State-Space Dynamics Distance for Clustering Sequential Data

TL;DR

This paper introduces a new similarity measure for clustering sequential data by constructing a shared state-space and comparing transition matrices, improving over existing methods in scalability and overfitting.

Contribution

It presents a novel state-space based distance measure that reduces overfitting and enhances scalability in sequence clustering tasks.

Findings

Effective on synthetic datasets

Outperforms existing methods in real-world data

Reduces overfitting and improves scalability

Abstract

This paper proposes a novel similarity measure for clustering sequential data. We first construct a common state-space by training a single probabilistic model with all the sequences in order to get a unified representation for the dataset. Then, distances are obtained attending to the transition matrices induced by each sequence in that state-space. This approach solves some of the usual overfitting and scalability issues of the existing semi-parametric techniques, that rely on training a model for each sequence. Empirical studies on both synthetic and real-world datasets illustrate the advantages of the proposed similarity measure for clustering sequences.