Research On Hybrid Ant Colony Clustering Algorithm

Ant colony clustering algorithm is a bionic optimization algorithm inspired by ant colony behavior.It is suitable for discrete real environments and is widely used in data mining,image segmentation and other fields.However,the movement of ants in the ant colony clustering algorithm is random.At this time,the random allocation strategy of the ant colony requires a large number of search processes to find the correct foraging path.Besides,in the local neighborhood of the ants,the similarity and density of the data will affect the search for the correct foraging path,which leads to the reduction of the computational efficiency of the ant colony clustering algorithm and the possibility of falling into the local optimal result.Secondly,the ant colony clustering algorithm requires manual intervention on the number of clusters,which cannot automatically identify the cluster centers,where the algorithm is poorly adaptive.Aiming at the above-mentioned problems of ant colony algorithm,this thesis designs a corresponding hybrid ant colony clustering algorithm to improve the performance of ant colony clustering algorithm.The specific research content is as follows:(1)Taking into account the drawbacks of ant colony clustering algorithm's allocation strategy and local search,we propose a new hybrid ant colony clustering algorithm based on the means of K-means++ algorithm.First,the initial center point of K-means++ algorithm is calculated by the non-random data point.Then,we introduce a pre-clustering process in the new hybrid ant colony clustering algorithm with the help of clustering process in Kmeans++ algorithm,where the pre-clustering result is obtained and the center point matrix is updated to obtain the distance matrix.Secondly,the ant colony clustering algorithm is used to re-cluster the pre-clustering results.The allocation strategy is changed by considering the distance from the data point to the pre-clustering center point and the distance matrix as well as the pheromone concentration.We combine the advantage of simulated annealing algorithm,where the local search process is improved and the ant colony clustering result is generated.The pre-clustering result and the ant colony clustering result are compared,where the best is used to update the pheromone.Finally,the center point obtained by each ant colony clustering is used as the input of pre-clustering in the next hybrid ant colony clustering.Experimental results show that the proposed algorithm improves the allocation strategy of data points in ant colony clustering,reduces the randomness of ant movement,improves computational efficiency and avoids finding local optimal solutions.(2)Taking into account the drawbacks of central point identification and adaptability of ant colony clustering algorithm,we propose a new hybrid ant colony clustering algorithm based on improved density peak clustering.The algorithm is divided into two stages: data division and data merging.In the first stage,a method of determining the value inflection point is proposed to improve the density peak algorithm and obtain the initial cluster center point,and generate a distance matrix.The new density peak algorithm can make more accurate automatic selection of appropriate cluster center points.Then,the ant colony clustering algorithm combines the pheromone matrix and the distance matrix to divide the data set,where the number of clusters obtained by the pre-clustering is larger than the real number.In the second stage,based on the algorithmic idea of hierarchical clustering,the similarity between the clusters in the pre-clustering is calculated,and the pair of clusters with the largest similarity is merged each time,and the clustering evaluation index is used to determine the best merging result of the sub-categories,so as to accurately merge the redundant sub-categories divided in the first stage and obtain the final clustering result.Experiments on multiple artificial data sets and UCI data sets respectively prove that the algorithm is efficient,suitable for center point recognition of arbitrary shape data sets,and of more adaptability.