Learning multivariate shapelets with multi-layer neural networks for interpretable time-series classification

Shapelets are discriminative subsequences extracted from time-series data. Classifiers using shapelets have proven to achieve performances competitive to state-of-the-art methods, while enhancing the model's interpretability. While a lot of research has been done for univariate time-series shapelets, extensions for the multivariate setting have not yet received much attention. To extend shapelets-based classification to a multidimensional setting, we developed a novel architecture for shapelets learning, by embedding them as trainable weights in a multi-layer Neural Network. We also investigated the introduction of a novel learning strategy for the shapelets, comprising of two additional terms in the optimization goal, to retrieve a reduced set of uncorrelated shapelets. This paper describes the proposed architecture and presents results on ten publicly available benchmark datasets, as well as a comparison with existing state-of-the-art methods. Moreover, the proposed optimization objective leads the model to automatically select smaller sets of uncorrelated shapelets, thus requiring no additional manual optimization on typically important hyper-parameters such as number and length of shapelets. The results show how the proposed approach achieves competitive performance across the datasets, and always leads to a significant reduction in the number of shapelets used. This can make it faster for a domain expert to match shapelets to real patterns, thus enhancing the interpretability of the model. Finally, since the shapelets learnt during training can be extracted from the model they can serve as meaningful insights on the classifier's decisions and the interactions between different dimensions.