Latent semantic KNN algorithm for multi-label learning

Exploiting label structures or label correlations is an important issue in multi-label learning, because taking into account such structures when learning can lead to improved predictive performance and time complexity. In this paper, a multi-label lazy learning approach based on k-nearest neighbor and latent semantics is presented, which is called LsKNN. Firstly, latent semantic analysis is applied to discover some semantic correlations between instances and class labels and the semantic features of each training sample are obtained. Then for each unseen instance, its k-nearest neighbors in the latent semantic subspace are identified and finally its proper label set is determined by resembling the votes of neighbors. Meanwhile, a support vector machine based pruning strategy called SVM-LsKNN, is proposed to deal with the slow testing of LsKNN. Experiments on three multi-label sets show that LsKNN needs no training, but can achieve at least comparable performance with some state-of-art multi-label learning algorithms. Extra experiments also verify the testing efficiency of the pruning technique.