Sparse Model-Space Learning for Multivariate Sequence Classification

The analysis of high-dimensional multivariate sequential data is pivotal across various fields. For efficient classification of such data, this paper presents Sparse Model-Space Learning (SMSL) based on Model-Space Learning (MSL) methods, which shift the analysis from data space to model space by representing each instance with a well-fitted model. By integrating reservoir computing and substituting ridge regression with lasso regression, each sequence is efficiently fitted with a sparse model solution. Afterward, the dimension of sparse models can be reduced through a reduction process. Thus, the original sequences are represented by the corresponding reduced models that fully capture the sequence-inherent dynamic characteristics. Finally, a Hilbert space is defined based on the fitted models, where classification methods can be utilized on these models instead of the original sequences. The proposed approach not only improves the efficiency of processing multivariate sequences but also minimizes memory usage and training resources. The effectiveness of the proposed method is validated by experimental results on the Tennessee Eastman Process (TEP) dataset, demonstrating its ability to manage multivariate sequential data.