Hybrid Swarm Intelligence-Based Biclustering Approach for Recommendation of Web Pages

Introduction

Web usage mining (WUM) systems use data mining algorithms on both usage and clickstream data from one or more websites to discover usage patterns/user profiles. These patterns are analyzed to determine user’s behavior and these patterns are used by the recommendation systems to provide the recommendation of appropriate pages to the web users. Web personalization is the process of personalizing web sites or web services according to the specific user’s profile to achieve more efficient web browsing. User’s browsing efficiency can be increased by altering the web sites’ structures, and by employing recommender systems to produce user-tailored recommendations.

User profiles or Usage models are the generalization of the collected data about the user behavior from web usage data (such as clickstream data). The main goal of using user profiling is to increase the efficiency of user activities by delivering more personalized information when users interacting the web site. Usage-based Personalized Recommendation systems (Cooley et al., 1997;Srivastava et al., 2000) analyzed the user’s navigation pattern to provide personalized recommendations of web pages according to the current interests of the user.

Recently, web usage mining has gained much attention as it is found to fulfill the needs of web personalization. In the literature, various clustering algorithms can be applied to detect the user profiles as well as other web mining techniques such as association rule, have been explored by several research groups. Clustering is the process of grouping the users into clusters such that users within a cluster have high similarity compared to each other but dissimilar to users in other clusters.

The discovery of patterns from usage data by clustering the web transaction into clusters of user sessions or pages, by itself is not sufficient for performing the personalization tasks (Mobasher et al., 2002). The critical step in the recommendation system is the effective derivation of good quality and useful usage profiles from the web usage data. Clustering, in general tries to partition the set of web users according to their browsing interest of all pages of a web site. Assuming that users may show interest only for a particular subset of pages, conventional clustering approaches may not be sensitive and/or specific enough to find and present correlation between users in an appropriate and comprehensive manner.

To address the limitation of traditional clustering, the biclustering method was introduced. In contrast to traditional clustering, a biclustering method produces biclusters, which consists of subset of users and a subset of pages under which these users behave similarly.

Greedy search algorithms are used as the promising approach in the biclustering algorithms. Greedy search algorithms start with an initial solution and find a locally optimal solution by successive transformations that improve some fitness function. Most of the times, the results of biclustering suffer from local optima problem. Meta-heuristics optimization algorithms such as Particle Swarm Optimization (PSO) (Kennedy & Eberhart, 1995) Genetic Algorithm (GA) (Goldberg, 1989) and Simulated Annealing (SA) are used along with greedy biclustering to improve the results because it has potential to escape local optima.

The term Swarm Intelligence (SI) was coined by Beny and Wang in late 1980s in the context of cellular robotics. Swarm Intelligence is a collection of nature based algorithms which has attracted several researchers from the field of pattern recognition, information retrieval and clustering. SI systems are typically made up of a population of simple agents interacting locally with one another and with their environment. The efforts to mimic such behaviors through computer simulation finally resulted into the fascinating field of SI. An example of swarm intelligence is Particle Swarm Optimization (PSO) and it is a very popular SI algorithm for global optimization over continuous search spaces and also for discrete optimization problems.