From Prediction to Actionable Understanding: Knowledge-Grounded Counterfactual Explanations for Multivariate Time Series

#XAI #time-series #decision-making #machine-learning #estimation #predictive-models #monitoring #state-estimation #neural-networks #intelligence #research #technology #uae

Deep learning has become a central tool in Industry 4.0, where models trained on multivariate time series can detect complex patterns that are difficult for human experts to perceive. However, in operationally sensitive applications, predictive accuracy alone is not sufficient: a prediction that cannot be understood or justified can be difficult to trust and act upon.

This talk presents a post-hoc explainable AI framework that aims to bridge the gap between predictive performance and actionable understanding. The central idea is to move beyond feature attribution and generate counterfactual explanations that answer a more practical question: what should change for the prediction to change? Rather than relying only on numerical proximity, the proposed approach uses domain knowledge to ensure that counterfactual explanations are plausible and semantically consistent with the application context.

The framework combines predictive models for multivariate time series, counterfactual explanation generation, ontology-based validation, and semantic verification using rules. It also integrates large language models to extract structured knowledge from domain documents and to transform validated explanation artefacts into accessible, human-readable explanations. The language model therefore acts as a synthesis layer over validated knowledge rather than as the source of scientific correctness.

The approach is illustrated through two industrial applications: lithium-ion battery state estimation and production-line quality monitoring. The predictive models span several architectures, including CNNs, LSTMs, spiking neural networks, and Transformer-based models. The talk will highlight how combining machine learning with domain knowledge can make explanations not only more understandable, but also more relevant to real-world industrial decision-making.

Click to join (event starts on Oct 21 at 11AM, GMT+4)