Uncovering Behavioural Patterns in Multivariate Time Series through Latent Space Analysis

Abstract

Multivariate time series (MTS) data, consisting of multiple time series observed concurrently across multiple variables, has become increasingly common in a variety of fields ranging from finance to healthcare. However, due to its complex nature, analysing and extracting meaningful insights from it remains a challenging task. This project aims to address the challenge of extracting valuable insights from high-complexity data by combining machine learning with visualisation tools, with a particular focus on latent space analysis. The framework is applied to a general energy sustainability project to understand how behavioural interventions affect energy consumption. We evaluate desirable features of the framework by conducting testing and comparing the model’s performance using various combinations of parameters. Additionally, we apply the framework to real use case scenarios, demonstrating its practical applicability and effectiveness. In conclusion, our findings indicate that the framework effectively facilitates the exploration of MTS data. Latent space analysis proves valuable in providing deeper insights into the data, allowing us to investigate the effects of sociodemographic factors, detect anomalies, and uncover behavioural patterns.