Improving Performance in Recommender System: Collaborative Filtering Algorithm and User’s Rating Pattern

Abstract

With the advance of IT technologies, a recommender system in online commerce environment has been introduced as personalized services (Schafer et al., 2006). The recommender system is used in E-commerce for recommending a product, an item or even any web service to each customer based on customer’s preference. Since a recommender system can predict customers’ preference and forecast the future degree of customer’s fondness for a certain item and services, it is used as a conspicuous service which distinguishes an on-line commerce service from an off-line commerce service. In predicting each user’s preference, a recommender system essentially provides enough information of items and users because it is able to predict the specific user’s preference for a target item and suggest the result to users. One of the classic recommender systems is a content-based filtering system which uses textual contents. In the recommender system for an on-line movie rental process, two types of profiles are usually used; movie profile and customer profile. The movie profile describes a movie category, main actors, and performance movie. The customer profile is created with the historical experienced textual information, which is stored in the system, of items or users for seeking the best fits. This type of approach works well in initial systems, but there are some drawbacks for expanding the scales of recommender systems due to the following reasons. First, there are difficulties in converting features of all traded items into textual data. Additionally, if the number of traded items extremely increases, it is not easy to automatically convert all items’ information into textual forms. Second, since content-based systems only recommend items based on the past experience of user, it cannot help the user choose items for specific cases. This problem is called over-specification for recommendations. Such drawbacks can be eliminated by collaborative filtering recommender systems, which use relationships between users and items that can be represented on numerical scales (i.e. preference rating). This preference rating information can be collected from tracks of clients who surf web and purchase items. Typically, such types of recommender systems utilize neighbour users’ data, using a set of data that has similar characteristics for 10