Classifying biomedical knowledge in PubMed using multi-label vector machines with weaker optimization constraints

In this paper, we have developed an automated multi-label linking scheme for PubMed citations with gene ontology (GO) terms, which enables users to have easy access to relevant publications according to various biomedical ontological terms (in particular, GO terms). We propose a maximum margin approach derived from ranking support vector machine (Rank-SVM), called SCRank-SVM. In this scheme, we remove the term bias “b” and recast the decision boundary and the separating margin to improve the margin of Rank-SVM. Due to the weaker optimization constraints, SCRank-SVM has better generalization performance and lower computational complexity. Experiments on our lung cancer data set and 6 diverse multi-label data sets show that SCRank-SVM is quite suitable to solve our problem. The performance of SCRank-SVM is superior to that of the original Rank-SVM and some other well-established multi-label learning algorithms.