Multi-label classification using non-iterative learning and deep learning based approaches

Abstract

Multi-label classification is a generalization of traditional classification techniques where classes are not mutually independent and multiple classes are assigned to each instance. Multi-label classification is applied in various domains such as image classification, functional annotation of protein sequences, and text categorization. The main objective of this thesis is to propose multi-label classifiers using non-iterative learning and deep learning approaches. We propose non-iterative multi-label classifiers with an adaptive threshold to consider the correlation among labels. The non-iterative approaches provide a fast and accurate solution for multi-label classification. We propose non-iterative randomization-based neural networks for multi-label classification. These multi-label neural networks are named as Multi-label Broad Learning System (ML-BLS), Multi-label Fuzzy Broad Learning System (ML-FBLS), Multi-label Random Vector Functional Link Network (ML-RVFL), and Multi-label Kernelized Random Vector Functional Link Network (ML-KRVFL). The output weights of these neural networks are computed using pseudoinverse. At the output layer, multi-label classification is performed by using an adaptive threshold function. The computation of output weights using pseudoinverse retains the faster computation power of these algorithms compared to iterative learning algorithms. The adaptive threshold function used in the proposed approach can consider the correlation among the output labels and the whole dataset for threshold computation. Five multi-label evaluation metrics, hamming loss, ranking loss, one error, coverage, and average precision, evaluate the proposed multi-label neural networks on 12 benchmark datasets of various domains such as text, image, and genomics. The ML-KRVFL provides the overall best Friedman rankings on five evaluation metrics, followed by ML-RVFL, ML-FBLS, and ML-BLS, respectively. Based on the experimentation results, the proposed ML-BLS, ML-FBLS, ML-RVFL, and ML-KRVFL perform better than other relevant multi-label approaches in this mentioned order. The non-iterative approaches use the inverse to compute the parameters of algorithms, and it is cumbersome to compute the inverse for large data.