Date/Time
Date(s) - 05/02/2026
1:30 pm - 2:30 pm

Speaker: Kostas Banos (McMaster University)

Location: Hamilton Hall, Room 217

Title: Power-Law Non-Parametric Temporal Models for Continuous-Time Document Streams

Abstract: Modelling continuous-time document streams requires capturing both textual structure and temporal dynamics. The Dirichlet–Hawkes Process (DHP) combines Dirichlet process clustering with Hawkes processes, but its reliance on a Dirichlet prior induces exponentially decaying cluster sizes, which poorly reflects the heavy-tailed topic distributions observed in real-world corpora. We propose the Pitman–Yor–Hawkes Process (PYHP), which replaces the Dirichlet prior with a Pitman–Yor process, enabling power-law behaviour and improved modelling of both frequent and rare topics. We develop a sequential Monte Carlo algorithm for efficient online inference and evaluate the model on a corpus of 150,000 online news articles, demonstrating coherent topic discovery and strong temporal structure. We also introduce a hierarchical extension motivated by empirical limitations of the base model. This work presents a new family of power-law temporal topic models with practical relevance to streaming applications such as news monitoring and social media analysis.