ARTICLE N PRESS xpert Systems with Applications xxx(2011)xXX-XXX Contents lists available at ScienceDirect Expert Systems with Applications ELSEVIER journalhomepagewww.elsevier.com/locate/eswa Making use of associative classifiers in order to alleviate typical drawbacks in recommender systems Joel Pinho Lucas", Saddys Segrera, Maria N Moreno Departamento de informatica y Automatica, Universidad de salamanca, Plaza de la Merced s/n 37008,. Salamanca, spain ARTICLE FO ABSTRACT Nowadays, there is a constant need for perso in e-commerce systems. Recommender systems make suggestions and provide information Is available, however, many recommender tech- Associative classification niques are still vulnerable to some shorto his work, we analyze how methods employed in these systems are affected by some typical drawbacks. Hence, we conduct a case study using data gath- ed from real recommender systems in order to investigate what machine learning methods can all ate such drawbacks. Due to some especial features inherited by associative classifiers, particular attention to this category of methods to test their capability of dealing with typical G 2011 Elsevier Ltd. All right 1 Introduction other domains with similar drawbacks(Moreno, Ramos, Garcia, Toro, 2008), we try classification based on association, which is a Nowadays, e-commerce systems present loads of products machine learning technique that combines concepts from classifi- available for sale. Thus, users would probably have difficulty to cation and association, in recommender systems. As it will be de- choose the products they prefer and, consequently, to purchase scribed along the next sections, such technique may present them. Due to such facts and to a more and more competitive indus- several advantages if applied for building a recommender model try. these systems need to personalize their products'presentation. Moreover, in a preliminary study made in Lucas, Segrera, and A way to reach such personalization is by means of therecom- Moreno(2008), we proved that classification based on association mender systems", which according to Bae and Kim(2010) have have a promising potential in the recommender systems domain. emerged in e-commerce applications to provide advice to cus- n order to perform such analysis, we accomplished a case study tomer about items they might wish to purchase or examine. In this using two recommender systems databases. The first consists of nse, recommender systems aim at enabling the creation of a new ratings of movies made by movielens users and the second con- ore personally designed for each consumer (Schafer, Konstan, sists of book ratings made the BookCrossing community. We con- sidered essential to use MovieLens data on the case study we Taking into account that machine learning techniques are ap- propose, because almost every case study on recommender sys- plied for identifying patterns with different purposes, such as clas- tems we found employed mainly movie rating datasets(mostly sification or prediction and knowledge discovery, according MovieLens). Cheung, Kwok, Law, and Tsui(2003), these techniques can lowever, employing just one type of data from a single domain cessfully applied to predict users' preferences in recommender sys- may limit the scope of many case studies. This shortcoming on this tems. Machine learning methods can provide several research field is also confirmed by herlocker, Konstan, Terveen, provements on these systems and then provide more and Riedl(2004), who argues that the lack of variety in publicly ization. However, even employing machine learning methods, rec- available collaborative filtering datasets (particularly with signi ommender systems still suffer from innumerous limitations and cant numbers of ratings) remains one of the most significant chal may be very susceptible to produce errors. lenges in the field. In this way, we decided to employ, in addition to In this work we investigate the main shortcoming presented by a classic movie rating data, a different data base within a different recommender systems and how they affect recommendations' domain in order to enhance the quality of the case study presented quality. Afterwards, we analyze how they may be alleviated. There- in this work. fore, after obtaining successful results with association rules Herlocker et al.