Adaptive Generalized Multi-View Canonical Correlation Analysis for Incrementally Update Multiblock Data

One of the major problems in real-life multiblock dynamic data analysis is that all the available modalities may not be relevant. Some of them may provide noisy or even inconsistent information with respect to other modalities. So, it is necessary to evaluate the quality of a new modality before considering it for feature extraction. In this regard, the paper introduces a new multiset canonical correlation analysis (MCCA), termed as incremental MCCA (IMCCA). When a new modality is available for the analysis, the IMCCA generates the new canonical variables from that of the earlier modalities, without repeating the same procedure with the original data augmented by the new modality. The proposed IMCCA deals with the “curse of dimensionality” problem associated with multidimensional data sets, by using the ridge regression optimization technique. Using the proposed IMCCA model, a new feature extraction algorithm is introduced, which considers a new modality for the analysis if it has relevant and significant information with respect to existing modalities. The proposed algorithm starts with the two most relevant modalities, and the remaining modalities are added sequentially according to their relevance. The optimum regularization parameters for the proposed algorithm are estimated based on the supervised information of sample categories. The effectiveness of the proposed algorithm, along with a comparison with state-of-the-art multimodal data integration methods, is established on several real-life multiblock data sets.