Experimentt-fixed 10 users Experiment2. fixed 50 users d付 6: Results for Experiment3.random 10 Experiment 4. random 5o users 35701131517 Figure 7: Results for experiment 3 Figure 8: Results for experiment 4 Experiment 1: Each of the first 10 users was picked as the to the pearson algorithm on 8 active users out of 10 active user in turn, and the first 10 users(fixed) were used Figure 6 shows that in the second experiment, out of the to provide recommendations 50 users the accuracy for the ga recommender fell below Experiment 2: Each of the first 50 users was picked as the that of the Pearson algorithm for 14 active users. On the active user in turn, and the first 50 users( fixed) were used rest of the active users, the accuracy for the ga to provide recommendations recommender was found to be better in some cases the Experiment 3: Each of the first 10 users was picked as the difference was as great as 31% active user in turn, and 10 users were picked randomly and The random sampling for experiment 3 showed great used to provide recommendations(same 10 used per run) improvement on the prediction accuracy for the ga Experiment 4: Each of the first 50 as picked as the recommender, see figure 7. All 10 active users performe active user in turn, and 50 users were picked randomly and better than the Pearson algorithm used to provide recommendations(same 50 used per run The results for the last experiment show that the accuracy for the ga recommender was significantly better 3.1. Results for all but 4 active users Figures 5 to 8 show the results for experiments 1 to 4, 3.2. Analysis of Results espectively. Each graph shows the percentage of the number of ratings that the system predicted correctly out Figure 5 indicates that the prediction accuracy for the of the total number of available ratings by the current active user 3 and 8 on the ga recommender was worse active user. Whilst the predictions computed with the than that obtained from using the Pearson algorithm. But Pearson algorithm al ways remain the same given the same when the number of users was increased to 50 in parameter values, those obtained from the GA vary experiment 2, the accuracy for the two mentioned active according to the feature weights of that run Out of the 30 users rose and outperformed the other algorithm. This was runs for each active user in each experiment, the run with expected-as the number of users goes up, the probability the best feature weights(that gave the highest percentage of finding a better matched profile should be higher and of right predictions) was chosen and plotted against the hence accuracy of the predictions should also increase result from the Pearson algorithm The patterns in both experiments 3 and 4 for the active Figure 5 shows that in the first experiment, the Ga users I to 10 look very similar. Both show an improved recommender performed equally well (or better)compared accuracy compared to the Pearson algorithm but in experiment 4 there seems to be a greater improvement The best rather than average was plotted since this is closest to Again, this is likely to be because of the increase in the the real world scenario where this system could be run off-line number of users. The results suggest that random sampling and the current best set of feature weights would be set as the is a good choice for the profile selection task of retrieving initial preference of the active user. The evolved weights could en be stored on the users local machine. A local copy of the expected to be better than fixing which users to select stem would be responsible for fine-tuning the weights to suit because it allowed the search to consider a greater variety that user's preferences further. This way the processing load or of profiles (potentially 10 30 runs 300 users in the server would be reduced and parallelism can be achievedExperiment 1: Each of the first 10 users was picked as the active user in turn, and the first 10 users (fixed) were used to provide recommendations. Experiment 2: Each of the first 50 users was picked as the active user in turn, and the first 50 users (fixed) were used to provide recommendations. Experiment 3: Each of the first 10 users was picked as the active user in turn, and 10 users were picked randomly and used to provide recommendations (same 10 used per run). Experiment 4: Each of the first 50 users was picked as the active user in turn, and 50 users were picked randomly and used to provide recommendations (same 50 used per run). 3.1. Results Figures 5 to 8 show the results for experiments 1 to 4, respectively. Each graph shows the percentage of the number of ratings that the system predicted correctly out of the total number of available ratings by the current active user. Whilst the predictions computed with the Pearson algorithm always remain the same given the same parameter values, those obtained from the GA vary according to the feature weights of that run. Out of the 30 runs for each active user in each experiment, the run with the best feature weights (that gave the highest percentage of right predictions) was chosen and plotted against the result from the Pearson algorithm.1 Figure 5 shows that in the first experiment, the GA recommender performed equally well (or better) compared 1 The best rather than average was plotted since this is closest to the real world scenario where this system could be run off-line and the current best set of feature weights would be set as the initial preference of the active user. The evolved weights could then be stored on the user’s local machine. A local copy of the system would be responsible for fine-tuning the weights to suit that user's preferences further. This way the processing load on the server would be reduced and parallelism can be achieved. to the Pearson algorithm on 8 active users out of 10. Figure 6 shows that in the second experiment, out of the 50 users the accuracy for the GA recommender fell below that of the Pearson algorithm for 14 active users. On the rest of the active users, the accuracy for the GA recommender was found to be better – in some cases the difference was as great as 31%. The random sampling for experiment 3 showed great improvement on the prediction accuracy for the GA recommender, see figure 7. All 10 active users performed better than the Pearson algorithm. The results for the last experiment show that the accuracy for the GA recommender was significantly better for all but 4 active users, see figure 8. 3.2. Analysis of Results Figure 5 indicates that the prediction accuracy for the active user 3 and 8 on the GA recommender was worse than that obtained from using the Pearson algorithm. But when the number of users was increased to 50 in experiment 2, the accuracy for the two mentioned active users rose and outperformed the other algorithm. This was expected – as the number of users goes up, the probability of finding a better matched profile should be higher and hence accuracy of the predictions should also increase. The patterns in both experiments 3 and 4 for the active users 1 to 10 look very similar. Both show an improved accuracy compared to the Pearson algorithm but in experiment 4 there seems to be a greater improvement. Again, this is likely to be because of the increase in the number of users. The results suggest that random sampling is a good choice for the profile selection task of retrieving profiles from the database. Random sampling was expected to be better than fixing which users to select because it allowed the search to consider a greater variety of profiles (potentially 10*30 runs = 300 users in Experiment1 – fixed 10 users 0 20 40 60 80 100 1 2 3 4 5 6 7 8 9 10 Active User Pearson GA Recommender Figure 5: Results for experiment 1 Experiment2 - fixed 50 users 0 20 40 60 80 100 13579 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 Active User Pearson nGA Recommender Figure 6: Results for experiment 2 Experiment3 - random 10 users 0 20 40 60 80 100 1 2 3 4 5 6 7 8 9 10 Active User Pearson GA Recommender Figure 7: Results for experiment 3 Experiment 4 - random 50 users 0 20 40 60 80 100 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 Pearson GA Recommender Figure 8: Results for experiment 4