M.Y.H. Al-Shamri, K.K. Bharadwaj/ Expert Systems with Applications 35(2008)1386-1399 The most familiar and most widely implemented filter- as fuzzy. But at the item level it is difficult to fuzzify the ing is the collaborative filtering. The initial CF was intro- profile because it would require prohibitively large space duced by Tapestry(Goldberg, Nichols, Oki, Terry, and long processing time 1992), and was automated by GroupLens(Resnick et al Our work in this paper is an attempt towards introduc- 1994)and Ringo( Shardanand Maes, 1995). Typically, ing hybridization at four different levels, namely feature- CF explores similar users(neighbors), recognizes common- level, model-level, CF algorithm-level, and approach-level alities between the user and his neighbors on the basis of Firstly, we proposed hybrid features that exploit both user their ratings, and then accordingly generates new recom- ratings for high rated items and some content description mendations based on inter-user comparisons(Adomavicius of the items. At the model-level we built a user model from Tuzhilin, 2005; Eirinaki Vazirgiannis, 2003). Further, the set of hybrid features and DMf profile. Hybridization CF and dMf have the unique capacity to identify cross- between model-based and memory-based algorithms of CF genre niches and can entice users to jump outside of the is done at the CF algorithm-level. The user model is used to familiar outside the box'( Burke, 2002) find a set of like-minded users within which a memory- Breese, Heckerman, and Kadie (1998)classified CF based search is carried out. This set is much smaller in size algorithms as either memory-based(Resnick et al., 1994; than the original set, thus making the technique scalable Shardanand Maes, 1995), the entire data is used for rec- Finally, we developed a hybrid fuzzy-genetic Rs by ommendations, or model-based (Shahabi et al., 2001)in employing Ga to evolve appropriate weights for each fea- which a model is derived offline from the data to be used ture of the user model and proposing a novel fuzzy distance for online recommendations. While the former is more metric to match users accurate, its scalability compared to model-based is poor. The contributions of this paper are three-fold: Practically, CF faces two fundamental challenges, namely accuracy and scalability. Although memory-based algo- A novel user model is built that enables hybrid filtering, rithms are simple, provide high accuracy recommend reduces the system complexity and the computational tions, and admit easy addition of new data, but they are time by roughly a factor of six. computationally expensive as the size of the input dataset A novel fuzzy distance function is proposed for users increases. Eventually, the user will leave the Web site before the processing completes. On the other hand, apply- .A hybrid fuzzy-genetic approach to recommender sys ing the model-based algorithm alone on such sparse data tems is developed though reduces the online processing cost, often comes at the cost of recommendation accuracy. One common threat The rest of this paper is organized as follows: some in current RS research is the need to combine recommenda- background on the rs is given in the next section. In Sec tion techniques to achieve peak performance because each tion 3, the proposed user model is presented, while the technique has its own pros and cons fuzzy and hybrid fuzzy-genetic approaches are given in Sec Essentially, Rs keep a profile for each user Without any tion 4. The experimental results of the proposed information about the user, the rs are not able to assist approaches and classical approach are discussed in Section the user. The user profile contains raw information about 5. Finally, in the last section we conclude our work with a the user. The terms user profile and user model are often review of our contributions along with some future used as synonyms. However, recent researchers(Froschl, research directions 2005: Koch, 2000) differentiate between them according to the level of sophistication. The user profile is a collection 2. Background of raw personal information represents preferences, back ground, personal details, and interactions with the system. 2. 1. Recommender systems Depending on the user profile, a user Thus, the user profile is used to retrieve the needed infor Recommender systems have gained an increasing mation to build up a model for the user. Koch(2000) importance since the early work on CF in the mid-1990s describes the user model as the representation of the sys- when researchers started focusing on RS that explicitly tems beliefs about the user and the user profile as a simple rely on the ratings structure(Adomavicius Tuzhilin, 2005). Normally, explicit ratings from users are binary rat Some efforts have been made towards introducing fuzz- ings (like/dislike)or follow a specified numerical scale less in RS. Nasraoui and Petenes(2003)used fuzzy indicating the degree of preference (e.g, 1-bad to pproximate reasoning to develop a general framework MAX-excellent, where MAX is the maximum possible for the recommendation process while Suryavanshi, Shiri, rating for a given system). Usually, one of the following and Mudur(2005)used relational fuzzy subtractive cluster- information filtering techniques is employed to generate ing Shahabi et al.