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

Deep learning models for time series analysis lack transparency: While these AI systems are increasingly accurate at identifying patterns in complex data over time, they function as "black boxes" that don't explain why they make specific predictions, creating challenges for decision-makers who need to understand and trust the system's reasoning.

Counterfactual explanations offer practical "what-if" insights: Rather than just saying what a model decided, counterfactual methods show what would need to change in the data for the model to reach a different conclusion—for example, "if this patient's heart rate had been 5 beats per minute lower, the diagnosis would have been different." This helps stakeholders understand decision boundaries in actionable terms.

CONFETTI balances three competing priorities for real-world use: The proposed method simultaneously optimizes for confidence (reliable predictions), proximity (minimal changes needed), and sparsity (focusing on few key factors). Previous approaches typically excelled at only one or two of these, limiting their practical applicability in high-stakes domains like healthcare or finance.

Minimal modifications make explanations more trustworthy: By identifying the smallest possible changes needed to alter a prediction, CONFETTI provides more credible explanations than methods requiring extensive data modifications, which may introduce unrealistic or implausible scenarios that undermine user confidence.

Significant performance improvements across multiple evaluation metrics: CONFETTI achieved at least 10% higher confidence in its explanations while improving sparsity in 40% or more of test cases compared to existing counterfactual explanation methods, suggesting practical readiness for deployment.