M. Y.H. Al-Shamri, KK Bharadwaj/ Expert Systems with Applications 35(2008)1386-1399 some filtering technique and then apply on the constructed 3. 1. Necessary formulae model another filtering technique? In our approaches only one online filtering process(CF)is needed, while the other The total ratings(TR)of user u; is filtering techniques (CBF and DMF) are used to dense the data by building a compact user model, which is an offline TR(=>ris (6) ocess Further, a sparse user-item matrix causes a scalability Here S; is the set of movies rated by user u The genre rat problem for CF. However, we can overcome this problem ing(GR)(resp. genre frequency(GF) for high rated items to some extent by combining or integrating many informa- of genre G; corresponding to user u; is computed using for- tion sources. In our work, we develop a set of hy brid fea- mula (7)(resp. formula( 8) tures that combines some of the users' and items properties. An example of a hybrid feature(Basu, Hirsh. GR(i,j) Cohen, 1998)would be the set of Movies of the Comedy Genre User x Highly Rated. We derive the hybrid feature GF(i,j)=25(is), kE(3, 4, 5) based on the user's ratings for a set of high rated movies s∈GCS and the content descriptions of the genres corresponding where to this set of movies. The hybrid features are utilized as the basis for formulating a genre interestingness measure S(ris) (GIM). Once we have an appropriate formula for GIM, a user model can be constructed from DMF user profile It is to be noted that only the items having rating r>3 are and GIMs considered, i.e. the items rated as good (3), very good (4), Block diagram of the proposed work is given in Fig. I. or excellent(5). Such items would be called items with high First of all, a hybrid user model is built using the hybrid ratings. Finally, relative genre rating (RGR)(resp. relative features and DMF user profile, thereafter CF recom- genre frequency(RGF)), the ratio of u;'s ratings(resp. fre- mender generates a neighborhood set according to quency) for high rated items of G; to his total ratings(resp model-based CF. Finally the entire database of this set is frequency), is computed as retrieved to recommend items according to memory-based CF. Before going to details, let us present necessary formu- RGR(i, j)=TR( GR(i,j) (10) lae for the Movie Lens dataset and illustrate GIM compu- tations through an example RGF(i,j) GF(i,j TF() where TF(=Si, the cardinality of Si 3.2. Example I CBF USer profile Content For the sake of simplicity, we consider a table(Table 1) Descriptions of only three users who have rated movies belonging to four genres. In columns Gi, i= 1, 2, 3, 4, one(1)indicates the belongingness of a given movie to G, and 0 otherwise Also a non-zero value in the user ratings columns indicates Hybrid Features Table I Data of Example 1 Hybrid User Model Movie Corresponding genres User ratings User.l User.2 User-3 CF Recommender 23456789 0 Set of Neighbors 0 G001111 Database G010010110010 3154000000 003040050 0 Recommendatio o12 1011l0 02002430 Fig. I. Block diagram of the proposed worksome filtering technique and then apply on the constructed model another filtering technique? In our approaches only one online filtering process (CF) is needed, while the other filtering techniques (CBF and DMF) are used to dense the data by building a compact user model, which is an offline process. Further, a sparse user-item matrix causes a scalability problem for CF. However, we can overcome this problem to some extent by combining or integrating many informa￾tion sources. In our work, we develop a set of hybrid fea￾tures that combines some of the users’ and items’ properties. An example of a hybrid feature (Basu, Hirsh, & Cohen, 1998) would be the set of Movies of the Comedy Genre User X Highly Rated. We derive the hybrid feature based on the user’s ratings for a set of high rated movies and the content descriptions of the genres corresponding to this set of movies. The hybrid features are utilized as the basis for formulating a genre interestingness measure (GIM). Once we have an appropriate formula for GIM, a user model can be constructed from DMF user profile and GIMs. Block diagram of the proposed work is given in Fig. 1. First of all, a hybrid user model is built using the hybrid features and DMF user profile, thereafter CF recom￾mender generates a neighborhood set according to model-based CF. Finally the entire database of this set is retrieved to recommend items according to memory-based CF. Before going to details, let us present necessary formu￾lae for the MovieLens dataset and illustrate GIM compu￾tations through an example. 3.1. Necessary formulae The total ratings (TR) of user ui is TRðiÞ ¼ X s2Si ri;s: ð6Þ Here Si is the set of movies rated by user ui. The genre rat￾ing (GR) (resp. genre frequency (GF)) for high rated items of genre Gj corresponding to user ui is computed using for￾mula (7) (resp. formula (8)) GRði;jÞ ¼ X s2GjSi;rP3 ri;s; ð7Þ GFði;jÞ ¼ X s2GjSi dk ðrisÞ; k 2 f3; 4; 5g; ð8Þ where, dk ðri;sÞ ¼ 1; k ¼ ri;s; 0; k 6¼ ri;s: ð9Þ It is to be noted that only the items having rating r P 3 are considered, i.e. the items rated as good (3), very good (4), or excellent (5). Such items would be called items with high ratings. Finally, relative genre rating (RGR) (resp. relative genre frequency (RGF)), the ratio of ui’s ratings (resp. fre￾quency) for high rated items of Gj to his total ratings (resp. frequency), is computed as RGRði;jÞ ¼ GRði;jÞ TRðiÞ ; ð10Þ RGFði;jÞ ¼ GFði;jÞ TFðiÞ ; ð11Þ where TF(i) = jSij, the cardinality of Si. 3.2. Example 1 For the sake of simplicity, we consider a table (Table 1) of only three users who have rated movies belonging to four genres. In columns Gi, i = 1, 2, 3, 4, one (1) indicates the belongingness of a given movie to Gi, and 0 otherwise. Also a non-zero value in the user ratings columns indicates CBF User Profile DMF User Profile Content Descriptions Explicit Ratings Hybrid Features Hybrid User Model CF Recommender Recommendations Set of Neighbors Database Fig. 1. Block diagram of the proposed work. Table 1 Data of Example 1 Movie Corresponding genres User ratings G1 G2 G3 G4 User-1 User-2 User-3 1 1010150 2 1011300 3 1100153 4 0110510 5 011 1 4 0 4 6 0100020 7 1101000 8 001 1 0 0 5 9 0110020 10 1 1 1 0 0 4 1 11 1 1 0 1 0 3 3 12 1 0 0 0 0 0 3 1390 M.Y.H. Al-Shamri, K.K. Bharadwaj / Expert Systems with Applications 35 (2008) 1386–1399