(2004)also affirm that even early I recognized that, in a recommender scenario it can be Corresponding author. Tel. +34 923294653: fax: +34 923294514 able to measure how often the system leads its users to wrong choices and that accuracy differences are usually tiny even when ucas).saddyseusales(S. Segrera) they are measurable. Based on such observations, we consider other 0957-4174/s- see front matter o 2011 Elsevier Ltd. All rights reserved doi:10.1016/eswa2011.07.136 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/jeswa2011.07.136Making use of associative classifiers in order to alleviate typical drawbacks in recommender systems Joel Pinho Lucas ⇑ , Saddys Segrera, María N. Moreno Departamento de Informática y Automática, Universidad de Salamanca, Plaza de la Merced s/n 37008, Salamanca, Spain article info Keywords: Recommender systems Associative classification Sparsity abstract Nowadays, there is a constant need for personalization in e-commerce systems. Recommender systems make suggestions and provide information about items available, however, many recommender tech￾niques are still vulnerable to some shortcomings. In this work, we analyze how methods employed in these systems are affected by some typical drawbacks. Hence, we conduct a case study using data gath￾ered from real recommender systems in order to investigate what machine learning methods can allevi￾ate such drawbacks. Due to some especial features inherited by associative classifiers, we give a particular attention to this category of methods to test their capability of dealing with typical drawbacks. 2011 Elsevier Ltd. All rights reserved. 1. Introduction Nowadays, e-commerce systems present loads of products available for sale. Thus, users would probably have difficulty to choose the products they prefer and, consequently, to purchase them. Due to such facts and to a more and more competitive indus￾try, these systems need to personalize their products’ presentation. A way to reach such personalization is by means of the ‘‘recom￾mender systems’’, which according to Bae and Kim (2010) have emerged in e-commerce applications to provide advice to cus￾tomer about items they might wish to purchase or examine. In this sense, recommender systems aim at enabling the creation of a new store personally designed for each consumer (Schafer, Konstan, & Riedl, 2001). Taking into account that machine learning techniques are ap￾plied for identifying patterns with different purposes, such as clas￾sification or prediction and knowledge discovery, according to Cheung, Kwok, Law, and Tsui (2003), these techniques can be suc￾cessfully applied to predict users’ preferences in recommender sys￾tems. Machine learning methods can provide several improvements on these systems and then provide more personal￾ization. However, even employing machine learning methods, rec￾ommender systems still suffer from innumerous limitations and may be very susceptible to produce errors. In this work we investigate the main shortcoming presented by recommender systems and how they affect recommendations’ quality. Afterwards, we analyze how they may be alleviated. There￾fore, after obtaining successful results with association rules in other domains with similar drawbacks (Moreno, Ramos, García, & Toro, 2008), we try classification based on association, which is a machine learning technique that combines concepts from classifi- cation and association, in recommender systems. As it will be de￾scribed along the next sections, such technique may present several advantages if applied for building a recommender model. Moreover, in a preliminary study made in Lucas, Segrera, and Moreno (2008), we proved that classification based on association have a promising potential in the recommender systems domain. In order to perform such analysis, we accomplished a case study using two recommender systems databases. The first consists of ratings of movies made by MovieLens users and the second con￾sists of book ratings made the BookCrossing community. We con￾sidered essential to use MovieLens data on the case study we propose, because almost every case study on recommender sys￾tems we found employed mainly movie rating datasets (mostly MovieLens). However, employing just one type of data from a single domain may limit the scope of many case studies. This shortcoming on this research field is also confirmed by Herlocker, Konstan, Terveen, and Riedl (2004), who argues that the lack of variety in publicly available collaborative filtering datasets (particularly with signifi- cant numbers of ratings) remains one of the most significant chal￾lenges in the field. In this way, we decided to employ, in addition to a classic movie rating data, a different data base within a different domain in order to enhance the quality of the case study presented in this work. Herlocker et al. (2004) also affirm that even early researchers recognized that, in a recommender scenario, it can be more valu￾able to measure how often the system leads its users to wrong choices and that accuracy differences are usually tiny even when they are measurable. Based on such observations, we consider other 0957-4174/$ - see front matter 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.07.136 ⇑ Corresponding author. Tel.: +34 923294653; fax: +34 923294514. E-mail addresses: joelpl@usal.es (J. Pinho Lucas), saddys@usal.es (S. Segrera), mmg@usal.es (M.N. Moreno). Expert Systems with Applications xxx (2011) xxx–xxx Contents lists available at ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa 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