ARTICLE IN PRESS even us sil g fu uzzy ns than other methods racy than classic orative fil- sociation rule sociation rules which is very e they can cluster. R m 1, also liked item nder sys-on items’ features. However, nowadays item-based collaborative methods are generally known as model-based methods. Since most machine learning techniques employed in recom￾mender systems are supervised learning methods, they need to be ran off-line in order to build a learning model used for classifi- cation. Thus, great part of the processing is performed off-line and hence, these methods do not require much on-line processing time, letting recommendations to be provided faster. Moreover, they may reduce shortcomings caused by sparsity, because they do not need ratings of all items available in order to build the classi- fication model. However, as highlighted in Section 2, due to difficulties on obtaining considerable information of evaluation from users about system’s items, model-based methods are commonly mixed with content-based methods in order to enhance recommendation’s quality. Despite being the first machine learning method applied for col￾laborative filtering, nowadays neural networks are not vastly em￾ployed in recommender systems. The ‘‘black box’’ learning approach of such method, where no information known except in￾put and output, is a crucial issue for building recommender models. In addition, the training time of a neural network is costly, which might be a critical issue in current recommender systems, whose data, according to Koren (2010), is changing over time and models should be continuously updated to reflect its present nature. As classifiers may be implemented employing many different machine-learning techniques, naturally recommender systems ex￾panded the model-based techniques they apply. Probably the most popular machine learning method for recommender systems are the Bayesian networks. They are an effective method employed in data mining that are widely applied for building recommender models (Condliff et al., 1999). The use of Bayesian networks for col￾laborative filtering was first suggested in Breese et al. (1998). It makes use of a training set to build a probabilistic structure for predicting user ratings. A distinct Bayesian network is used for rep￾resenting every user who recommendations will be provided to and a learning algorithm is responsible for processing every net￾work. Each node of the network represents, by means of a decision tree, a domain item and the edges represent information about the user. The states of a node correspond to rating values for the item being represented by the node. Since the classification model used for recommendations need to be build and, depending on the amount of data of the system, it may be very costly. On the other hand, the output model is small and its efficiency is similar to the nearest neighbor algorithm. However, it may still be not effi- cient enough depending on the number of items, because in order to predict the rating of a given item, the algorithm needs to calcu￾late the conditional probabilities of all possible ratings for such item given all possible ratings for other items. Another popular data mining technique widely employed in recommender systems, as a model-based method, is clustering. It consists on performing non-supervised learning in order to identify groups of users who appear to have similar preferences. Therefore, recommendations provided to the active user will be related to the opinions (or ratings) given by users of the group he owns to. Most clustering algorithms require a distance metric or similarity met￾ric, obtained by computing similarity between items, to guide the clustering process (Connor & Herlocker, 2001). Depending on the system, sometimes, after the clusters being created, opinions of other users may be averaged in the cluster in order to provide recommendations for the active user. Nevertheless, clustering techniques may use fuzzy logics and then every user may own to more than one cluster. In this way, each user will have partial participation in several clusters. At this point, a membership measure is assigned to each user in every cluster. Recommendation will be an average across the clusters, weighted by the membership value. However, even using fuzzy logics, according to Breese et al. (1998), clustering techniques usu￾ally produce less-personal recommendations than other methods, and in some cases, the clusters have worse accuracy than classic collaborative filtering (i.e., nearest neighbor) algorithms. Therefore, nowadays clustering techniques are usually applied in collabora￾tive filtering together with other methods. In this context, they are applied as a first step in order to distribute the data for differ￾ent recommender methods or to reduce the initial dataset. While dividing the population into clusters may hurt the accuracy of rec￾ommendations to users near the fringes of their assigned cluster, pre-clustering may be a worthwhile trade-off between accuracy and throughput (Schafer, 2005). Taking into account machine learning techniques mainly ap￾plied in collaborative filtering methods, the Support Vector Ma￾chines (SVMs) are quite frequent. Such technique consists in a supervised learning method for building a lineal classifier. In order to build a SVM, every user is seen as a vector composed by ratings about items. Such vectors are associated to a geometric space in which a hyperplane of separation between the possible classes is built. In this context, such classes are related to groups of users with similar preferences. Unlike other learning methods, SVM’s performance is related not to the number of features in the system, but to the margin with which it separates the data (Cheung et al., 2003). Association rule-based methods for classification are also em￾ployed in recommender systems. They are being more and more popular due to some benefits they present for the collaborative fil￾tering context. In the next subsection we detail association rule discovery applied for recommender systems. 3.2. Association rules As well as most data mining techniques, association rules induction algorithms can be employed to enhance personalization in recommendation systems, such as the ones in Lee et al. (2001), Lazcorreta, Botella, and Fernández-Caballero (2008). Association rules were first introduced by Agrawal, Imielinski, and Swami (1993) aiming at discovering consuming patterns in retail dat￾abases. Thus, the task of discovering association rules in retail data was termed as ‘‘market basket analysis’’. Data stored in market basket are items bought on a per-transaction basis for a retailing organization. The representation of an association rule may be de￾clared as A ? B, where A and B are item sets. Such representation states that, in a transaction, the occurrence of all items from ‘‘A’’ (antecedent side of the rule) results in the occurrence of items belonging to ‘‘B’’ (consequent side of the rule), such as A # I and B # I, where ‘‘I’’ is an item set. An association rule describes an association relation between item sets that occurs together on transactions of a data set. Thus, association, unlike classification, is not considered as a prediction task, because it aims at describing data. Therefore, association rule mining is not a machine learning method. There are several association rule mining algorithms available in the literature, such as ECLAT (Zaki, Parthasarathy, Ogihara, & Li, 1997), DIC (Brin, Motwani, Ullman, & Tsur, 1997) and FP-Growth (Han, Pei, & Yin, 2000). However, the Apriori (Agrawal et al., 1993) is certainly the most popular, and widely employed, association rules discovery algorithm nowadays (Neves, 2003). Its concepts and techniques are used by almost every algorithm proposed cur￾rently, which, are mostly mere extension of the Apriori (Neves, 2003). Our motivation to apply association rules in recommender sys￾tems is based on the structure of these rules, which is very appropriated for recommendation purposes, because they can learn patterns like ‘‘70% of users who liked item 1, also liked item J. Pinho Lucas et al. / Expert Systems with Applications xxx (2011) xxx–xxx 3 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