Abstract. Content-based recommender systems can overcome many problems related to collaborative filtering systems, such as the new-item issue. However, to make accurate recommendations, content-based recommenders require an adequate amount of content, and external knowledge sources are used to augment the content. In this paper, we use Wordnet synsets to enrich a contentbased joke recommender system. Experiments have shown that content-based recommenders using K-nearest neighbors perform better than collaborative filtering, particularly when synsets are used.
IntroductionRecommender systems (RSs) have been an active research area in recent years, and most university computer science departments now offer RS-related courses [1]. Basically, what recommender systems do is to predict which items (e.g., movies, cameras, books, etc.) suit a user who has not seen them, and suggest those items to the user. RSs effectively solve the problem of having too many products on the internet toThe process of making recommendations requires three components: items, users and user feedback on items. Things that RSs recommend to users are called items, regardless of if they are a service, a trip or any other product. The second component of a recommendation process is users. Users are the center of the system, as some researchers define RSs as software that develops and uses customers' profiles [1] . The information needed to build a user profile varies from system to system; RSs might exploit users' demographic information, ratings or personality. Feedback, the third component, is how users interact with items, and it can be collected explicitly or implicitly. Rating items can have many classes, such as 1 to 5 stars, two classes (like/dislike), or only one class (like). After receiving suggestions, a user may provide feedback to indicate whether the user likes or dislikes the item; the system stores the user's opinion in a database, and uses it for future recommendations [1] [2].
Collaborative Filtering (CF)Collaborative filtering is the most widely used approach of recommender systems [3]. CF systems exploit the available ratings of some users to predict and recommend 296 H. Alharthi and D. Inkpen items to another group of users. In general, CF predicts the preferences of users by exploiting a user-item matrix that has the ratings of m users {u 1 , u 2 ,… u m } on n items {i 1 , i 2 ,…, i n }. A user u i has rated a list of items I ui . If the user has not rated the item, a missing value is shown in the matrix. There are two approaches of CF algorithms: user-based and item-based [4]. In the former, it is assumed that two users with similar tastes or rating histories will rate items similarly in the future. Each user is represented as a vector of items' ratings, which is called the user profile. The system finds users with similar profiles to the target user and exploits their ratings to predict the likeliness that the user likes a particular item [5]. The item-based CF system computes the similarity between two co-rated items, and...