Interpretable Prototype Discovery in Deep Learning-Based Time Series Classification

Multivariable time series classification problems are increasingly prevalent and complex across various domains, such as biology and finance. While deep learning methods are an effective tool for these problems, they often lack interpretability. In this work, we propose a modular prototype learning framework for multivariable time series classification. In the first stage of our framework, we use contrastive learning to train a set of long short-term memory (LSTM) encoders for each variable. We then use prototype matching layers to learn representative prototypes for the variables that contribute to the classification task. The visualization of these prototypes offers an interpretation of the representative patterns within each variable. We validated our framework on a simulated dataset with embedded patterns. On this simulated dataset, we found that our model produces interpretations consistent with the embedded patterns. Overall, our findings pave the road for constructing interpretable deep learning-based models for time series classification.