Exploring Bayesian Models for Multi-level Clustering of Hierarchically Grouped Sequential Data

TL;DR

This paper introduces a generalized Bayesian hierarchical clustering model that accounts for sequential and temporal data structures, providing a unified framework and demonstrating improved performance in hierarchical segmentation tasks.

Contribution

The paper proposes a new generalized hierarchical Bayesian model with a novel Degree of Sharing concept, unifying existing models and incorporating sequential data analysis.

Findings

Model outperforms existing methods in hierarchical segmentation of news transcripts.

Introduces Degree of Sharing to analyze and classify hierarchical Bayesian models.

Provides a Gibbs Sampling inference algorithm for the proposed model.

Abstract

A wide range of Bayesian models have been proposed for data that is divided hierarchically into groups. These models aim to cluster the data at different levels of grouping, by assigning a mixture component to each datapoint, and a mixture distribution to each group. Multi-level clustering is facilitated by the sharing of these components and distributions by the groups. In this paper, we introduce the concept of Degree of Sharing (DoS) for the mixture components and distributions, with an aim to analyze and classify various existing models. Next we introduce a generalized hierarchical Bayesian model, of which the existing models can be shown to be special cases. Unlike most of these models, our model takes into account the sequential nature of the data, and various other temporal structures at different levels while assigning mixture components and distributions. We show one specialization of this model aimed at hierarchical segmentation of news transcripts, and present a Gibbs Sampling based inference algorithm for it. We also show experimentally that the proposed model outperforms existing models for the same task.