Article Preview
TopIntroduction
E-commerce is developing so fast that nearly everyone who use internet would like to buy or sale items (products and services) or get information about items. E-commerce websites and open rating systems such as epinions.com allow registered users to rate items they have bought and to get recommendations about the items which are thought of having more possibilities of fitting their preferences. For instance, users can rate items with numerical scores and publish trust or distrust relationships with other users in the epinions.com. However, users are suffering from an information overload problem (Chen, Shang, & Kao, 2009), which means that too much information makes it too hard for users to select the most suitable or favored ones because they do not have enough time to know or even browse all the information. In the area of item transaction such as e-commerce, recommender systems emerge to response to the aforementioned information overload problem by recommending users with the items they might like automatically.
Recommender systems are relevant with data mining, which is another important way to resolve information overload problem. Data mining is defined as the process of extracting implicit, valuable and interesting rules (or patterns) from large sets of data (Shekhar, Lu, Chawla, & Zhang, 2000). Some data mining technologies including classification, clustering and association rules have been applied in the area of recommender systems (Schafer, 2005). For instance, Good et al. utilize classification technology of data mining based on the vector of movie features to analyze uses’ interest in movies (Good et al., 1999). Another example of data mining technology used in recommender systems is that the association rule discovery is used to find out the items associated with the ones target users have expressed interest in (Sarwar, Karypis, Konstan, & Riedl, 2001). However, recommender systems are also different from data mining. Recommender systems are expected to generate recommendation to fit users’ interest rather than extracting rules by mining data. The key problem of recommender systems is to analyze historical user data and then predict users’ interests. Collaborative filtering (CF) is one of the most popular methods used in predicting users’ interest in recommender systems. CF methods do not utilize data mining technologies. They recommend target users with the items generated by aggregating their neighbors’ interest.
In both research area and industry area, recommender systems are becoming more and more popular. However, recommender systems are still suffering from at least two major challenges. The first is data sparsity problem, which results from ratings given by users are often very less compared to the massive amounts of items. The second is cold start problem, which is due to the users (usually new users) who review few items and provide little information about themselves. The two problems cause recommender systems hard to discover user interest. Many recommender systems have taken into account the data sparsity problem and the cold start problem in user interest analyzing (Huang, Chen, & Zeng, 2004). To resolve the two problems, trust has been incorporated into recommender systems to help leverage the performance of recommender systems (Jamali & Ester, 2009;Massa & Bhattacharjee, 2004).