User Driven Model Adjustment via Boolean Rule Explanations

Authors: Elizabeth M. Daly, Massimiliano Mattetti, Öznur Alkan, Rahul Nair5896-5904

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental Evaluation
Researcher Affiliation Industry Elizabeth M. Daly, Massimiliano Mattetti, Oznur Alkan, Rahul Nair IBM Research Dublin, Ireland
Pseudocode Yes Algorithm 1: Generate Response, Algorithm 2: Evaluate Feedback Rules, Algorithm 3: Example of a transformation on a categorical feature when the class label is preserved, Algorithm 4: Example of a transformation on a numeric feature when the class label is changed
Open Source Code No The paper uses an open-source library AIX360 (footnote 3: https://github.com/Trusted-AI/AIX360) for BRCG implementation, but does not provide access to the code for their own interactive overlay solution.
Open Datasets Yes We select four well known binary classification benchmarks from the UCI repository TIC-TAC-TOE, BANKNOTE, BANK-MKT and BREAST CANCER. https://archive.ics.uci.edu/ml/datasets/
Dataset Splits No The paper states: 'The data is divided into 80% for training and 20% for the holdout test set.' It describes training and test splits but does not explicitly mention a separate validation set split.
Hardware Specification Yes We implement all algorithms in Python using scikitlearn (Pedregosa et al. 2011) and perform the experiments on a cluster of Intel Xeon CPU E5-2683 processors at 2.00GHz with 8 Cores and 64GB of RAM.
Software Dependencies No The paper mentions using scikit-learn and the AIX360 library for BRCG implementation, but does not specify their version numbers or the Python version used.
Experiment Setup Yes For the purposes of these experiments the underlying machine learning algorithm used is a logistic regression with 500 iteration limit1. Numeric features are pre-processed with Standard Scaler and the categorical one with a One Hot Encoder. At each iteration a batch of 10 instances is selected from the pool, labelled by the oracle with the ground truth, and then used for retraining the ML model.