  • --output_suffix OUTPUT_SUFFIX Output suffix for output in batch mode

Currently supported datasets:

  • adult : UCI Income dataset
  • iwpc : Warfarin dosage
  • nlsy97 : Arrest prediction from the NLSY 97

Currently supported measures:

  • discrim : Unary QII on discrimination
  • average-unary-individual : Average unary QII
  • unary-individual : Unary QII on individual outcome (use -i k) for kth individual
  • general-inf : Influence on average classification
  • shapley : Shapley QII (use -i k) for kth individual
  • banzhaf : Banzhaf QII (use -i k) for kth individual