Agglomerative Hierarchical Clustering (AHC) is a general type of Hierarchical Clustering (HC) that forms clusters from the “bottom-up.” This paper focuses on the development of AHC methods based on ensemble-based approaches. Accordingly, we develop an AHC framework based on clusters clustering along with an innovative similarity criterion that performs clustering through ensemble approaches. The proposed algorithm consists of three main steps. In the first step, a group of single AHC methods are combined to detect relationships between samples and as well as the formation of initial clusters. The similarity of the samples is calculated using an innovative similarity criterion based on the clusters created. In the second step, all the initial clusters created by different methods are re-clustered to form hyper-clusters. After clusters clustering, each sample is assigned to a hyper-cluster with maximum similarity to create the final clusters in the third step. The comprehensive experimental study has been performed to evaluate the performance of the proposed algorithm based on several benchmark datasets from the UCI machine learning repository. The results clearly show that the proposed ensemble AHC-based framework performs better than the state-of-the-art methods.

Many meta-heuristic algorithms have been proposed to solve continuous optimization problems. Hence, researchers have applied various techniques to change these algorithms for discrete search spaces. Artificial bee colony (ABC) algorithm is one of the well-known algorithms for real search spaces. ABC has a good ability in exploration but it is weak in exploitation. Several binary versions of ABC have been proposed so far. Since the methods are based on the standard ABC, they have the disadvantage of ABC. In this article, a new binary ABC called binary multi-neighborhood ABC (BMNABC) has been introduced to enhance the exploration and exploitation abilities in the phases of ABC. BMNABC applies the near and far neighborhood information with a new probability function in the first and second phases. A more conscious search than the standard ABC is done in the third phase for those solutions which have been not improved in the previous phases. The performance of algorithm has been evaluated by low- and high-dimensional functions and the 0-1 multidimensional knapsack problems. The proposed method has been compared with state-of-the-art algorithms. The results showed that BMNABC had a better performance in terms of solution accuracy and convergence speed.