Web Recommender Systems

Web recommender systems are used to locate relevant items in which the user is interested. This can be done based on the user’s data that is collected implicitly (Web access logs) or explicitly (ratings). Generally, it is more efficient and user-friendly to provide users with what they need automatically and without asking them explicitly for it [1]. Web recommender systems are used frequently in e-commerce and information access in order to assist the user in locating relevant products, items or services such as on Amazon and CDNow. A number of knowledge discovery and statistical techniques are generally used in advanced recommendation systems. The latter can be divided, depending on the techniques used, into content-based filtering, collaborative filtering, and hybrids, which are summarized below. Other approaches such as demographic or knowledge based, also exist (but tend to require extensive private user information or manual construction of knowledge). Automatic recommendation implies that the user profiles are created and eventually maintained dynamically by the system without explicit user information. Examples include amazon.com’s personalized recommendations and music recommenders like Mystrand.com in commercial systems [2], and smart recommenders in e-learning [3], and Hybrid Recommender System for E-Learning Personalization [4], etc. In general, such systems differ in the input data, in user profiling strategies, and in prediction techniques. Several approaches for automatic personalization have been reported in the literature, such as content-based or item-based filtering, collaborative filtering, rule-based filtering, and techniques relying on Web usage mining, etc [5]. In the e-learning area, one of the new forms of personalization is to give recommendations to learners in order to support and help them through the e-learning process.  In this paper, we present our proposed personalization approach taking into account both the Web access history of learners as well as the content of the learning material. Our approach is based on applying Web usage mining techniques in combination with an open source Web information retrieval system to enable an implementation that is not only open and scalable, but also fast to deploy. 

Content based filtering: Content-based filtering (or item-based filtering) systems recommend items to a given user based on the correlation between the content of these items and the preferences of the user [6]. This means that the recommended items are considered to be similar to those seen and liked by the same user in the past. Thus, there is no notion of a community of users, rather only one user profile is considered while making recommendations.
Collaborative based filtering: Collaborative filtering systems have tried to remedy for the shortcomings listed earlier. Collaborative filtering system recommends items that are liked by other users with similar interests. So, the exploration of new items, in this approach, is assured by the fact that other similar user profiles are also considered. Thus, the history of community of users is combined. Examples of such systems include GroupLens [7] and [8]. This approach relies on a historic record of all user preferences and interests which can be obtained explicitly by asking users to give a rating on each item (product, Web page, etc) or implicitly by observing the user’s behaviour when browsing a Website, for instance via their clickstreams.
Hybrids: Hybrid recommender systems combine several recommendation strategies to provide better performance than either strategy alone. Most hybrids work by combining several input data sources or several recommendation strategies. There are many hybridization methods reported in the state of the art. Generally, content/collaborative hybrids are the most popular hybrid strategies.