a b s t r a c tData clustering is related to the split of a set of objects into smaller groups with common features. Several optimization techniques have been proposed to increase the performance of clustering algorithms. Swarm Intelligence (SI) algorithms are concerned with optimization problems and they have been successfully applied to different domains. In this work, a Swarm Clustering Algorithm (SCA) is proposed based on the standard K-Means and on K-Harmonic Means (KHM) clustering algorithms, which are used as fitness functions for a SI algorithm: Fish School Search (FSS). The motivation is to exploit the search capability of SI algorithms and to avoid the major limitation of falling into locally optimal values of the K-Means algorithm. Because of the inherent parallel nature of the SI algorithms, since the fitness function can be evaluated for each individual in an isolated manner, we have developed the parallel implementation on GPU of the SCAs, comparing the performances with their serial implementation. The interest behind proposing SCA is to verify the ability of FSS algorithm to deal with the clustering task and to study the difference of performance of FSS-SCA implemented on CPU and on GPU. Experiments with 13 benchmark datasets have shown similar or slightly better quality of the results compared to standard K-Means algorithm and Particle Swarm Algorithm (PSO) algorithm. There results of using FSS for clustering are promising.

Abstract-Association is widely used to find relations among items in a given database. However, finding the interesting patterns is a challenging task due to the large number of rules that are generated. Traditionally, this task is done by post-processing approaches that explore and direct the user to the interesting rules of the domain. Some of these approaches use the user's knowledge to guide the exploration according to what is defined (thought) as interesting by the user. However, this definition is done before the process starts. Therefore, the user must know what may be and what may not be interesting to him/her. This work proposes a general association rule post-processing approach that extracts the user's knowledge during the post-processing phase. That way, the user does not need to have a prior knowledge in the database. For that, the proposed approach models the association rules in a network, uses its measures to suggest rules to be classified by the user and, then, propagates these classifications to the entire network using transductive learning algorithms. Therefore, this approach treats the post-processing problem as a classification task. Experiments were carried out to demonstrate that the proposed approach reduces the number of rules to be explored by the user and directs him/her to the potentially interesting rules of the domain.

No abstract

No abstract