MM-CPred: A Multi-task Predictive Model for Continuous-Time Event Sequences with Mixture Learning Losses

Sequence prediction is a well-defined problem with a proliferation of applications, such as recommendation systems, social media monitor, economic analysis, etc. Recently, RNN-based methodologies have shown their superiority in time-series data analysis and sequence modeling. The question of which event would happen next is not difficult to answer anymore, but the prediction of when it would happen is still a mountain to climb. In this paper, we propose a Multi-task model to predict both event and their continuous timestamps at the same time. Specifically, (1) we design a two-layer RNN encoder for event sequences and a CNN encoder for time sequences, both equipped with multi-head self-attention to align history features; (2) we form multiple generative adversarial models for predicting future time sequences to solve the problem of multi-modal time distribution; (3) Mixture learning losses are adopted to conduct a 3-step learning strategy for training our model, the cross-entropy loss for events, Huber loss and adversarial classification loss which induces the Wasserstein distance for times. Due to these characteristics, we name it MM-CPred. The experiments on 4 real-life datasets confirmed its improvements compared with the baselines.