An evidence accumulation based block diagonal cluster model for intent recognition from EEG. (August 2022)

• Record Type: Journal Article
• Authors: Fu, Rongrong; Li, Zheyu
• Is Part Of: Biomedical signal processing and control. Volume 77(2022)
• Publication Date: 2022-08
• Subjects: Electroencephalogram (EEG) -- Clustering -- Evidence accumulation (EA) -- Block-diagonal model; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103835
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A novel block clustering model that combines CE and probabilistic hybrid model is developed for clustering EEG. A similarity matrix is constructed at the sample level, reducing the workload of features processing. The approach takes full advantage of the similarity measure between samples and mitigates the effects of some unreliable features. The similarity within the class is reflected in the block-diagonal structure, which provides an explanation for the final clustering results. Abstract: Most of the probabilistic mixture models perform clustering by observing the eigenvectors of the data sample and these models rely on the layout of features. Clustering ensemble based on similarity matrices avoids complex processing of samples by only accessing basic clusters. However, while there are many literatures on the probability mixture model for clustering, there is almost no study focusing on applying the similarity matrix to the probability mixture model. Therefore, a new clustering method called block clustering structure of evidence accumulation matrix (BEAM) is proposed in this study by combining the clustering ensemble and the probability mixture model. Specifically, evidence accumulation (EA) is developed to obtain a similarity matrix of samples. The interpretability of the similarity matrix can be improved due to sample-based similarity measures, and then the diagonal block model is designed to identify representative block cluster structures from the similarity matrix. The proposed method has been evaluated on the BCI Competition IV Data set 1 and the block-diagonal structure of the similarity matrix is discovered, which ensures high similarity within the same cluster and large separation between the clusters. In addition, the Davies-Bouldin index (DBI) and adjusted rand index (ARI) are used to evaluate BEAM performance. The results show that the proposed method is superior to the state-of-the-art approaches.