Towards Explainable Time Series

: An Exploration Of Post-hoc Local Explanations From A Multi-disciplinary Perspective

Abstract

Black-box Artificial Intelligence, which describes systems so complex they are uninterpretable to humans, is now ubiquitous across society yet the lack of transparency surrounding how these models arrive at decisions is concerning. Explainable AI attempts to bridge the interpretability gap between human and black-box.

Within the Explainable AI landscape, post-hoc local explanations, which help users understand why a black-box returned a particular output for a particular input, have become the most widely applied tool. Despite their success, post-hoc explanations are still in their infancy and accompanied by limitations. In this thesis we provide four novel methods of explanation. We focus on LIME and SHAP, two of the most popular post-hoc local explanation methods. Our methods address the limitations of both LIME and SHAP in the following ways:

Both LIME and SHAP do not consider the complex temporal dependency of time series as they were not designed for this data-structure. We adapt both LIME and SHAP to provide two novel attribution methods which are specifically designed for univariate and multivariate time series.

The Shapley value is one of the most popular methods for value attribution. However, when applied for feature attribution it has been known to generate misleading explanations. By unifying SHAP with its game-theoretic context we explore why this behaviour occurs and propose a new attribution method which generates more robust explanations in the presence of interacting features.

The relationship between Explainable AI and the Philosophy of Causality is complex and debated throughout the community. We integrate both disciplines through the language of counterfactuals and propose an alternative causal perspective on a post-hoc explanation.

In this thesis, alongside methodological contributions, we unify post-hoc local explainability with the rich multi-disciplinary background underpinning the explanation sciences which, we believe, is fundamental in bridging the interpretability gap between human and black box.

Date of Award	5 Dec 2023
Original language	English
Awarding Institution	University of Bristol
Supervisor	Peter A Flach (Supervisor)