The multi-label classification problem indicates that multi-label samples possibly belong to different classes simultaneously, and labels may be overlapped or mutually exclusive. Multi-label classification algorithm’s performance partly depends on whether the parameters are optimal. We tune multi-label classification parameters based on a multi-objective evolutionary algorithm. This paper mainly focuses on parameter tuning of multi-label classification algorithm based multi-objective evolutionary algorithm.This paper conducts the correlation analysis and theory analysis between two evaluation criteria:Hamming loss and Ranking loss for the first time. We draw a conclusion that there’s a weak correlation and repellency between the two evaluation criteria mentioned above. In addition, we use the evaluation criteria as the object functions of multi-object genetic algorithm NSGA-II to research the parameter tuning problem. Depending on the mechanism of NSGA-II, which is able to randomly search solution set and skip the local optimization, we obtain a global optimum solutions efficiently. Therefore, our algorithms can improve the performance of the traditional strategy of parameter search, and reduce the parameter searching time as well.This paper studies the parameter optimization problems for three multi-label classification algorithms:ML-kNN, OVR-kNN, and OVR-SVM. Multi-label classification criteria are utilized in our experiments on ten datasets:Image, Scene, Emotions, Yeast, Genbase, Human, Plant, Langlog, Enron and Medical. In our experiments, firstly, we research the relationship between the performances and parameters, and then indicate the parameter optimization process through some figures. Secondly, we compare our multi-objective evolutionary parameter tuning algorithm with another parameter tuning algorithm:grid search. In the end, we draw a conclusion that our algorithms give good performance, and outperform grid search in running time. |