Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interactive dense pixel visualizations for time series and model attribution explanations

Udo Schlegel, Daniel A. Keim

TL;DR

The paper tackles interpretability in time-series XAI by introducing DAVOTS, a dense-pixel visualization that places raw time series, neural activations, and attributions in a single row augmented with histograms. It leverages hierarchical clustering with multiple distance metrics to order samples, enabling pattern discovery across large datasets. Demonstrated on a CNN trained on the FordA dataset, DAVOTS reveals how different data types reveal complementary patterns and how clustering choice affects pattern visibility. The work outlines concrete future directions to enhance interactivity and clustering methods for deeper model understanding in time-series contexts.

Abstract

The field of Explainable Artificial Intelligence (XAI) for Deep Neural Network models has developed significantly, offering numerous techniques to extract explanations from models. However, evaluating explanations is often not trivial, and differences in applied metrics can be subtle, especially with non-intelligible data. Thus, there is a need for visualizations tailored to explore explanations for domains with such data, e.g., time series. We propose DAVOTS, an interactive visual analytics approach to explore raw time series data, activations of neural networks, and attributions in a dense-pixel visualization to gain insights into the data, models' decisions, and explanations. To further support users in exploring large datasets, we apply clustering approaches to the visualized data domains to highlight groups and present ordering strategies for individual and combined data exploration to facilitate finding patterns. We visualize a CNN trained on the FordA dataset to demonstrate the approach.

Interactive dense pixel visualizations for time series and model attribution explanations

TL;DR

The paper tackles interpretability in time-series XAI by introducing DAVOTS, a dense-pixel visualization that places raw time series, neural activations, and attributions in a single row augmented with histograms. It leverages hierarchical clustering with multiple distance metrics to order samples, enabling pattern discovery across large datasets. Demonstrated on a CNN trained on the FordA dataset, DAVOTS reveals how different data types reveal complementary patterns and how clustering choice affects pattern visibility. The work outlines concrete future directions to enhance interactivity and clustering methods for deeper model understanding in time-series contexts.

Abstract

The field of Explainable Artificial Intelligence (XAI) for Deep Neural Network models has developed significantly, offering numerous techniques to extract explanations from models. However, evaluating explanations is often not trivial, and differences in applied metrics can be subtle, especially with non-intelligible data. Thus, there is a need for visualizations tailored to explore explanations for domains with such data, e.g., time series. We propose DAVOTS, an interactive visual analytics approach to explore raw time series data, activations of neural networks, and attributions in a dense-pixel visualization to gain insights into the data, models' decisions, and explanations. To further support users in exploring large datasets, we apply clustering approaches to the visualized data domains to highlight groups and present ordering strategies for individual and combined data exploration to facilitate finding patterns. We visualize a CNN trained on the FordA dataset to demonstrate the approach.
Paper Structure (6 sections, 2 figures)

This paper contains 6 sections, 2 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Clustering the data
  4. Dense-Pixel Attribution Visualization On Time Series
  5. Current preliminary findings
  6. Conclusions and Future Work

Figures (2)

  • Figure 1: Overview of DAVOTS. (V) shows the visualization approach, while (P) shows the adjustable parameters. Starting from the left (V), the pixel visualization of the raw time series, next the histogram of the raw time series, afterward the activations of the model, then the corresponding histogram, next the attributions, then the corresponding histogram, and lastly, the probabilities of the prediction. On the bottom at (P), the parameters can be changed. (P1) demonstrates the standard deviation of every dataset sample for the selected clustering base as a slider to select the data which should be shown. (P2) presents the options for selecting the data and clustering. Here the stage (Train, Test, ...), the attribution methods, the clustering base, and the clustering method can be selected. With (P3), the complete visualization can be redrawn to incorporate the new parameters.
  • Figure 2: Two findings we have during our initial exploration on a CNN trained for the FordA dataset. (U2) presents the data clustered by the raw time series using the Ward method wardjr_hierarchical_1963 with the normalized Euclidean distance vanwijk_cluster_1999. (U2a) presents on the left patterns in the raw time series, which are less obvious to observe in the attributions on the right. (U2b) shows clusters in the raw time series and smaller ones in the attributions. (U3) visualizes the clustering using the same settings as before on the activations. (U3a) shows mostly no patterns in the raw time series and attributions. (U3b) presents patterns on the activations, the raw time series, and also on the attributions.