Interactive Counterfactual Generation for Univariate Time Series

TL;DR

This work tackles the opacity of deep learning for univariate time series classification by introducing an interactive method to generate counterfactual explanations via 2D projections of data, activations, and attributions. It leverages $UMAP$ for projection and inverse projection techniques, with gradient-based optimization in activation and attribution spaces to synthesize plausible counterfactuals within an interactive visual analytics workspace that includes projection decision boundary maps and time-series line plots. Applied to the ECG5000 dataset with a Conv1D ResNet, the approach demonstrates improved interpretability and user-driven exploration of model decisions. The results point to a scalable path for explainability in time-series tasks and highlight future directions toward multivariate data integration and broader interpretability toolchains.

Abstract

We propose an interactive methodology for generating counterfactual explanations for univariate time series data in classification tasks by leveraging 2D projections and decision boundary maps to tackle interpretability challenges. Our approach aims to enhance the transparency and understanding of deep learning models' decision processes. The application simplifies the time series data analysis by enabling users to interactively manipulate projected data points, providing intuitive insights through inverse projection techniques. By abstracting user interactions with the projected data points rather than the raw time series data, our method facilitates an intuitive generation of counterfactual explanations. This approach allows for a more straightforward exploration of univariate time series data, enabling users to manipulate data points to comprehend potential outcomes of hypothetical scenarios. We validate this method using the ECG5000 benchmark dataset, demonstrating significant improvements in interpretability and user understanding of time series classification. The results indicate a promising direction for enhancing explainable AI, with potential applications in various domains requiring transparent and interpretable deep learning models. Future work will explore the scalability of this method to multivariate time series data and its integration with other interpretability techniques.

Figures (4)

• Figure 1: Overview of the application: On top, visualizes the decision boundary maps (DBM) of the projections of time series, activations, and attributions of a deep learning time series classifier. The arrows between data points visualize dragged points by a user towards counterfactual explanations. A line plot on the bottom (LP) presents the corresponding time series to the dragged data points in the scatter plot. The highlighted line (upfront with a black stroke) is also highlighted in the scatter plots. The dragline for the points demonstrates interesting patterns in the activations and attributions during a generation of a counterfactual.
• Figure 2: Based on a click on the projected activations, the inverse projections generate activations. These activations are used in an optimization loop that takes a generated time series and refines it based on the difference between the current time series activations of the model and the inverse-generated activation.
• Figure 3: Generating counterfactual explanations based on the projected attributions by slowly dragging a data point to a region with another class prediction on the dense decision map on the projected attributions. Generated time series seem plausible in the projected time series, and projected activations scatter plots never come close to borders and stay in regions with other data points.
• Figure 4: On top, visualizes the decision boundary maps (DBM) of the projections of time series, activations, and attributions of a deep learning time series classifier. The arrows between data points visualize dragged points by a user towards counterfactual explanations in the activations. A line plot on the bottom (LP) presents the corresponding time series to the dragged data points in the scatter plot. The highlighted line (upfront with a black stroke) is also highlighted in the scatter plots. The dragline for the points demonstrates interesting patterns in the original time series and attributions during a generation of a counterfactual in the activations. Especially interesting is the attribution DBM, as the generated time series jumps around quite heavily.