M. Y.H. Al-Shamri, KK Bharadwaj/ Expert Systems with Applications 35(2008)1386-1399 0000 10 988R588厉于导导÷寻 Active User 口PRs ■FRs 口FGRs Fig. 5. Correct predictions percentage for active users of split-3 Table 6 show that FRS outperforms PRS for all the five tent of the movies and the non-existence of neighborhood splits. Results of FRS are better than PRS values on 66% transitivity relationship in a classical PRS, prevent similar on the worst split (split-3)and 82% on the best split users from being matched unless they have rated exactly (split-1). The total MAE for FRS is always smaller than the same movies he corresponding value for PRS as shown in Table 7. whereas the coverage is greater for all the splits except 5.2. Experiment 2 split-3 where it is smaller. The higher prediction values for FRS obviously illustrate that better set of matching An elitist genetic algorithm is used in this experiment for users is found and therefore the accuracy of the rs gets evolving features weights using simple parameter values as enhanced. On the other hand, the lack of access to the con- shown in Table 8. A simple unsigned binary encoding is Table 6 Comparison between FSR with PRS, FGRS FRS with PRs FGRS with PRS FGRS with Frs Greater Same Smalle Greater Same Smaller Greater Smaller 1244 24333 Table 7 Total MAE and coverage for PRs FRs and Fgrs MAE Coverage PRS FRS FGRS PRS FRS FGRS 0.844795 768797 716272 0.791428 0.719163 9590455 663480 4.67433 0.770952 0.720282 95.7099 96.97650 Table 8 Ga parameter values used in Experiment 2 meter name Population size 10 The size of individuals in the population at each generation mination threshold 0.025 When the fitness score of the best individual is below the threshold, a good solution is found and this set of weights is used as the final result for the current run Maximum generation 30 If the number of generations reaches this value and the solution has not been found, the best individuals fo (per run) that ger neration are used as the final result Number of runs The number of times the system was run for each active userTable 6 show that FRS outperforms PRS for all the five splits. Results of FRS are better than PRS values on 66% on the worst split (split-3) and 82% on the best split (split-1). The total MAE for FRS is always smaller than the corresponding value for PRS as shown in Table 7, whereas the coverage is greater for all the splits except split-3 where it is smaller. The higher prediction values for FRS obviously illustrate that better set of matching users is found and therefore the accuracy of the RS gets enhanced. On the other hand, the lack of access to the con￾tent of the movies and the non-existence of neighborhood transitivity relationship in a classical PRS, prevent similar users from being matched unless they have rated exactly the same movies. 5.2. Experiment 2 An elitist genetic algorithm is used in this experiment for evolving features’ weights using simple parameter values as shown in Table 8. A simple unsigned binary encoding is 0 10 20 30 40 50 60 70 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 Active User % of correct predictions PRS FRS FGRS Fig. 5. Correct predictions percentage for active users of split-3. Table 6 Comparison between FSR with PRS, FGRS with PRS, and FGRS with FRS Split FRS with PRS FGRS with PRS FGRS with FRS Greater Same Smaller Greater Same Smaller Greater Same Smaller 1 41 1 8 48 0 2 38 9 3 2 37 2 11 42 4 4 44 4 2 3 33 4 13 46 1 3 47 2 1 4 35 4 11 45 2 3 39 6 5 5 29 10 11 45 2 3 48 2 0 Table 7 Total MAE and coverage for PRS, FRS, and FGRS Split MAE Coverage PRS FRS FGRS PRS FRS FGRS 1 0.805011 0.707672 0.663464 95.27965 96.82935 97.04311 2 0.844795 0.768797 0.716272 95.54466 95.88573 96.22681 3 0.791428 0.719163 0.675140 95.90455 95.77991 96.38533 4 0.776416 0.713459 0.663480 94.67433 95.02690 95.62071 5 0.770952 0.720282 0.660131 95.70990 96.97650 97.56895 Table 8 GA parameter values used in Experiment 2 Parameter name Parameter value Description Population size 10 The size of individuals in the population at each generation Termination threshold 0.025 When the fitness score of the best individual is below the threshold, a good solution is found and this set of weights is used as the final result for the current run Maximum generation (per run) 30 If the number of generations reaches this value and the solution has not been found, the best individuals for that generation are used as the final result Number of runs 10 The number of times the system was run for each active user 1396 M.Y.H. Al-Shamri, K.K. Bharadwaj / Expert Systems with Applications 35 (2008) 1386–1399