Semi-supervised Multivariate Sequential Pattern Mining

Multivariate sequence analysis is of growing interest for learning on data with numerous correlated time-stamped sequences. It is characterized by correlations among dimensions of multivariate sequences and may not be separately analyzed as multiple independent univariate sequences. On the other hand, labeled data is usually expensive and difficult to obtain in many real-world applications. We present a graph-based semi-supervised learning framework for multivariate sequence classification. The framework explores the correlation within the multivariate sequences, and exploits additional information about the distribution of both labeled and unlabeled data to provide better predictive performance. We also develop an efficient method to extend the graph-based learning approach to out-of-sample prediction. We demonstrate the effectiveness of our approach on real-world multivariate sequence datasets from three domains.