User and Item Preference Learning for Hybrid Recommendation Systems

With the abundance of items, features, services and products available on the web, it becomes quite tedious to sort through and find what is desired. A recommendation system, therefore, provides a solution for this by only retrieving desired information from relevant data sets. A similarity metric is applied on the retrieved data and recommendations are made based on the predicted similarity. Typically, there are three major recommendation methods-collaborative, content-based and hybrid. Collaborative (user-based) works on the premise that similar users are bound to like the same items. Whereas, content-based (item-based) filtering tries to recommend items that are similar to the ones the user has liked previously. While functioning independently, the above two methods have certain flaws. Content-based systems cannot make predictions outside of the user's content profile and collaborative systems cannot make recommendations for cold start problems. To alleviate these issues, we seek to combine these two methods into a hybrid recommender system. Hybrid recommender, as the name suggests, is a segmented approach to recommendation where each segment consists of either the collaborative or content-based model. In this paper, we propose meta-level cum switching hybrid model that incorporates the dual hybridization techniques into the above mentioned hybrid model. The type of hybridization method applied depends on the ability of each individual model to generate results. I.e. In situations where both the individ-ual segments can generate results, the meta-level hybridization is used. Whereas, in a situation where any of the individual methods are unable to provide a result, the switching hybridization is enabled to shift to the method that is able to function in that scenario.