With the ever growing importance of internet, people are becoming overwhelmed by information. More concretely, consider a situation where you find yourself with an evening alone and would like to rent a DVD to watch. For several reasons, this is a difficult problem. First, most people have limited knowledge about the alternatives. Second, the set of alternatives changes frequently. In our example, there could be thousands of movies to choose from, and new movies are released on a daily basis. Third, this is an example of a low user involvement decision process, where the user is not prepared to spend hours expressing her preferences. Recommender systems have been designed to help people in this situation by finding the most relevant items based on the person's preferences. Two kinds of techniques are used in eCommerce sites today. The most widely used technique is collaborative filtering, (CF), which recommends products to users based on the experience of others. Amazon.com, with its 29 million customers and several million catalog items, uses item-based collaborative filtering to recommend items to the user. Item-based collaborative filtering works by finding similar items to the ones rated by the user, and then combines those similar items into a recommendation list. Thus, the ability of CF to recommend items depends on the capability of identifying a set of similar users. Furthermore, it does not build an explicit model of the user's preferences. Instead, preferences remain implicit in the ratings that the user gives to some subset of products, either explicitly or by buying them. Despite its popularity, CF suffers from profound problems such as the cold-start problem, and scalability issues. The former problem is due to the fact that a user must rate (too) many items before getting a recommendation, while the latter is correlated to the complexity of finding and working with the list of similar users. The other technique is the preference-based approach, (PBA). Here, a user is asked to express explicit preferences for certain attributes of the product. If preferences are accurately stated, then multi-attribute utility theory provides methods for finding the most preferred product, even when the set of alternatives is extremely large and/or volatile. Preference based approaches do not suffer from the same problem as collaborative filtering. However, the user needs to express a potentially quite complex preference model. This may require a large number of interactions, and places a higher cognitive load on the user since he has to reason about the attributes that model the product. To be able to reason over the items, multi-attribute utility theory also requires that products are modeled by a set of well defined attributes, which is sometime impossible to obtain in real life situation. Despite being very popular, these recommender systems sometimes fail to achieve high recommendation accuracy in eCommerce environments; especially when users' preferences are rare. Our analysis shows that these problems are due to two fundamental issues. First, current recommender systems use inappropriate models of the items and of the users' preferences. Second, recommender systems must elicit too many preferences from the user in order to build the user's preference profile. This dissertation proposes the ontology filtering approach that can overcome most of the problems faced by previous approaches, while achieving a better prediction accuracy than item-based collaborative filtering. The intuition behind ontology filtering is that the information captured by the topology of an ontology can be used to estimate missing preferences. The main novelties of this technique are to model the content of the eCatalog and infer missing preferences using the ontology, and then use the inferred information for directly recommending items to the user.