Explaining Single Predictions: A Faster Method

Abstract

Machine learning has proven increasingly essential in many fields. Yet, a lot obstacles still hinder its use by non-experts. The lack of trust in the results obtained is foremost among them, and has inspired several explanatory approaches in the literature. In this paper, we are investigating the domain of single prediction explanation. This is performed by providing the user a detailed explanation of the attribute's influence on each single predicted instance, related to a particular machine learning model. A lot of possible explanation methods have been developed recently. Although, these approaches often require an important computation time in order to be efficient. That is why we are investigating about new proposals of explanation methods, aiming to increase time performances, for a small loss in accuracy.

Improving on Coalitional Prediction Explanation

“…On this way, obtaining explanations for predictive models, in a reasonable time, is essential. In a previous work [5], we proved the feasibility of lowering the computation time of existing solutions, with a very limited loss of explanation accuracy while saving a high computation time. In this paper, we continue this work to find better approximations of these solution, through the exploration of new ways to find groups of attributes.…”

“…In this work, we want to facilitate the generation of prediction explanation, without having to restrict ourselves to a given set of models. This paper is the continuity of [5] in which we already established possible methods of simplification. One of these methods relies on the automatic detection of groups of attributes.…”

“…A more efficient strategy is to only identify the groups of correlated attributes, as proposed in our previous work [5]. This strategy avoids having to calculate all possible subgroups of influence.…”

“…This is achieved through new way to find groups of attributes for the coalitional method described in Section 3. Experiments are presented in Section 5, showing the interest of our methods in terms of computation time and their limited impacts in terms of loss of accuracy, significantly improving the results of [5]. Finally, Section 6 concludes the paper by giving new perspective of works, including the new possibilities opened by our results.…”

“…Then, Section 3 describes the base methods used to generate predictions explanation. The extension of our work [5] is proposed in Section 4 to find faster methods of explanation. This is achieved through new way to find groups of attributes for the coalitional method described in Section 3.…”