A Multi-label Classification Algorithm Based On Evolutionary Multi-objective Optimization

According to the number of labels in the sample sets,the classification problems can be divided into single label and multi-label classification problems.Conventional multi-label classification approaches mainly optimize a single performance criterion or a heuristic function.But actually the trade-offs among multiple inconsistent objectives should be considered in the optimization of multi-label classification,so this thesis mainly studies multi-label classification algorithm based on evolutionary multi-objective optimization.To solve the problems of the poor distribution of population and negative impact of the overlapping individuals in Non-dominated Sorting Genetic Algorithm(NSGA-?),a genetic algorithm based on vector space model(VSGA)is proposed.VSGA introduces vector space model into the problems of evolutionary multi-objective optimization,forming a kind of crowding mechanism in which included angle between target weighted vectors is regarded as distribution estimation,and eliminating the overlapping individuals.The simulation experiment shows,compared with NSGA-?,Pareto fronts obtained from VSGA are smoother and distributivity of non-domination solution is much better.A novel multi-label classification method ML-EMO based on multi-objective optimization is proposed to solve the problems,that the objectives are inconsistent or even conflicting in the multi-label classification.The method randomly generates a set of ML-RBF classification models with different scale coefficient firstly,and chooses the evaluation metrics of classification as optimization objects.Then the above algorithm VSGA is used to optimize these models and generates a group of optimal models,which can be used to classify the multi-label data.At last,the final classification results are got by majority vote.ML-EMO can flexibly construct various predictive models according to the optimal models,and this algorithm provides more meaningful classification results in different application scenarios.The simulation experiment shows,comparing with four representative multi-label classification algorithms,such as ML-RBF?BP-MLL?ML-KNN and Ecc,ML-EMO dose provide a better rank on seven common evaluation metrics.