A graph theoretic approach to assess quality of data for classification task

Abstract

The correctness of predictions rendered by an AI/ML model is key to its acceptability. To foster researchers’ and practitioners’ confidence in the model, it is necessary to render an intuitive understanding of the workings of a model. In this work, we attempt to explain a model’s working by providing some insights into the quality of data. While doing this, it is essential to consider that revealing the training data to the users is not feasible for logistical and security reasons. However, sharing some interpretable parameters of the training data and correlating them with the model’s performance can be helpful in this regard. To this end, we propose a new measure based on Euclidean Minimum Spanning Tree (EMST) for quantifying the intrinsic separation (or overlaps) between the data classes. For experiments, we use datasets from diverse domains such as finance, medical, and marketing. We use state-of-the-art measure known as Davies Bouldin Index (DBI) to validate our approach on four different datasets from aforementioned domains. The experimental results of this study establish the viability of the proposed approach in explaining the working and efficiency of a classifier. Firstly, the proposed measure of class-overlap quantification has shown a better correlation with the classification performance as compared to DBI scores. Secondly, the results on multi-class datasets demonstrate that the proposed measure can be used to determine the feature importance so as to learn a better classification model.