Modeling Share Dynamics by Extracting Competition Structure

The document from 2004 introduces a probabilistic dynamical model for analyzing multivariate time-series data, with a focus on understanding the competitive dynamics among websites. The authors, Masahiro Kimura, Kazumi Saito, and Naonori Ueda, developed a learning algorithm that categorizes websites into competitive groups and predicts their future market shares. The model uses a replicator equation to represent the dynamics of website competition for users. The method's effectiveness was demonstrated through tests on synthetic data, where it accurately identified the true competition structure and outperformed conventional methods in prediction accuracy. When applied to real data from 20 streaming video websites, the method revealed a competition structure that conventional methods did not detect and provided superior predictions. The study highlights the significance of accounting for competitive dynamics when modeling the fluctuations in website traffic.