A Case-Based Reasoning and Explaining Model for Temporal Point Process

Event sequence data widely exists in real life, where each event can be typically represented as a tuple, event type and occurrence time. Combined with deep learning, temporal point process (TPP) has gained a lot of success for event forecasting. However, the black-box nature of neural networks makes them lack transparency for their forecasting decision. In this paper, we introduce case-based reasoning (CBR) into the modeling of temporal point process, yielding CBR-TPP. CBR is in line with human intuition and can provide explanation cases for decision-makers. Not using traditional similarity metrics (e.g., edit distance), we propose to employ the type-aware attention mechanism to retrieve the explanation cases as well as for cased-based reasoning. Experimental results on six datasets show that CBR-TPP outperforms existing TPP models on event forecasting task under both extrapolation and interpolation setting. Moreover, the results highlight the generalization ability and interpretability of our proposed model.