Collaborative Recommender Systems Essay

3.1 Collaborative Recommender Systems
It was in the article "PARC Tapestry system" (Goldberg et al., 1992) that the concept and term "collaborative" was first introduced. This system uses collaborative filtering which is the most popular, widely implemented and most mature among all the recommendation techniques. Even though the technique is just recently developed, the method that it employs can be seen since the very beginning of time (Schafer, Frankowski, Herlocker, & Sen, 2007), i.e. it uses the concept of "relying on the recommendations of other people". This filtering technique aggregates the recommendations of items, recognizes the recommendation commonalities between users and based on inter user comparisons it generates new recommendations. The user profile that is collected in the collaborative recommender system consists of a vector of items and its respective recommendations which are continuously growing as the user interacts with the system over a period of time. Collaborative systems can be further categorized based on the collaborative filtering algorithm used.
1. Memory-based collaborative filtering
This type of filtering is known as lazy recommendation algorithms, as the prediction computation of a user's interest is put off until the user requests for the recommendations (Eijlander & Bogers, 2009). This memory based filter is easy to incorporate and is already deployed in recommender systems of Amazon, Group Lens and Firefly. In the user-based version, memory based collaborative filtering recommends a user an item based on the average of the item ratings of similar users. In the item-based version, the filtering works by grouping similar items and removing items from the group that are already purchased by the user.
2. Model-based collaborative filtering
This type of filtering is known as eager recommendation algorithms, as most of the prediction computation is performed during the training phase. The model based filtering first creates a model for similar users by considering the rating matrix, following which it derives the particular rating for an item that is recommended by the model. In order to create the model, learning and statistical techniques are used (Su & Khoshgoftaar, 2009).
3.2 Content-based Recommender Systems
Content recommender systems provide recommendations for items based on the user's profile and item description. It learns the profile of the user's interests by analysing the features present in the objects that the user has rated in the past and creates a profile based on the rated items features. This type of relation is called "item-to-item correlation" (Schafer, Konstan & Riedl). The user profile type created by this recommender will vary based on the learning method (neural nets, decision trees) being used. The user profiles of content-based are long-term models and are continuously updated based on observed user preference evidences.
3.3 Knowledge-based Recommender Systems
Knowledge recommender systems provide recommendations for objects based on deducing the user's preferences and needs. This system uses information of recommendation criteria, user preference and item properties. Knowledge-based technique is distinguished because it has functional knowledge reasoning of how a particular item meets the need of a particular user which ensures the understanding