M.Y.H. Al-Shamri, K.K. Bharadwaj/ Expert Systems with Applications 35(2008)1386-1399 1397 used in the implementation, using 8-bits for each of the 21 but with a little difference in their priorities. Looking at the genes. The Ga begins with random genotypes. A decimal final weights obtained for each active user, many interest value for each weight is obtained by converting the alleles ing observations are made. Some users give more emphasis phenotype is then calculated by summing up the decimal interest for some features. Let us focus our discussion only values for all the 21 weights. Finally, by dividing the deci- on two splits, split-l(Fig. 4)and split-3(Fig. 5), and con mal value of each weight by the total value, the weighting sider the following observations value for each feature can be found training ratings of the active user. The evolved set of By degrading the effect of the gender feature this value hts is employed to find the predicted ratings for each further increased to 47.6% using FGRS movie in the training ratings set. The average of the differ-. When all features contributing equally, the correct pre- ences between the actual and predicted ratings of all movies dictions for the active user 18(split-3)are smaller in the training set is used as the fitness score for that set of(40.64%)than that of Prs(47.59%). However, by ignor weights. The fitness function used is the same as given in ing the gender feature and emphasizing more on the Section 4.2.2 occupation, the correct predictions increased to 47. 10% The population of size 10 is kept fixed for all the exper- For the active users 8(split-1)and 10 (split-3), the cor- iments. Ten more individuals are generated from genera- rect predictions of FRS are higher than that of PRS tion to generation. The best ones from these new 10 Further by ignoring the age these values increased to individuals replace the bad ones(if exist) of the previous 32.06% and 51.14%, respectivel fixed size population individuals. Parents are selected ran -. FRS performed better than PRS for the active user 1l domly from the top 8 individuals to generate 8 new individ (split-3). However, by emphasizing more on the demo- gle point crossover. The mutation is applied graphical data, the correct predictions increased to to generate the remaining two individuals 32.93% The predicted rating for a given movie in the test ratings By incorporating the demographical data for the active set is computed using formula(4)over the set of all neigh ser 37(split-3), the correct predictions are higher bors who have rated that movie. The predicted ratings (31.58%)for FRS as compared to PRS (25%). However, produced by the FRS and the hybrid FGRS, are compared by ignoring the demographical data and relying only on against those produced by the PRS as shown in Figs 4 and GIMs this value deceased to 28.95% using FGrs By ignoring the age and occupation features of the active user 43(split-3), the correct predictions of FGRS 5.2.1. Analysis of the results exceed that of Frs and Prs For FGrS, out of 10 runs for each active user, the run For the active user 10(split-l), the correct predictions with the best weights was chosen and plotted against the are higher(3952%) for FRs as compared to PRS results from the PRS and FRS as shown in Figs 4 and 5 (29.84%) Further, by emphasizing more on the age this for split-l and split-3, respectively. Results summarized in alue increased to 40. 32% Table 6 show that FGRS outperforms PRS and FRS for The feature weights of the active user 28(split-1)are all the five splits, where it is greater than PRS(FRS) values interesting. They show that when all features are con- on 92%(94%)for split-3 and 96%(76%)for split-1. The tributing equally using FRS, the number of correct pre- MAE for FGRS is always smaller than the corresponding dicted movies increases only by 4. As soon as the gender values for PRS and FRs, while the coverage value is feature is ignored and the age takes high weight, FGRS always greater for all the splits outperforms the Prs by 19 movies. The main features that reflect users' preferences are dif -. The correct predictions of the active user 32(split-1)are ferent for different users. For example, some users rely smaller(33.6%)under FRs than that of PRS (36.8%) mainly on their explicit ratings, others on the similar age However, when the weight of the gender feature gets and gender groups, whereas some others rely on all features higher, this value increased to 40% using FGRS A summary of the results corresponding to the 7th active user of split-1 Correct predictions (% MAE Coverage(7)(%) Pearson 29.268 0.900 97.56 19-70 48.780 0.675 97.56 22-51 51.219 0.675 97.56 22-51used in the implementation, using 8-bits for each of the 21 genes. The GA begins with random genotypes. A decimal value for each weight is obtained by converting the alleles of the binary genes to decimal. The total decimal value of phenotype is then