Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Counterfactual Explainable AI (XAI) Method for Deep Learning-Based Multivariate Time Series Classification

Alan G. Paredes Cetina, Kaouther Benguessoum, Raoni Lourenço, Sylvain Kubler

TL;DR

This work tackles the lack of transparency in deep learning for multivariate time series (MTS) classification by introducing CONFETTI, a multi-objective counterfactual explanation method. CONFETTI identifies a counterfactual target via a nearest unlike neighbor, locates influential subsequences with CAM-based weighting, and optimizes counterfactuals to jointly maximize prediction confidence, minimize changes (proximity), and enforce sparsity, all while ensuring plausibility and validity by design. The approach is evaluated on seven MTS datasets with two architectures (FCN and ResNet), showing superior or competitive performance across metrics (confidence, sparsity, proximity, plausibility, and validity) compared to state-of-the-art baselines, with robust sensitivity and ablation analyses. Practically, CONFETTI provides actionable, sparse, and high-confidence explanations that remain close to the data manifold, enabling more transparent and trustworthy decision support in real-world MTS tasks. The method also demonstrates model-agnostic viability, though leveraging CAM can enhance efficiency and interpretability when available.

Abstract

Recent advances in deep learning have improved multivariate time series (MTS) classification and regression by capturing complex patterns, but their lack of transparency hinders decision-making. Explainable AI (XAI) methods offer partial insights, yet often fall short of conveying the full decision space. Counterfactual Explanations (CE) provide a promising alternative, but current approaches typically prioritize either accuracy, proximity or sparsity -- rarely all -- limiting their practical value. To address this, we propose CONFETTI, a novel multi-objective CE method for MTS. CONFETTI identifies key MTS subsequences, locates a counterfactual target, and optimally modifies the time series to balance prediction confidence, proximity and sparsity. This method provides actionable insights with minimal changes, improving interpretability, and decision support. CONFETTI is evaluated on seven MTS datasets from the UEA archive, demonstrating its effectiveness in various domains. CONFETTI consistently outperforms state-of-the-art CE methods in its optimization objectives, and in six other metrics from the literature, achieving $\geq10\%$ higher confidence while improving sparsity in $\geq40\%$.

Counterfactual Explainable AI (XAI) Method for Deep Learning-Based Multivariate Time Series Classification

TL;DR

This work tackles the lack of transparency in deep learning for multivariate time series (MTS) classification by introducing CONFETTI, a multi-objective counterfactual explanation method. CONFETTI identifies a counterfactual target via a nearest unlike neighbor, locates influential subsequences with CAM-based weighting, and optimizes counterfactuals to jointly maximize prediction confidence, minimize changes (proximity), and enforce sparsity, all while ensuring plausibility and validity by design. The approach is evaluated on seven MTS datasets with two architectures (FCN and ResNet), showing superior or competitive performance across metrics (confidence, sparsity, proximity, plausibility, and validity) compared to state-of-the-art baselines, with robust sensitivity and ablation analyses. Practically, CONFETTI provides actionable, sparse, and high-confidence explanations that remain close to the data manifold, enabling more transparent and trustworthy decision support in real-world MTS tasks. The method also demonstrates model-agnostic viability, though leveraging CAM can enhance efficiency and interpretability when available.

Abstract

Recent advances in deep learning have improved multivariate time series (MTS) classification and regression by capturing complex patterns, but their lack of transparency hinders decision-making. Explainable AI (XAI) methods offer partial insights, yet often fall short of conveying the full decision space. Counterfactual Explanations (CE) provide a promising alternative, but current approaches typically prioritize either accuracy, proximity or sparsity -- rarely all -- limiting their practical value. To address this, we propose CONFETTI, a novel multi-objective CE method for MTS. CONFETTI identifies key MTS subsequences, locates a counterfactual target, and optimally modifies the time series to balance prediction confidence, proximity and sparsity. This method provides actionable insights with minimal changes, improving interpretability, and decision support. CONFETTI is evaluated on seven MTS datasets from the UEA archive, demonstrating its effectiveness in various domains. CONFETTI consistently outperforms state-of-the-art CE methods in its optimization objectives, and in six other metrics from the literature, achieving higher confidence while improving sparsity in .

Paper Structure

This paper contains 30 sections, 1 theorem, 3 equations, 3 figures, 9 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Limitations
  4. CE Problem Definition
  5. Problem Statement.
  6. CONFETTI
  7. Feature Weight Vector
  8. Retrieve the NUN
  9. Subsequence extraction
  10. Naive Stage
  11. Optimization
  12. Correctness of the approach.
  13. Experimental Evaluation
  14. Models & Datasets
  15. Evaluation criteria
  16. ...and 15 more sections

Key Result

Theorem 1

Let $f$ be the classifier, $X_i$ the instance to be explained, and $\theta \in [0,1]$ a confidence threshold. Suppose Confetti is executed with these inputs and returns a set of counterfactuals $C(X_i)$. Let $C_0 \in C(X_i)$ denote the initial counterfactual generated during the naive stage (Lines 5

Figures (3)

  • Figure 1: Counterfactual explanations generated by i) Confetti; ii) CoMTE; and iii) TsEVO for a single dimension of a multivariate time series. Below Confetti example is the Class Activation Map of the Nearest Unlike Neighbour as a heatmap, showing in red the most relevant timesteps.
  • Figure 2: Effect of parameter configurations on counterfactual generation. Panels show: (i) Sparsity as $\alpha$ varies; (ii) Sparsity as $\theta$ varies; (iii) Confidence as $\alpha$ varies; and (iv) Confidence as $\theta$ varies.
  • Figure 3: Critical Difference Diagram showing the average ranking (lower is better) of different objective combinations on counterfactual generation metrics. The lines group combinations that are not significantly different from each other. The different combinations are i) Confidence-Proximity (CO_PR); ii) Confidence-Sparsity (CO_SP); iii) Confidence-Sparsity-Proximity (CO_SP_PR); and iv) Sparsity-Proximity (SP_PR).

Theorems & Definitions (8)

  • Definition 1: Classifier, Instance, Prediction
  • Definition 2: Subsequence
  • Definition 3: Nearest Unlike Neighbor - NUN
  • Definition 4: Counterfactual Set, Counterfactual
  • Definition 5: Prediction Confidence
  • Definition 6: Minimality
  • Theorem 1
  • proof