Semi-supervised Multi-label k-Nearest Neighbors Classification Algorithms

Abstract

Classification is one of the most traditional tasks in machine learning. In supervised learning for classification, the goal is to learn a classifier function using a completely labeled dataset. Semi-supervised learning modifies the learning algorithm function allowing the use of partially labeled data. Single-label classification assigns only one label to each instance in the dataset, while multi-label classification can assign multiple labels for each instance. It would be relevant to develop techniques that are both multi-label and semi-supervised. However, few previous work has been devoted to semi-supervised multi-label classification. In the current work, we propose two new algorithms by extending the Multi-label k-Nearest Neighbors (MLkNN) algorithm to semi-supervised learning. The original MLkNN is a graph-based supervised algorithm. In our proposal, we augmented the graph structure and adapted two semi-supervised algorithms, label propagation and label spreading, for performing the label expansion in the augmented graph. We compare the proposed algorithms with a group of baseline supervised multi-label algorithms. The results for the metrics analyzed showed that the new algorithms were suitable for the multi-label semi-supervised scenarios.