Research Of Multi-label Feature Selection Algorithms In The Form Of Nonlinear Programming

In multi-label problem,an instance is not only related to multiple labels,but also often contains high-dimensional features.Some of these features are redundant or even irrelevant.Their existence reduces the performance of classifiers and increases the memory footprint Multi-label feature selection technology has become the most mainstream method to solve the above problems because it can select the most relevant original features.In this thesis,two multi-label feature selection algorithms are proposed on the basis of QIPcmi,a formal feature subset selection framework for non-linear programming based on conditional mutual information:(1)multi-label feature selection algorithm based on conditional mutual information combined with genetic algorithm;(2)quadratic programming multi-label feature selection algorithm based on normalized conditional cross-covariance operator.For the multi-label feature selection algorithm based on conditional mutual information combined with genetic algorithm,we first discretize the features to meet the requirements of conditional mutual information and mutual information.Then we use genetic algorithm with control strategy to solve QIPcmi.In the experiment,we compare the proposed algorithm with four existing multi-label feature selection algorithms on six benchmark multi-label datasets.Experiments show that our proposed multi-label feature selection algorithm based on conditional mutual information combined with genetic algorithm can select a better feature subset.For quadratic programming multi-label feature selection algorithm based on normalized conditional cross-covariance operator,we first introduce normalized conditional cross-covariance operator and normalized cross-covariance operator,which can measure conditional dependency and dependency respectively.In this thesis,we use them to replace conditional mutual information and mutual information in QIPcmi respectively,and form a feature subset selection method based on normalized conditional cross-covariance operator in the form of non-linear programming.Then we relax the constraints and transform the original NP-hard problem into a quadratic programming problem to solve.In the experiment,we compare this algorithm with the first one proposed in this paper and two other existing algorithms on five benchmark multi-label datasets.Experiment analysis shows that our quadratic programming multi-label feature selection algorithm based on normalized conditional cross-covariance operator has better classification effect in experiments.