Research On Ordinal Regression With Uncertain Multi-view Data

Ordinal regression(OR)is a paradigm which learns a prediction model on the data with ordered classes.Despite much progress in OR,the existing OR works learn the classifier from only one view and the multi-view learning in OR has not been considered.What is more,due to sampling errors,transmission errors,etc.,there may exist uncertain information in multi-view data.How to use the multi-view information in the ordered data and deal with the uncertain information in the ordered data is an urgent problem to be solved in this thesis.According to the limitations of the above OR problems,this thesis proposes ordinal regression for single-view uncertain data and multi-view support vector ordinal regression with uncertain data,which can improve the OR classifier by handling with the uncertainty information in the sample.The first algorithm aims at dealing with uncertain information data in ordinal regression.Based on the principle of maximum interval,support vector ordinal regression for uncertain data is constructed.The model can reduce the impact of uncertain data on the decision boundary and enhance the robustness of the model.The second algorithm is based on the ordinal regression for single-view uncertain data.The classification accuracy of the model is further improved by integrating multi-view data.Then,the algorithm adopts a heuristic framework to solve the OR learning problem.The main research contents of this thesis can be viewed in three aspects:(1)The existing OR methods mainly learn the classifier on deterministic data,and there is little work done on uncertain multi-view data.In this thesis,the OR algorithm for single-view uncertain data is proposed,which enriches the research work in this area.(2)This thesis presents a maximum-margin model and utilizes it to deal with the OR problem with uncertain multi-view data.OR for multi-view uncertain data is capable of improving the OR classifier by incorporating the multi-view information and handling the data uncertainty.(3)Extensive experiments are conducted on real-world data sets,and the mean zero-one error and mean absolute error are reported.The experiments have shown that the performance of MORU is explicitly better than the existing OR models.