On Several Issues Of Label Distribution Learning

How to solve the problem with label ambiguity is a popular research direction in the field of machine learning. Currently, the mature learning paradigm which can solve these problems is multi-label learning. Multi-label learning which allows an instance to associate multi labels is an expansion of single-label learning. Hence it is more suitable to solve the problem with label ambiguity than single-label learning. But there are still some problems which are not suitable to use multi-label learning to solve directly. Therefore, in order to address these problems with label ambiguity better and more directly, label distribution learning was put out. Label distribution learning is an expansion of multi-label learning. In theory, label distribution learning has more application scenarios.Therefore, further study of label distribution learning will be possible to make more problems with label ambiguity to be better solved. For this purpose, this paper will research some issues of label distribution learning in the following order.1. Same as multi-label learning, we can not only use a single criterion to evaluate the performance of prediction of a label distribution learning algorithm, it needs to develop a set of evaluation criterions to evaluate an algorithm from different aspects. Hence, we use a strategy to choose a set of evaluation criterions with diversity and representativeness based on the syntax and semantics relationship between different distances/similarities. From the theory and experimental results, these evaluation criterions can evaluate the prediction performance of an algorithm from different aspects correctly.2. Currently, there are three common strategy of designing label distribution learning algorithm. They are problem transformation strategy, algorithm adaptation strategy and specialized algorithms strategy. The algorithm using different design strategy may have different performance of prediction. Analyzing these differences will promote the development of algorithm designing of label distribution learning. Therefore, this paper launches a comparative study of three strategies based on some datasets of label distribution learning. Experimental results and theoretical analysis show that the specialized algorithms strategy is a move effective design strategy.3. In order to improve this design strategy, this paper launches a study on this strategy. After analyzing the characteristics of this strategy, this paper sums up a design framework of specialized algorithms. Then we study the target function part in this framework, and make some proposals on how to select the target function by experimental results and theoretical analysis.