Multivariate time series classification is a widely known problem, and its applications are ubiquitous. Due to their strong generalization capability, neural networks have been proven to be very powerful for the task, but their applicability is often limited by their intrinsic black-box nature. Recently, temporal decision trees have been shown to be a serious alternative to neural networks for the same task in terms of classification performances, while attaining higher levels of transparency and interpretability. In this work, we propose an initial approach to neural-symbolic temporal decision trees, that is, an hybrid method that leverages on both the ability of neural networks of capturing temporal patterns and the flexibility of temporal decision trees of taking decisions on intervals based on (possibly, externally computed) temporal features. While based on a proof-of-concept implementation, in our experiments on public datasets, neural-symbolic temporal decision trees show promising results.
