Two-Stage Multilayer Perceptron Hawkes Process

Abstract

Many social activities can be described as asynchronous discrete event sequences, such as traffic accidents, medical care, financial transactions, social networks and violent crimes, how to predict the event occurrence probabilities, times of occurrence and types of events is a challenging and upmost important problem. It has broad application prospects in urban management, traffic optimization and other fields. Hawkes processes are used to simulate complex sequences of events. Recently, in order to expand the capacity of Hawkes process, neural Hawkes process (NHP) and transformer Hawkes process (THP) were proposed. We argue that the complexity of the model is high due to the introduction of recurrent neural networks or attention mechanisms. While the attention mechanism can achieve good performance, it is not essential. Therefore, in this paper, we propose a Two-stage Multilayer Perceptron Hawkes Process (TMPHP). The model consists of two types of multilayer perceptrons: one that applies MLPs (learning features of each event sequence to capture long-term dependencies between different events) independently for each event sequence, and one that applies MLPs to different event sequences MLP (capturing long-term and short-term dependencies between different events). Our model is simpler than other state-of-the-art models, but it has better predictive performance. Especially for MIMI data sets, our model outperforms RMTPP (4.2%), NHP (2.2%) and THP (2.2%) in terms of prediction accuracies.