Improved Dragonfly Algorithm Based On Competition Operator And Destruction Operator

With the popularization of computer technology,people produce a lot of data every day.However,not all data is necessary,as much of it is duplicative,irrelevant,or of low quality.These redundant data will bring many problems,such as interfering data analysis and reducing algorithm performance^[1].Therefore,data mining becomes more and more important.Data mining^[2]is the cross-fusion of multiple disciplines（such as databases,machine learning,pattern recognition,algorithms and statistics）,aiming to extract useful information from massive data.Feature selection is a common data preprocessing method^[3].By eliminating repetitive,unnecessary or noisy features,the most relevant subset is selected from the data set to better capture data attributes and conduct proper classification.Feature selection is considered to be a NP-hard search problem^[4],which means that it is not easy to find the optimal minimum feature subset when dealing with high-dimensional data.Assuming the data set has N features,it takes 2^N attempts to test all possible feature subsets to determine which features can be used to distinguish data records^[3].This approach of generating all possible subsets in detail becomes impractical for high-dimensional data and computationally expensive.Therefore,in order to solve this problem,meta-heuristic algorithms^[5]are introduced,which reduce computational costs and improve search efficiency through some strategies and techniques to find the optimal feature subset.The Binary Dragonfly Algorithm（BDA）[6]is a population-based meta-heuristic that simulates dragonfly foraging and migration behavior.Through the study of the ecological habits and behavior rules of dragonflies,it is found that they have efficient and optimized search strategies,which can find the best foraging targets and habitats in a very short time.However,BDA,like most meta-heuristic algorithms,has the problem of convergence being too slow and falling into the local optimal solution too early.These problems will seriously affect the performance of the algorithm and reduce the classification accuracy.Therefore,this paper proposes a competitive operator,introduces a destruction operator,discretizes it and applies it to BDA to optimize BDA and develop a new meta-heuristic algorithm,the Competitive Swarm Binary Dragonfly Algorithm（CSBDA）based on competitive group optimization behavior.Finally,the excellent performance of CSBDA is verified through experiments.The Competitive Operator（CO）uses the competition mechanism of the real competition to divide the solution into winners and losers,the winner goes directly to the next iteration,and the loser learns from the winner in this competition.But to avoid falling prematurely into a local optimal solution,the loser must not only learn based on the performance of the winner in this competition,but also learn from the winner’s personal best score so far.In addition,losers are converted to binary after learning through an improved binary variant method.This improved binary variant method adds global optimization to the original V-shaped function,which binds each other between the individual optimal and the global optimal.The algorithm uses this behavior to jump out of the local optimal and further increase the probability of finding a better solution.In order to make the population more diverse,explore more search space,and further increase the probability of finding the optimal solution,this paper also introduces the Disruption Operator（DO）.Moreover,the cooperation between exploration and development of competing operators and destruction operators can make the algorithm maintain stable acceleration and continuously explore the optimal solution.Based on the above two operators,a new meta-heuristic algorithm CSBDA is generated.Then,this paper experiments with CSBDA to observe its performance.Firstly,this paper simulates CSBDA on 23 datasets from the UCI database to evaluate the performance of CSBDA.Secondly,based on the results of the CSBDA experiment obtained earlier,it is compared with 11 algorithms in recent years.The results show that the proposed CSBDA has better performance than the original BDA and better performance than other comparison algorithms.