A novel multi-view learning developed from single-view patterns

Globalized and localized matrix-pattern-oriented classification machine

“…The same matrixized reshaping strategy without overlapping among the components of the pattern is adopted in our experiments. In practice, the original image matrix or vector pattern is partitioned into many sub-vectors, and then re-arranged column-by-column into the corresponding matrix following the same way used in the literature [26], e.g., the original dimension for BCW is 10, which means that each sample of BCW is a ten-dimension vector. After the processing of matrixized reshaping, there are four new forms of each sample created, including 2 × 5, 5 × 2, 1 × 10, and 10 × 1.…”

“…(27), it can be found that: (1) both GLMatMHKS and MatMHKS share the same decision function; (2) it has been proved that the solution space for the weights in MatMHKS is contained in that of MHKS [26]. Thus MatMHKS is viewed as an MHKS imposed by Kronecker product decomposability constraint [26]. Therefore, the solution sets of GLMatMHKS, MatMHKS, and MHKS have the following relationship:…”

“…Thirdly, both the vector-and matrix-based structural learning for classifier design gain their popularity in recent decade. After discussing both the superiority and inferiority between the Single View Learning (SVL) and the Multiple View Learning (MVL), Wang et al [26] find the difficulty in getting multi-view information from single view pattern in reality and thus propose a novel MVL based on MHKS named MultiV-MHKS. On the other side, Mehmet et al [10] try to introduce the localized information into multi-kernel learning based on the Support Vector Machine (SVM).…”

Double-fold localized multiple matrixized learning machine

“…For example, AdaMatLSSVC [35] is an ensemble-based strategy constructed by multiple MatLSSVCs with different reshaping ways. In addition, both the multi-view learning developed from single-view patterns with Ho-Kashyap linear classification strategy (MultiV-MHKS) [37] and the regularized multi-view machine based on the response surface technique (RMultiV-MHKS) [38] fuse multiple matrix representations into a joint learning machine, respectively. In the joint learning machine, the classifier derived from a corresponding matrix representation is named as a matrixized sub-classifier.…”

“…But MatMHKS is a single-matrixized learning machine which is based on one matrix representation. It has been validated that the multiple-matrixized learning machine has a superior performance than the corresponding single one [37,38].…”