Self-organized Hawkes Processes

Shen Yuan, Hongteng Xu

Abstract: In this paper, we propose a novel self-organized Hawkes process (SOHP) to model complex event sequences based on extremely few observations. Motivated by the fact that the complicated global relations among events are often composed of simple local relations, we model the event sequences by a set of heterogeneous local Hawkes processes rather than a single Hawkes process. In the training phase, we learn the Hawkes processes with a self-organization mechanism, selecting training sequences adaptively for each Hawkes process by a bandit algorithm. The reward used in the algorithm is originally defined based on an optimal transport distance. Additionally, we leverage the superposition property of the Hawkes process to enhance the robustness of our algorithm to the data sparsity problem. We apply our SOHP method to sequential recommendation problems in the continuous-time domain and achieve encouraging performance in various datasets. The code is available at https://github.com/UESTC-DaShenZi/MHP .