ARTICLE IN PRESS netrics, besides accuracy, to evaluate algorithms in a er systems usually do not employ only collaborative filtering meth- He d ods. Likewise, content-based methods are usually not employed ot effective due to the lack of mecha- res. Hence, each of these ap- Bae Kim, 2010). g methods are nender sys- ntages formance d items to usersmetrics, besides accuracy, to evaluate algorithms in a recommender context. Hence, we also consider the false positive occurrence and number of rules generated (if applicable for the algorithm being analyzed). Moreover, we also took into account some implicit fea￾tures related to shortcomings of recommender systems. In the next section we present the main classes of methods em￾ployed in recommender systems. Subsequently, we depict the most important ones. At this time, we also describe the use of asso￾ciation rules for classification problems. In Section 4, we quote and explain the main shortcomings and research challenges related to recommender systems. Finally, Section 5 contains the case study accomplished in order to experiment methods’ capability for alle￾viating effects of shortcomings produced in recommender systems. 2. Types of recommender methods Methods employed in recommender systems have their foun￾dations in different areas of knowledge. Cheung et al. (2003) and Lee, Kim, and Rhee (2001) classify recommender systems depend￾ing on the type of method used for making recommendations. The two main categories of recommender methods are collaborative filtering algorithms and content-based approaches. Content-based methods compare text documents to a specific user profile, where web objects are recommended to a user based on those he has been interested in the past (Lee et al., 2001). On the other hand, collaborative filtering methods were proposed aiming to provide more consistency to recommendations by means of information related to the social context of each user. To do so, the recommen￾dation will be given to the active user according to item’s opinions given by other users who have similar profile (similar preferences about other items). The collaborative filtering approach was originally based on the nearest neighbor algorithm (Sarwar, Karypis, Konstan, & Reidl, 2001), which recommends products to a target user according to the opinions of users who have similar purchase patterns. Thus, recommended products will be the ones liked by users with similar interests. This approach appeals to the notion that when we are looking for information, we often seek the advice of friends with similar tastes or other people whose judgment we trust (Condliff, Lewis, Madigan, & Posse, 1999). In this way, information about items that other people have already found and evaluated is taken into account. Breese, Heckerman, and Kadie (1998) classified col￾laborative filtering methods into two groups: memory-based methods and model-based methods. In memory-based methods the nearest neighbors of a target user is found by matching the opinions of such user to the opinions of all system’s users. On the other hand, model-based methods build a predictive model by means of a training set comprising opinions acquired from a small portion of the system’s users. Such methods have been devel￾oped more recently in order to avoid the sparsity problem, which usually arises when memory-based methods are employed, be￾cause e-commerce systems generally offer millions of products for sale, so that it is not feasible to obtain opinions about all of them (Sarwar, Karypis, Konstan, & Riedl, 2000). Current recommender systems usually do not hold a substantial number of evaluations comparing with the number of items avail￾able in the system. As a consequence, the system seldom encom￾passes a complete database with evaluations of all items available. That is why model-based methods generally employ a training set with evaluations gathered from just a part of users of the system. Even though, generally the number of evaluations available is pro￾portionally short and, as result, it is still necessary to develop more techniques to solve such shortcoming. Due to difficulties on obtaining considerable information of evaluation from users about system’s items, current recommender systems usually do not employ only collaborative filtering meth￾ods. Likewise, content-based methods are usually not employed solely, because they are not effective due to the lack of mecha￾nisms to extract Web objects features. Hence, each of these ap￾proaches has its strengths and weaknesses (Bae & Kim, 2010). Therefore, content-based and collaborative filtering methods are commonly combined or employed together in recommender sys￾tems. Combining different methods to overcome disadvantages and limitations of a single method may improve the performance of recommenders (Göksedef & Gündüz-Ögüdücü, 2010). 3. Methods employed in recommender systems Seeing that recommender systems recommend items to users based on ratings or past customer behavior and also that there are usually several items and users, they need to be grouped in or￾der to make recommendations feasible. In this way, classification methods are mostly employed to classify every user and/or item in one of those groups. Techniques used for classification are con￾sidered predictive methods because they aim at predicting the va￾lue of an attribute, called label-attribute, in a certain dataset. Each value of the label-attribute must be discrete since it is responsible for representing a class. The prediction of the label-attribute value is achieved by means of other attributes’ values (the descriptive attributes). The values of these attributes must be known in all samples, so that they may build a training set defining clearly the characteristics of all classes. Thus, classification is considered as a supervised learning approach. Furthermore, a test set is used to verify the consistence of the learning model. In this section we will describe the main techniques employed in recommender systems, which may be either supervised or unsu￾pervised learning techniques. In the next subsection we depict the most common machine learning methods employed in recom￾mender systems. Subsequently, we describe some foundations of association rules and how they can be employed in these systems. Finally, in Section 3.3, we describe classification based on associa￾tion methods. 3.1. Machine learning methods According to Bae and Kim (2010), most researchers have been using data mining techniques in recommender systems aiming at predicting the customer’s future behavior and to increase the chance of repurchasing. In this way, and since data mining ap￾proaches are basically made of machine learning methods, we may say that such methods are quite popular in the recommender systems area. Billsus and Pazzani (1998) were the first to apply machine learning techniques in recommender systems, the authors proposed to transform the traditional collaborative filtering recom￾mender problem (nearest neighbor) into a machine learning prob￾lem. To do so, authors firstly employed neural networks and considered the recommender problem as a classification problem. In this sense, a neural network constructs a model, defining classes of items, for recommendation using the ratings given by users and then classifying the non-rated items. More recent neural networks approaches (Chou, Li, Chen, & Wu, 2010) uses consumer knowl￾edge upon items in order to provide personalization in e￾commerce systems. They assume that costumers have some amount of experience or information about items they are using or plan to purchase. In this context, collaborative filtering may be seen as a prediction task, because the basic idea is to predict how a user will rate a new item (not rated before). That is the rea￾son why this new approach for collaborative filtering was called item-based methods, because recommendations are given based 2 J. Pinho Lucas et al. / Expert Systems with Applications xxx (2011) xxx–xxx Please cite this article in press as: Pinho Lucas, J., et al. Making use of associative classifiers in order to alleviate typical drawbacks in recommender sys￾tems. Expert Systems with Applications (2011), doi:10.1016/j.eswa.2011.07.136