(2001)proposed a Yoda Rs that softly recommendations: classifying active user based on typical patterns of users and then generating soft recommendations for him. The Demographic filtering (DMF): The user will be recom- user profile includes many features that can be described mended items similar to the ones other people with sameThe most familiar and most widely implemented filter￾ing is the collaborative filtering. The initial CF was intro￾duced by Tapestry (Goldberg, Nichols, Oki, & Terry, 1992), and was automated by GroupLens (Resnick et al., 1994) and Ringo (Shardanand & Maes, 1995). Typically, CF explores similar users (neighbors), recognizes common￾alities between the user and his neighbors on the basis of their ratings, and then accordingly generates new recom￾mendations based on inter-user comparisons (Adomavicius & Tuzhilin, 2005; Eirinaki & Vazirgiannis, 2003). Further, CF and DMF have the unique capacity to identify cross￾genre niches and can entice users to jump outside of the familiar ‘outside the box’ (Burke, 2002). Breese, Heckerman, and Kadie (1998) classified CF algorithms as either memory-based (Resnick et al., 1994; Shardanand & Maes, 1995), the entire data is used for rec￾ommendations, or model-based (Shahabi et al., 2001) in which a model is derived offline from the data to be used for online recommendations. While the former is more accurate, its scalability compared to model-based is poor. Practically, CF faces two fundamental challenges, namely accuracy and scalability. Although memory-based algo￾rithms are simple, provide high accuracy recommenda￾tions, and admit easy addition of new data, but they are computationally expensive as the size of the input dataset increases. Eventually, the user will leave the Web site before the processing completes. On the other hand, apply￾ing the model-based algorithm alone on such sparse data, though reduces the online processing cost, often comes at the cost of recommendation accuracy. One common threat in current RS research is the need to combine recommenda￾tion techniques to achieve peak performance because each technique has its own pros and cons. Essentially, RS keep a profile for each user. Without any information about the user, the RS are not able to assist the user. The user profile contains raw information about the user. The terms user profile and user model are often used as synonyms. However, recent researchers (Froschl, 2005; Koch, 2000) differentiate between them according to the level of sophistication. The user profile is a collection of raw personal information represents preferences, back￾ground, personal details, and interactions with the system. Depending on the user profile, a user can be modeled. Thus, the user profile is used to retrieve the needed infor￾mation to build up a model for the user. Koch (2000) describes the user model as the representation of the sys￾tem’s beliefs about the user and the user profile as a simple user model. Some efforts have been made towards introducing fuzz￾iness in RS. Nasraoui and Petenes (2003) used fuzzy approximate reasoning to develop a general framework for the recommendation process while Suryavanshi, Shiri, and Mudur (2005) used relational fuzzy subtractive cluster￾ing. Shahabi et al. (2001) proposed a Yoda RS that softly classifying active user based on typical patterns of users and then generating soft recommendations for him. The user profile includes many features that can be described as fuzzy. But at the item level it is difficult to fuzzify the profile because it would require prohibitively large space and long processing time. Our work in this paper is an attempt towards introduc￾ing hybridization at four different levels, namely feature￾level, model-level, CF algorithm-level, and approach-level. Firstly, we proposed hybrid features that exploit both user ratings for high rated items and some content descriptions of the items. At the model-level we built a user model from the set of hybrid features and DMF profile. Hybridization between model-based and memory-based algorithms of CF is done at the CF algorithm-level. The user model is used to find a set of like-minded users within which a memory￾based search is carried out. This set is much smaller in size than the original set, thus making the technique scalable. Finally, we developed a hybrid fuzzy-genetic RS by employing GA to evolve appropriate weights for each fea￾ture of the user model and proposing a novel fuzzy distance metric to match users. The contributions of this paper are three-fold: • A novel user model is built that enables hybrid filtering, reduces the system complexity and the computational time by roughly a factor of six. • A novel fuzzy distance function is proposed for users matching. • A hybrid fuzzy-genetic approach to recommender sys￾tems is developed. The rest of this paper is organized as follows: some background on the RS is given in the next section. In Sec￾tion 3, the proposed user model is presented, while the fuzzy and hybrid fuzzy-genetic approaches are given in Sec￾tion 4. The experimental results of the proposed approaches and classical approach are discussed in Section 5. Finally, in the last section we conclude our work with a review of our contributions along with some future research directions. 2. Background 2.1. Recommender systems Recommender systems have gained an increasing importance since the early work on CF in the mid-1990s when researchers started focusing on RS that explicitly rely on the ratings structure (Adomavicius & Tuzhilin, 2005). Normally, explicit ratings from users are binary rat￾ings (like/dislike) or follow a specified numerical scale indicating the degree of preference (e.g., 1 – bad to MAX – excellent, where MAX is the maximum possible rating for a given system). Usually, one of the following information filtering techniques is employed to generate recommendations: • Demographic filtering (DMF): The user will be recom￾mended items similar to the ones other people with same M.Y.H. Al-Shamri, K.K. Bharadwaj / Expert Systems with Applications 35 (2008) 1386–1399 1387