calculated by summing up the decimal values for all the 21 weights. Finally, by dividing the deci￾mal value of each weight by the total value, the weighting value for each feature can be found. A supervised learning is used to learn weights. The GA learns weights using guidance of the actual ratings in the training ratings of the active user. The evolved set of weights is employed to find the predicted ratings for each movie in the training ratings set. The average of the differ￾ences between the actual and predicted ratings of all movies in the training set is used as the fitness score for that set of weights. The fitness function used is the same as given in Section 4.2.2. The population of size 10 is kept fixed for all the exper￾iments. Ten more individuals are generated from genera￾tion to generation. The best ones from these new 10 individuals replace the bad ones (if exist) of the previous fixed size population individuals. Parents are selected ran￾domly from the top 8 individuals to generate 8 new individ￾uals using a single point crossover. The mutation is applied to generate the remaining two individuals. The predicted rating for a given movie in the test ratings set is computed using formula (4) over the set of all neigh￾bors who have rated that movie. The predicted ratings, produced by the FRS and the hybrid FGRS, are compared against those produced by the PRS as shown in Figs. 4 and 5. 5.2.1. Analysis of the results For FGRS, out of 10 runs for each active user, the run with the best weights was chosen and plotted against the results from the PRS and FRS as shown in Figs. 4 and 5 for split-1 and split-3, respectively. Results summarized in Table 6 show that FGRS outperforms PRS and FRS for all the five splits, where it is greater than PRS (FRS) values on 92% (94%) for split-3 and 96% (76%) for split-1. The MAE for FGRS is always smaller than the corresponding values for PRS and FRS, while the coverage value is always greater for all the splits. The main features that reflect users’ preferences are dif￾ferent for different users. For example, some users rely mainly on their explicit ratings, others on the similar age and gender groups, whereas some others rely on all features but with a little difference in their priorities. Looking at the final weights obtained for each active user, many interest￾ing observations are made. Some users give more emphasis on specified features more, while others do not show any interest for some features. Let us focus our discussion only on two splits, split-1 (Fig. 4) and split-3 (Fig. 5), and con￾sider the following observations: • The correct predictions of the active user 15 (split-1) are higher (43.81%) for FRS as compared to PRS (27.62%). By degrading the effect of the gender feature this value further increased to 47.6% using FGRS. • When all features contributing equally, the correct pre￾dictions for the active user 18 (split-3) are smaller (40.64%) than that of PRS (47.59%). However, by ignor￾ing the gender feature and emphasizing more on the occupation, the correct predictions increased to 47.10%. • For the active users 8 (split-1) and 10 (split-3), the cor￾rect predictions of FRS are higher than that of PRS. Further by ignoring the age these values increased to 32.06% and 51.14%, respectively. • FRS performed better than PRS for the active user 11 (split-3). However, by emphasizing more on the demo￾graphical data, the correct predictions increased to 32.93%. • By incorporating the demographical data for the active user 37 (split-3), the correct predictions are higher (31.58%) for FRS as compared to PRS (25%). However, by ignoring the demographical data and relying only on GIMs this value deceased to 28.95% using FGRS. • By ignoring the age and occupation features of the active user 43 (split-3), the correct predictions of FGRS exceed that of FRS and PRS. • For the active user 10 (split-1), the correct predictions are higher (39.52%) for FRS as compared to PRS (29.84%). Further, by emphasizing more on the age this value increased to 40.32%. • The feature weights of the active user 28 (split-1) are interesting. They show that when all features are con￾tributing equally using FRS, the number of correct pre￾dicted movies increases only by 4. As soon as the gender feature is ignored and the age takes high weight, FGRS outperforms the PRS by 19 movies. • The correct predictions of the active user 32 (split-1) are smaller (33.6%) under FRS than that of PRS (36.8%). However, when the weight of the gender feature gets higher, this value increased to 40% using FGRS. Table 9 A summary of the results corresponding to the 7th active user of split-1 RS Correct predictions (%) MAE (7) Coverage (7) (%) Neighbors Age range Male Female Pearson 29.268 0.900 97.56 23 7 19–70 Fuzzy 48.780 0.675 97.56 1 29 22–51 Fuzzy-genetic 51.219 0.675 97.56 0 30 22–51 M.Y.H. Al-Shamri, K.K. Bharadwaj / Expert Systems with Applications 35 (2008) 1386–1399 1397