oa Kernel collaborative label power set system for multi-label classification

A traditional multi-class classification system assigns each example x a single label l from a set of disjoint labels L. However, in many modern applications such as text classification, image/video categorization, music categorization etc [1, 2], each instance can be assigned to a subset of labels Y ⊆ L. In text classification, news document can cover several topics such as the name of movie, box office ratings, and/or critic reviews. In image/video categorization, multiple objects can appear in the same image/video. This problem is known as multi-label learning. Figures 1 shows some examples of the multi-label images. Collaborative Representation with regularized least square (CRC-RLS) is a state-of-the-art face recognition method that exploits this collaborative representation between classes in representing the query sample [3]. The basic idea is to code the testing sample over a dictionary, and then classify it based on the coding vector. While the benefits of collaborative representation are becoming well established for face recognition or in general multi-class classification, the corresponding use for multi-label classification needs to be investigated. In this research, a kernel collaborative label power set multi-label classifier (ML-KCR) based on regularized least square principle is proposed. ML-KCR directly introduces the discriminative information of the samples using l2-norm\sparsity" and uses the class specified representation residual for classification. Further, in order to capture co-relation among classes, the multi-label problem is transformed using label power set which is based on the concept of handling sets of labels as single labels and thus allowing the classification process to inherently take into account the correlations between labels. The proposed approach is applied to six publicly available multi-label data sets from different domains using 5 different multi-label classification measures. We validate the advocated approach experimentally and demonstrate that it yields significant performance gains when compared with the state-of-the art multi-label methods. In summary, following are our main contributions * A kernel collaborative label powerset classifier (ML-KCR) based on regularized least square principle is proposed for multi-label classification. ML-KCR directly introduces the discriminative information and aim to maximize the margins between the samples of di_erent classes in each local area. * In order to capture correlation among labels, the multi-label problem is transformed using label powerset (LP). The main disadvantage associated with LP is the complexity issue arise due to many distinct label sets. We will show that this complexity issue can be avoided using collaborative representation with regularization. * We applied the proposed approach to publicly available multi-label data sets and compared with state-of-the-art multi-label methods. The proposed EML method is compared with the state-of-the-art multi-label classifiers: RAkEL, ECC, CLR, MLkNN, IBLR [2]. References [1] Tsoumakas, G., Katakis, I., Vlahavas, I., 2009. Data Mining and Knowledge Discovery Handbook. Springer, 2nd Edition, Ch. Mining Multilabel Data. [2] Zhang, M.-L., Zhou, Z.-H., 2013. A review on multi-label learning algorithms. IEEE Transactions on Knowledge and Data Engineering (preprint). [3] Zhang, L., Yang, M., 2011. Sparse representation or collaborative representation: Which helps face recognition? In: IEEE International Conference on Computer Vision (ICCV).