RLTime: Reinforcement Learning-Based Feature Attribution for Interpretable Time Series Models

Fuyuan Lyu, Wesley Chung, Melissa Mozifian, Ankit Vani, Edward J. Smith

Keywords: time series classification, explainable ai, reinforecement-learning

TL;DR: This paper utilize RL to replace the discrete optimization techniques in explainable ai

Abstract: Deep time-series models are widely used in healthcare and finance, where interpretability is essential. Explaining these models is challenging due to temporal dependencies, nonadditive feature interactions, and high-dimensional inputs. Recent approaches learn continuous masks under sparsity constraints to generate attribution maps. While effective, this method has two key limitations: it explores the combinatorial space of feature subsets myopically, often missing synergistic features, and suffers from a soft-to-hard gap, where soft masks used during training misalign with the discrete selections needed at inference. We introduce RLTime, a framework that learns discrete attributions through sequential information acquisition. A masked reconstruction network recovers the latent representation of the reference model from partially observed inputs, such that the change in the reconstructed latent after revealing a feature can be leveraged to quantify its marginal value. This signal defines rewards for a distributional reinforcement learning agent that iteratively unmasks features, balancing exploration and exploitation while operating directly in the discrete action space. The agent’s value function scores the utility of revealing each feature, enabling a clear ranking of features and a non-myopic acquisition policy. Experiments on synthetic and real-world datasets demonstrate that RLTime significantly improves attribution quality, exploration-exploitation balancing, and interpretability.

Primary Area: learning on time series and dynamical systems

Submission Number: 9215