Research On Fast Multi-label Learning Algorithms Via Hashing Transformation

In traditional supervised learning, each instance is associated with only one label. However, in many real world applications, each instance maybe associated with multiple labels simultaneously. The framework to solve such problem is referred to multi label learning(MLL). During the last several years, many MLL algorithms have been proposed and achieves excellent performance in multiple applications. However, these approaches are usually time-consuming and cannot handle large-scale data. With the advent of the era of big data, further investigating the fast multi label learning algorithm has great realistic significance.Firstly, the basic concepts of MLL and hashing learning are elaborated in this thesis. The thesis summarizes some related MLL algorithms and analyzes their strengths and weaknesses. Then, a fast multi-label learning algorithm based on hashing schemes, named HashMLL, is proposed. This algorithm exploits locality sensitive hashing to accelerate the k nearest neighbor algorithm, and thus can improve the efficiency of MLL. Moreover, due to the approximation of hash learning, so we also take the label correlation calculated by MinHash into account to improve the performance of HashMLL. Moreover, the HashMLL framework has a strong generalization ability, which can be extended by a variety of methods to enchance.This thesis verifies the performance of the HashMLL algorithm on the various published data sets in the multi-label field, and the extracted metagenomic data sets respectively. Experimental results shows that HashMLL algorithm can simultaneously perform a better accuracy and higher learning efficiency. Moreover, compared to the previous MLL algorithms, HashMLL algorithm can guarantee a certain accuracy, and achive a much faster leaming speed. HashMLL can achive a perfect compromise between accuracy and efficiency.