Learning Fair and Interpretable Rules Based on Linear Programming by Hakan Akyuz  

This talk being part of the Social Impact of AI expert practice makes an introduction to the concepts of interpretability and fairness in machine learning. We introduce two algorithms for interpretability and fairness based on using rule sets. Rules embody a set of if-then statements which include one or more conditions to classify a subset of samples in a dataset. In various applications, such classification rules are considered interpretable by the decision-makers. Fairness is usually built upon interpretability and needs to be simultaneously integrated within it. Both algorithms take advantage of linear programming (LP), and hence, they are scalable to large data sets.

The first algorithm extracts rules for the interpretation of trained models that are based on tree/rule ensembles. The second algorithm generates a set of classification rules through a column generation approach. The proposed algorithms return a set of rules along with their optimal weights indicating the importance of each rule for classification. Fairness requires additional conditions to be imposed on the LPs and restricts the set of rules generated. This boils down to obtaining a fair subset of rules in the end.

Our algorithms allow assigning cost coefficients, which could relate to different attributes of the rules, such as rule lengths, estimator weights, number of false negatives, discrimination ratio and so on. Thus, the decision-makers can adjust these coefficients and constraints to divert the training process and obtain a set of rules that are more appealing for their needs. We have tested the performances of our algorithms on a collection of datasets.

Our results show that a good compromise between interpretability and accuracy can be obtained by the proposed algorithms. We have more promising outcomes for fairness compared to a state-of-the-art approach on average.