FREQuency ATTribution: benchmarking frequency-based occlusion for time series data

TL;DR

FreqAtt is presented — a framework that enables post-hoc interpretation of time-series analysis and shows that using frequency-based attribution, especially in combination with traditional attribution on top of the frequency-optimized signal, provides strong performance across different metrics.

Abstract

Deep neural networks are among the most successful algorithms in terms of performance and scalability across different domains. However, since these networks are black boxes, their usability is severely restricted due to a lack of interpretability. Existing interpretability methods do not address the analysis of time-series-based networks specifically enough. This paper shows that an analysis in the frequency domain can not only highlight relevant areas in the input signal better than existing methods but is also more robust to fluctuations in the signal. In this paper, FreqAtt is presented - a framework that enables post-hoc interpretation of time-series analysis. To achieve this, the relevant frequencies are evaluated, and the signal is either filtered or the relevant input data is marked. FreqAtt is evaluated using a wide range of statistical metrics to provide a broad overview of its performance. The results show that using frequency-based attribution, especially in combination with traditional attribution on top of the frequency-optimized signal, provides strong performance across different metrics.

Figures (16)

• Figure 1: Dataset Statistics: Shows different statistical values computed for the full dataset and its subset the CharacterTrajectories dataset highlighting that there is no statistical evidence indicating that the subset differs from the full dataset.
• Figure 2: Class Distribution: Shows the class distribution for the full CharacterTrajectories dataset and its subset, highlighting that the subset maintains same distribution as the full dataset.
• Figure 3: Area under the curve: Shows the area under the curve (AUPC) for a deletion test comparing frequency attribution with traditional attribution. Frequency_x_Attribution refers to first optimizing the input using frequency attribution and then applying traditional attribution on top. Across all datasets, random attribution exhibited the worst performance.
• Figure 4: Infidelity, Sensitivity, and Continuity: Shows the Sensitivity, Infidelity, and Continuity across all evaluated datasets. The results indicate that the ranking varies significantly depending on the metric; for example, frequency attribution achieves the best performance for Infidelity but performs poorly in terms of Continuity and Sensitivity.
• Figure 5: Visual comparison of the input in time series space: Shows the comparison of traditional attribution, frequency attribution, and their combination applied to different character trajectory samples. The frequency-optimized signal exhibits a smoother appearance while preserving the overall shape of the original signal.