Learning Human Activity Patterns using Clustered Point Processes with Active and Inactive States

TL;DR

This paper introduces a novel clustered point process model with active and inactive states, offering a flexible and interpretable way to analyze heterogeneous event patterns like human activities and social media behaviors.

Contribution

It proposes a new class of clustered point processes with a composite likelihood and EM algorithm for efficient parameter estimation, addressing heterogeneity in event data.

Findings

Applied to Donald Trump's Twitter data to analyze behavioral changes.

Analyzed Sina Weibo data to identify user groups with distinct behaviors.

Demonstrated the model's effectiveness in capturing activity pattern heterogeneity.

Abstract

Modeling event patterns is a central task in a wide range of disciplines. In applications such as studying human activity patterns, events often arrive clustered with sporadic and long periods of inactivity. Such heterogeneity in event patterns poses challenges for existing point process models. In this article, we propose a new class of clustered point processes that alternate between active and inactive states. The proposed model is flexible, highly interpretable, and can provide useful insights into event patterns. A composite likelihood approach and a composite EM estimation procedure are developed for efficient and numerically stable parameter estimation. We study both the computational and statistical properties of the estimator including convergence, consistency, and asymptotic normality. The proposed method is applied to Donald Trump's Twitter data to investigate if and how his behaviors evolved before, during, and after the presidential campaign. Additionally, we analyze large-scale social media data from Sina Weibo and identify interesting groups of users with distinct behaviors.