Multivariate time-series explanations using paths in a time-dependent latent space

Abstract

Explainable AI (XAI) has gained increasing importance amid the widespread adoption of AI technologies across diverse domains, yet significant challenges persist in explaining neural networks, particularly for multivariate time series data. In this work, we introduce Conditional Time Variational Autoencoder (CTimeVAE), a novel Variational Autoencoder (VAE) architecture designed to effectively capture complex temporal dynamics and accurately reconstruct time series data. Building upon CTimeVAE, We propose Time-Enhanced Integrated Gradients (T-EIG), an advanced XAI approach that can identify critical features in multivariate time series prediction. Our framework is validated through experiments that integrate CTimeVAE with LSTM models in a two-phase process: representation learning via CTimeVAE followed by downstream prediction tasks with LSTM models using these representations. This approach demon- strates how latent representations can improve predictive performance while increasing explainability. Our experimental results across multiple datasets show that both CTimeVAE and T-EIG consistently outperform existing approaches, offering superior reconstruction capabilities and more reliable feature attributions for time series data. This work contributes significantly to the field of time series explainability, providing more effective tools for interpreting complex temporal models in real-life applications.