Feature Selection Method Based On Multi-objective Cuckoo Search Algorithm

Feature selection aims to improve the efficiency of the model by minimizing the number of subset features,while ensuring that the selected subset features have good classification accuracy for the model.Therefore,feature selection can reduce the feature space,decrease the computational overhead of the model and reduce the risk of dimensional catastrophe,while avoiding the occurrence of overfitting.However,there is a correlation and conflict between the two objectives of classification accuracy and the number of feature subsets.In this paper,we address these issues by a multi-objective optimization-based approach and an improved cuckoo algorithm.The specific feature selection problem is constructed by using multi-objective optimization theory and solved by applying a swarm intelligence optimization algorithm.The obtained results form the Pareto frontier,which provides multiple alternative feasible solutions to the multi-objective feature selection problem.An improved binary multi-objective cuckoo search algorithm is proposed to solve the constructed multi-objective feature selection problem under the conventional data set.The algorithm proposes a full-coverage initialization strategy based on the normal distribution,which takes the normal distribution as the probability basis for initialization and effectively ensures the uniform distribution of individuals in the solution space;the algorithm further introduces an iterative update based on the elite strategy,which takes the parent and child generations as the overall ordering to avoid the discarding of outstanding individuals in the iterative process,accelerates the convergence of the algorithm and improves the stability of the algorithm;the algorithm introduces a The algorithm introduces the advantage inheritance operator based on logical operations,which finds the advantageous features as much as possible through logical operations and has the probability to make the next generation of individuals to inherit,avoiding the directionlessness of the algorithm in individual iteration and improving the stability and efficiency of the algorithm.Simulation experiments using regular datasets from the UCI machine learning repository demonstrate the stronger performance of the proposed IBMOCS algorithm,and also validate the effectiveness of the three operators introduced in IBMOCS.An improved binary multi-objective quantum cuckoo search algorithm is proposed to solve the multi-objective feature selection problem in ultra-high-dimensional datasets.The algorithm introduces a Q-OBL-based initialization strategy to make the population coverage as large as possible by generating quasi-opposite individuals to merge and sort with the initial population to fit the huge solution space of ultra-highdimensional datasets.The algorithm introduces a T-distribution perturbation strategy based on a check mechanism,which can expand the search range of individuals in the huge solution space and jump out of the local optimal solutions,improving the global search performance of the algorithm,and thus can have a better solution capability in the ultra-high-dimensional dataset.The algorithm further employs a novel binary operator,which is applicable to the discretized solution space while enabling the algorithm to be more compatible with the mechanism of frequent short-distance movement and small-probability long-distance migration of Lévy flights.Simulation experiments were conducted using ultra-high-dimensional datasets from the UCI Machine Learning Library and other commons.The simulation results show that IBMOQCS has stable performance and achieves better results with ultra-highdimensional datasets.