On the Safety of Interpretable Machine Learning: A Maximum Deviation Approach

Authors: Dennis Wei, Rahul Nair, Amit Dhurandhar, Kush R. Varshney, Elizabeth Daly, Moninder Singh

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 5, we present case studies that illustrate the deviation maximization methods in Section 4 for decision trees, linear and additive models, and tree ensembles. It is seen that deviation maximization provides insights about models through studying the feature combinations that lead to extreme outputs. These insights can in turn direct further investigation and invite domain expert input. We also quantify how the maximum deviation depends on model complexity and the size of the certification set. For tree ensembles, we find that the obtained upper bounds on the maximum deviation are informative, showing that the maximum deviation does not increase with the number of trees in the ensemble. Two datasets are featured: a sample of US Home Mortgage Disclosure Act (HMDA) data (see Appendix D.2 for details), meant as a proxy for a mortgage approval scenario, and the UCI Adult Income dataset [47], a standard tabular dataset with mixed data types.
Researcher Affiliation Industry Dennis Wei IBM Research dwei@us.ibm.com Rahul Nair IBM Research rahul.nair@ie.ibm.com Amit Dhurandhar IBM Research adhuran@us.ibm.com Kush R. Varshney IBM Research krvarshn@us.ibm.com Elizabeth M. Daly IBM Research elizabeth.daly@ie.ibm.com Moninder Singh IBM Research moninder@us.ibm.com
Pseudocode Yes Appendix B.3 presents the full algorithm.
Open Source Code No The code is proprietary at this time due to our institutional obligations.
Open Datasets Yes Two datasets are featured: a sample of US Home Mortgage Disclosure Act (HMDA) data (see Appendix D.2 for details), meant as a proxy for a mortgage approval scenario, and the UCI Adult Income dataset [47], a standard tabular dataset with mixed data types.
Dataset Splits Yes For the Adult Income dataset, we split the data into 80% training and 20% test sets. We use a 70/30 split for the HMDA dataset, also without a validation set, due to its size.
Hardware Specification Yes All experiments are conducted on a 3.1 GHz Intel Xeon W-2145 CPU with 128 GB RAM.
Software Dependencies No Experiments use the InterpretML package [48] and scikit-learn [65]. No specific version numbers for these software dependencies are provided.
Experiment Setup Yes EBMs are trained with max_bins=32, and for Random Forests, we use the default parameters in scikit-learn, with the number of estimators varied from 1 to 500.