Research On Robust Graph-based Learning Methods For Relationship Exploitation

Relationships are ubiquitous in real world.There are two kinds of relationships among data in the area of machine learning research:1)relationships between exam-ples;2)relationships between labels.According to abundant results of research works,appropriately these two kinds of relationships is the key to improving the predictive performance of trained model.Graph-based learning is a popular paradigm of methods that exploit relationships.The representative work in this area won the best 10-year paper award of International Conference on Machine Learning.After decades of research,many achievements of graph-based methods have been made.However,the learning performance depends heavily on the problem of graph construction.In practical tasks,it is difficult to effec-tively determine the graph construction,making learning performance less robust and sometimes performance damaging.This dissertation focuses on the important issue of improving the robustness of relationship exploitation,and the main innovative achievements are as follows:(1)In terms of the issue that the example relationship exploitation is sensitive to the graph construction,the method of graph quality judgment based on large margin criterion is proposed.In this method,the problem of robust example relationship ex-ploitation is formalized into the framework of the classical semi-supervised support vector machine(S4VM).An efficient algorithm for optimization is given.Experimen-tal results show that the proposed method significantly improves the robustness of the example relationship exploitation and effectively avoids the performance degeneration of the traditional methods.In this dissertation,the large margin criterion is further extended to the noisy example relationships,and an efficient learning algorithm is pro-posed to prevent the hurt of noisy example relationships to the performance.(2)In terms of the issue that the label relationship exploitation is sensitive to the graph construction,the method of label relationship exploitation based on classifier circle is proposed.By constructing the classifiers in the form of cycles,the proposed method overcomes the serious effect of the label order in label relationship exploita-tion on the performance of traditional learning methods.The time complexity of this method is equivalent to that of the traditional methods,and the computational cost is not significantly increased.The experimental results show that the proposed method can significantly improve the robustness of label relationship exploitation and avoid the poor performance may happen on the traditional methods of label relationship ex-ploitation.