ProtoTSNet: Interpretable Multivariate Time Series Classification With Prototypical Parts

TL;DR

ProtoTSNet addresses the need for interpretable multivariate time series classification by introducing prototypical parts within an ante-hoc framework. It extends ProtoPNet with a grouped-convolution encoder, feature masking, encoder pretraining, and a prototype layer that computes subsequence distances via a $\max$ over all subsequences to learned prototypes, enabling direct input-to-prototype mapping. The approach yields explicit feature importance through a 1x1 mixer and transparent explanations in the form of prototype matches, termed the 'this resembles that' rationale. Empirical evaluation on the UEA multivariate archive shows ProtoTSNet achieving best performance among ante-hoc explainable methods and competitive results relative to non-explainable and post-hoc baselines, with thorough ablation and parameter-sensitivity analyses guiding design choices. This work provides a reproducible, interpretable framework suitable for deployment in safety-critical domains demanding transparent decision-making.

Abstract

Time series data is one of the most popular data modalities in critical domains such as industry and medicine. The demand for algorithms that not only exhibit high accuracy but also offer interpretability is crucial in such fields, as decisions made there bear significant consequences. In this paper, we present ProtoTSNet, a novel approach to interpretable classification of multivariate time series data, through substantial enhancements to the ProtoPNet architecture. Our method is tailored to overcome the unique challenges of time series analysis, including capturing dynamic patterns and handling varying feature significance. Central to our innovation is a modified convolutional encoder utilizing group convolutions, pre-trainable as part of an autoencoder and designed to preserve and quantify feature importance. We evaluated our model on 30 multivariate time series datasets from the UEA archive, comparing our approach with existing explainable methods as well as non-explainable baselines. Through comprehensive evaluation and ablation studies, we demonstrate that our approach achieves the best performance among ante-hoc explainable methods while maintaining competitive performance with non-explainable and post-hoc explainable approaches, providing interpretable results accessible to domain experts.

Figures (13)

• Figure 1: The general idea of how ProtoPNet and ProtoTSNet classify instances. In the case of ProtoPNet, it can be read as "this patch from the test image looks like that prototypical part from the training set." In the case of ProtoTSNet, it is "this subsequence from the test instance resembles that prototypical part from the training set." The ProtoPNet part of the image is sourced from its authors chen2019looks.
• Figure 2: Model Architecture. The input multivariate time series is processed through $l$ binary masks, each preserving a different subset of features controlled by reception parameter $r$. These masked variants are fed into a grouped convolution encoder ($l$ groups), where each group processes a distinct feature subset independently, producing a single latent feature per group. The encoder outputs are mixed through a $1 \times 1$ convolution layer, whose weights enable feature importance calculation. The prototype layer then computes $L_2$ distances between latent subsequences and learned prototypes, applying max pooling over all subsequences to obtain similarity scores per prototype. Finally, these similarity scores are processed by a dense layer to produce class predictions.
• Figure 3: The training process of ProtoTSNet and its different stages mapped to the learning curve. Numbers in brackets represent order of the training stage.
• Figure 4: Prototypical parts (blue background) learned on our synthetic dataset and calculated feature importance. Features are split into separate plots for clarity (each row represents a single prototypical part). Each prototype correctly captures the significant part (initial 40 time steps), however, some of them cover the insignificant area as well. This is specific to the design of our model and the experiment setup since capturing a small portion of the significant part is enough to distinguish the classes. There are three features, of which two are significant for the class distinction and one is not. The feature importance calculated reflects this, and the insignificant feature is dimmed out.
• Figure 5: Critical difference diagram for methods evaluated on the UEA multivariate datasets ($\alpha = 0.05$). ProtoTSNet achieves competitive performance, ranking best among ante-hoc explainable methods (ProtoTSNet, PETSC, Shapelet-based) and showing comparable performance to non-explainable methods (ROCKET, TapNet) and post-hoc explainable method (LITEMVTime). Note that non-explainable methods and post-hoc explainable methods naturally have accuracy advantages as they are not constrained by ante-hoc explainability requirements.