FreqRISE: Explaining time series using frequency masking

TL;DR

FreqRISE introduces a model-agnostic masking framework that explains time series in the frequency and time-frequency domains by applying masks in a dual domain via an invertible transform. It demonstrates that salient information for many time-series tasks is localized in frequency content, enabling more faithful and localized explanations than traditional time-domain saliency methods. Across synthetic data and AudioMNIST, FreqRISE achieves strong localization and faithfulness, particularly in domains where salient features are sparse, while offering a controllable post-processing approach to reduce explanation complexity. The work highlights the practical potential of domain-aware explainability for time-series models and outlines avenues for tuning, domain selection, and efficiency improvements in future research.

Abstract

Time-series data are fundamentally important for many critical domains such as healthcare, finance, and climate, where explainable models are necessary for safe automated decision making. To develop explainable artificial intelligence in these domains therefore implies explaining salient information in the time series. Current methods for obtaining saliency maps assume localized information in the raw input space. In this paper, we argue that the salient information of a number of time series is more likely to be localized in the frequency domain. We propose FreqRISE, which uses masking-based methods to produce explanations in the frequency and time-frequency domain, and outperforms strong baselines across a number of tasks. The source code is available here: \url{https://github.com/theabrusch/FreqRISE}.

Figures (11)

• Figure 1: The FreqRISE framework shown when, as an example, using the Short-Time Discrete Fourier Transform (STDFT) as the invertible mapping, $g$. $g$ maps $\mathbf{X}^T$ into the domain of interest, $g: \mathbf{X}^T \rightarrow \mathbf{X}^S$. Here, we sample $N$ masks and produce $N$ masked versions of $\mathbf{X}^S$. Using the inverse of $g$, we map back to the time domain, obtain $\hat{y}_c$, and compute the relevance map as a weighted average over the masks.
• Figure 2: A sample from the synthetic dataset with salient features at $k=\{5, 16, 53\}$ marked in blue in the frequency domain.
• Figure 3: AudioMNIST: Deletion plots for both tasks in frequency and time-frequency domains. We delete features according to the importance and measure model outputs. FreqRISE outperforms the baselines in both domains on the digit task and the frequency domain on the gender task.
• Figure 4: AudioMNIST: Gender task. We show relevance maps computed by LRP and FreqRISE in the time and frequency domain. The salient information is more localized in the frequency domain.
• Figure 5: AudioMNIST: Digit task. STDFT of the signal on the left and the FreqRISE relevance map in the time-frequency domain on the right. The relevance map shows the benefit of having both the time and frequency axis, when the describing the digit data.