Article Preview
Top1. Introduction
Collaborative filtering (CF) (Adomavicius & Tuzhilin, 2005; Koren, 2008) recommendation technique predicts the interest of an active user by collecting rating information from other similar users or items. These techniques have been extensively applied in some popular commercial systems, such as Amazon (Linden et al., 2003), AdaptiveInfo.com (Billsus et al., 2002). More precisely, CF approach finds a set of similar users based on their preferences for items and recommends those items which have preferred by similar users. The CF based recommendation methods are based on two types of approaches: model-based approaches and memory-based approaches (Adomavicius & Tuzhilin, 2005; Breese et al., 1998). The model-based CF is based on a compact model inferred from the database, which is then used to make rating predictions. The model can be learned from the underlying data using data mining or machine learning techniques. For example, Bayesian belief nets (BNs) CF models (Breese et al., 1998; Miyahara & Pazzani, 2000; Su & Khoshgoftaar, 2006), clustering CF models (Ungar & Foster, 1998; Chee et al., 2001; Xue et al., 2005), and latent semantic CF models (Hofmann, 2004) are the well-known model-based CF techniques. Memory-based approaches utilize the entire user-item database to generate a prediction based on user or item similarity. These methods are particularly deployed into commercial systems such as http://www.amazon.com/ and Barnes and Noble, because they are easy-to-implement and highly effective (Hofmann, 2004; Linden et al., 2003). Further, two types of memory-based methods have been studied: user-based (Breese et al., 1998; Xue et al., 2005) and item-based (Linden et al., 2003). User-based methods try to find the similar users for an active user and then employ the ratings from those similar users to predict the ratings for the active user. Whereas, item-based methods attempt to find similar items for each active item (the item not rated by the active user) and then utilize the ratings of those similar items to predict the ratings of the active item.
The major concern of memory-based approaches is the data sparsity of the user-item rating matrix. Generally, the user-item rating matrix is quite sparse, which directly leads to the poor quality of predictions. In literature several algorithms have been developed to address the issue of data sparsity in CF based recommendation techniques (Billsus et al., 2002; Breese et al., 1998; Xue et al., 2005; Tyagi & Bharadwaj, 2010; Tyagi & Bharadwaj, 2012). Xue et al. (2005) proposed a scalable CF approach using cluster-based smoothing to tackle the problem of sparsity. Wang et al. (2006) suggested a generative probabilistic framework to exploit more of the data available in the user-item rating matrix by fusing all ratings with a predictive value for making recommendations. Ma et al. (2007) designed an effective scheme which predicts the missing data and then generates recommendations.
In this article a novel approach is proposed to alleviate the problem of sparsity in the framework of CF. The main emphasis of the proposed work is on the use of particle swarm optimization (PSO) algorithm (Kennedy & Eberhart, 1995) to learn the weights of different attributes used to embody the preferences of a user. This enables the fuzzy case based reasoning (CBR) system to represent the user (case) more precisely and to retrieve more appropriate case or cases. In this way, the proposed approach employs PSO based Fuzzy CBR called FCBR for filling missing values in the sparse user-item rating matrix and then generates recommendations by applying CF.