Partial Label Learning Algorithms Based On Maximum Margin And Semi-supervised Learning

Partial label learning is an important weakly-supervised learning framework in machine learning.The training example in this framework is only associated with a set of candidate labels,among which only one label is the ground-truth label.The difficulty of the partial label learning problem is that the ground-truth label of the training example cannot be obtained,and other label in the candidate label set will cause some interference to the learning process.There are two strategies for solving the problem of partial label learning: disambiguation strategy and disambiguation-free strategy.The disambiguation strategy can be divided into average disambiguation and identification disambiguation.The partial label learning method based on disambiguation strategy is to find the ground-truth label from the candidate label set of each training example.For the purpose of disambiguation,the average disambiguation assigns the same weight to each of the candidate label of the example and learns the output of the model on each candidate label.In order to achieve the purpose of disambiguation,the identification disambiguation is to regard the ground-truth label of the example as a hidden variable,and the objective function with hidden variables is optimized by iterative update.The partial label learning method based on the disambiguation-free strategy processes the candidate label set of the training example as a whole,and uses the error correction output code to solve the multi-classification problem in the partial label learning.In this thesis,we focus on the research of the partial label learning algorithm based on the identification disambiguation and the partial label learning algorithm based on the disambiguation-free,which mainly include:(1)Based on the strategy of identification disambiguation,a new maximum margin based on partial label learning algorithm PL-MM is proposed in this thesis.The PL-MM algorithm improves the PL-SVM algorithm by using the margin between the candidate labels of the example as part of the model training.The PL-MM algorithm optimized the margin between the maximum output over the candidate label set and the maximum output over the non-candidate label set,and the margin between the output for the ground-truth label and that for any other candidate label.An improved sub-gradient Pegasos method was developed to optimize the proposed algorithm.(2)Based on the strategy of disambiguation-free,semi-supervised learning based on disambiguation-free partial label learning algorithm PL-S2 ECOC is proposed in this thesis.By constructing a non-redundant encoding matrix,the PL-S2 ECOC algorithm splits the multi-classification tasks in partial label learning into a series of two classification tasks.In the encoding stage,the PL-S2 ECOC algorithm constructs a non-redundant encoding matrix,so that the same or complementary column encoding does not exist in the encoding matrix,and the two classifiers obtained by the training are different.Then the PL-S2 ECOC algorithm uses the semi-supervised learner as the two classification learner,so that each of the two classifiers can make full use of the entire partial label training set.Finally,a trained semi-supervised learner is used to predict the category of the test example in the two classification problem.In the decoding stage,the PL-S2 ECOC algorithm uses the loss-weighted decoding strategy to decode the code words generated by the test example,and the category with the least loss is marked as the prediction label of the test example.Experimental results on the UCI datasets and the real-world datasets show the superiority of PL-MM and PL-S2 ECOC,compared with other partial label learning algorithms.