TY - JOUR AU - Sliwinski, Jakub AU - Strobel, Martin AU - Zick, Yair PY - 2019/07/17 Y2 - 2024/03/28 TI - Axiomatic Characterization of Data-Driven Influence Measures for Classification JF - Proceedings of the AAAI Conference on Artificial Intelligence JA - AAAI VL - 33 IS - 01 SE - AAAI Special Technical Track: AI for Social Impact DO - 10.1609/aaai.v33i01.3301718 UR - https://ojs.aaai.org/index.php/AAAI/article/view/3850 SP - 718-725 AB - <p>We study the following problem: given a labeled dataset and a specific datapoint <em>∼x</em>, how did the <em>i</em>-th feature influence the classification for <em>∼x</em>? We identify a family of <em>numerical influence measures</em> — functions that, given a datapoint <em>∼x</em>, assign a numeric value <em>φ</em><sub><em>i</em></sub>(<em>∼x</em>) to every feature <em>i</em>, corresponding to how altering <em>i</em>’s value would influence the outcome for <em>∼x</em>. This family, which we term <em>monotone influence measures (MIM)</em>, is uniquely derived from a set of desirable properties, or axioms. The MIM family constitutes a provably sound methodology for measuring feature influence in classification domains; the values generated by MIM are based on the dataset alone, and do not make any queries to the classifier. While this requirement naturally limits the scope of our framework, we demonstrate its effectiveness on data.</p> ER -