Conflict Resolution for Improving ML Accuracy

This paper investigates how to make practical use of conflict resolution (CR) to enhance the accuracy of ML classifiers $\mathcal{M}$ on relational data. We show that applying CR to influential attributes and/or tuples can substantially improve the performance of downstream model $\mathcal{M}$. Based on this, we formulate two problems for identifying influential attributes and tuples in the data to maximize model accuracy. Although we show that both problems are intractable, we develop effective algorithms to pinpoint these critical factors. To mitigate the impact of noise introduced by imprecise CR methods, we propose a creator-critic framework that iteratively applies CR and trains $\mathcal{M}$ with the corrected data. We prove that the creator-critic process guarantees convergence to a more accurate model. Using real-life datasets, we experimentally verify that our approach improves the relative accuracy of various ML classifiers by an average of 40.5%.