Research On Machine Learning Methods By Using Implicit Constraints

In many machine learning tasks, some additional constraints, named "implicit constraints" can be gathered during the learning process or from the learning environment. As the advances of technologies, it has become much easier to cumulate a huge number of data and the data environment has become more and more complicated. Therefore, how to efficiently utilize these implicit constraints to improve the efficiency and generalization performance has attracted much attention during the past few years. Particularly, when there are few labeled data, which is very common in real-world applications, implicit constraints can be much more helpful.This thesis tries to use implicit constraints to improve the learning performance when there are few labeled data, and makes several contributions summarized as follows:First, propose a new variant of co-training method named CoSnT. This method is inspired by the real-world teaching-learning system to incorporate both the teacher’s teaching confidence and the learner’s learning need. As a result, by using the implicit constraint of teaching-learning constraint within the learning process, the interactions between those two learners within co-training are considered in CoSnT. Experiments validate CoSnT’s superiority of learning performance compared to standard co-training method.Second, propose a new pairwised specific distance metric learning method named PSD. PSD utilizes the structure information of linkage constrains, which are implicit constraints gathered from the learning environment, and tries to distinguish different semantic meanings possessed by different linkages between different instance pairs. So each instance pair’s distance is measured separately. Experiments on both multi-class and multi-label datasets validate more appropriate distances learned from PSD, and better classification performance facilitated by them.