2". In this context, the rules can denot er on different beings can easier comprehend le algorithms than an output As highli ree data n recommender recommendations. An error in a2’’. In this context, the rules can denote two types of associations: between users and items (in the case of content-based methods) and associations between items or users (in the case of collabora￾tive filtering methods). The first type of association rules in recom￾mender systems consists on associate active user’s profile data with items’ features. The second one, employed in collaborative fil￾tering, consists on associating data of the active user with data of other users or on associating data of items the active user is inter￾ested in with data of other items available on the system. In Sun, Kong, and Chen (2005), for example, association rules are applied to mine relationships between items and then make the prediction about an item for an active user by computing the weighted sum of the ratings given by the user about the items similar to the target item. Géry and Haddad (2003) propose the use of implicit opinions (instead of ratings) of users. The problem of finding Web pages vis￾ited together is similar to find associations rules among itemsets in transaction databases. In most cases, rules may be discovered off￾line by processing data related to users’ opinions. Recommender models based on association rules are easy to be interpreted and are usually faster to be built than most machine learning models. A recommender model using Bayesian networks, for example, considers the conditional probabilities of all the pos￾sible opinions for an item given all possible opinions for other items. Moreover, association rules may alleviate the sparsity draw￾back, which will be described into the next section, because rules do not need to consider all opinions from users in order to build a recommender model. Furthermore, the system may take into ac￾count just the best rules for the recommender model. Rules may be evaluated and ranked by means of statistical measures like support and confidence. In the case study depicted in the next section, we will analyze in a more detailed way how association rules can re￾duce sparsity. However, Zhang and Chang (2005) affirm that recommender models built employing only association generally present poor￾quality recommendations. In Zhang and Chang (2005), association rules were combined with sequential rules to enhance recommen￾dations efficiency. While in Lee et al. (2001), a method combining the nearest neighbor algorithm with association rules was devel￾oped. Such method employs information acquired from transac￾tions of groups of users with similar preferences (neighbors) in order to discover association rules about Web objects. Alternatively, other machine learning methods could be applied for recommender systems, as the one proposed by Liu, Hsu, and Ma (1998). In such work, authors adapted association rules in order to play the role of a classifier. Such method will be described in the next subsection and it will be tried for recommender systems in Section 5. 3.3. Classification based on association methods As highlighted in the previous subsection, association rules aim at describing data and, consequently, they are seen a non-super￾vised learning method. On the other hand, a classification method is seen as a prediction technique, because it aims at predicting the value of an attribute (label) in a data set. The joining of concepts from classification and association (Liu et al., 1998) is an alterna￾tive approach for performing classification tasks, where association rules are employed as the basis of a classification method. Seeing that association models are commonly more effective than classifi- cation models, a crucial matter that encourages the use of associa￾tion rules in classification is the high computational cost that current classification methods present. Several works (Li, Han, & Pei, 2001; Liu et al., 1998; Thabtah, Cowling, & Peng, 2005; Yin & Han, 2003) verify that classification based on association methods presents higher accuracy than traditional classification methods. Differing from association rules, decision trees, for example, do not consider simultaneous correspondences occurring on different attribute values. Moreover, human beings can easier comprehend an output provided by association rule algorithms than an output provided by usual classification techniques, such as artificial neural networks (Sarwar et al., 2000). Thabtah et al. (2005) sustain that a few accurate and effective classifiers based on associative classification have been presented recently, such as CBA (Classification Based in Association) (Liu et al., 1998), CPAR (Classification based on Predictive Association Rules) (Yin & Han, 2003), and CMAR (Classification based on Multi￾ple class-Association Rules) (Li et al., 2001). Taking into account that for classification rule mining there is one and only one prede￾termined target, while for association rule mining the target of dis￾covery is not pre-determined (Liu et al., 1998), it is necessary to constrain the rules’ consequent terms to encompass only one attri￾bute. Thus, the consequent term of an association rule will repre￾sent the target, or class, attribute. Therefore, such rule can play a prediction role in a given system: in order to classify an item, the rule’s properties are matched to every rule’s antecedents and the attribute value of the consequent term (from one or more selected rules) will be the predicted class. Generally, the classification mod￾el is a set of ordered rules. In the CBA algorithm, for example, the rules are ordered by means of the confidence measure and it uses only one rule for per￾forming classification. However, in this case some scenario in which could exist multiples rules with similar confidence mea￾sures may occur and, at the same time, with greatly different sup￾port measures. Hence, a rule A with much higher confidence than a rule B could be the one chosen for classification even if B had a much higher support (Li et al., 2001). The MCAR algorithm solves such drawback by means of an approach that considers, in addition to the confidence, the rules’ support. The CMAR algorithm has a fine approach for selecting association rules for classification, in￾stead of using just one rule, it makes use of all rules that match the case to be classified. If the consequent term of all selected rules is the same, the predicted class will obviously be the value of the rules’ consequent term. Though, in a different scenario, rules are divided in groups according to the consequent terms’ values. The value chosen for classification is acquired through the group in which its elements hold the highest correlation value depending on the weighted v2 measure. Similarly to CMAR, the CPAR algo￾rithm also divides rules in groups, though, instead of using all rules that match to the object to be predicted, it uses the ‘‘k’’ best rules that represent each class. Afterwards, the algorithm chooses a group, by means of the Laplace Accuracy measure, that will be the one used for classification. The drawbacks presented by association rules induction algo￾rithms are, in general, the same ones of classification based on asso￾ciation algorithms. A critical drawback of these algorithms is due to those rules that have few attributes. Seeing that such rules ex￾presses narrow information, an object which has few attributes would be ineffectively classified. Another critical drawback is due to the large number of rules that algorithms commonly produce (Sarwar et al., 2001), as a consequence, much of them do not supply relevant information or are contradictory. Such drawback is a crit￾ical issue related to associative classifiers, because the performance of the algorithm may be affected when retrieving, storing, pruning and sorting a large number of rules (Li et al., 2001). The CMAR algo￾rithm tries to solve such drawback by implementing a FP-Tree data structure to store the association rules’ frequent itemsets. 4. Recommender systems shortcomings Shortcomings inherited by methods employed in recommender systems are reflected in erroneous recommendations. An error in a 4 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