Title: Multistate model for trajectories clustering
Authors: Rossella Miglio - Bologna University (Italy) [presenting]
Abstract: Clustering of temporal or sequential data is challenging, especially when dealing with discrete data. The motivating problem is to find patterns of drug use trajectories over time. It is essential to have standard measures of change, define appropriate similarities among trajectory observations, obtain appropriate data representation and use methods that are suitable for this kind of data or using information extracted from them to apply classical methods. We analysed data across 5 years on a sample of 70000 drug users to identify transition patterns in drug use trajectories during this time. Data were collected every three months for a total of 20 measurements for each subject, demographic and some clinical covariates are available. Optimal matching and a three-step procedure proposed previously to identify clusters of individual longitudinal trajectories were used in the preliminary analysis. We propose to address the problem of unsupervised classification of sequences using a multi-state approach, to obtain measures to quantify the change in drug use behaviours; these approaches allow us to consider also the effects of covariates. These set of measures could be used as input for classical clustering methods but could also help to provide effective visualizations for applied use. The results obtained by the other proposed method are compared with this new proposal.