Semi-Supervised Multi-Label Learning Based On Label Propagation

Weak label learning is becoming a research focus in the field of machine learning in recent years. The precondition of better performance in the traditional multi-label learning is the sheer volume of training data and the full label out there. However, training data usually be small size and incomplete label which may be due to large scale, variety label and heavy work in the practical application. Given a partial or incomplete label, a complete label set prediction(i.e., the missing label) problem is defined as the problem of weak label learning". Through introducing label propagation algorithm, semi-supervised learning in the small size labeled data, using large number of unlabeled data to improve the algorithm performance respond to the actual application. But semi-supervised learning is insufficient when dealing with weak label data sets. We improve the method based on graph, making it suitable for weak label data sets. The experiments on open multiple group multi-label data sets, verified the validity of the improved algorithm. The main work is as follows:1. Each sample in weak label data corresponding to more than one label. And it is very important to each sample data that the similarity measure is relation to the degree of label completion of weak sample data. Semi-supervised method based on graph needs to structure the graph before, and the extent of label propagation was determined by edge weight. It is spatially limited that the measurement method of edge weight was the Euclidean metric between the samples. For this purpose, we construct a graph based on k-means for weak label data sets. The composition method is used to determine the similarity of samples by the whole distribution of samples and the distance between samples.2. In order to deal with classification of weak label data sets, the multi-label learning and based on graph learning method is combined. And an effective label propagation based on weak label algorithm is put forward by improved the existing one. Not only labels of unlabeled data to be filled, missing labels of labeled data also need to complement due to the weakly labeled data set characteristics. The improved method populates sample labels by label propagation, and ensures better utilization of weak label data to improve the algorithm effect.The comparison experiments on multiple data sets indicate the effectiveness of the proposed method in dealing with weak label datasets. Simultaneously, the experiments on datasets under different label rates analysis the effect label rate bring in weak label data classification.