experiment 3 and 50* 30= 1500 users in experiment 4) features: rating and age. It is also clear that this user does and hence find a better set of well matched profiles not show any interest in the 3d feature which is gender. So As mentioned earlier, only the run(s) with the best as long as the people that are giving him recommendations feature weights for each active user were considered for have similar opinions and are in the same age group as this analysis. We now look into these runs in more detail him, he does not care whether they are male or female to see how the feature weights obtained and users selected The feature weights obtained for active user 8 were also for the neighbourhood in these runs played a part in interesting. They show that for this user, age and gender determining user preference are more significant. By looking further at the movie Looking at experiment 1, when more than I run for an genres, we found that people who have similar opinions as active user achieved the same best performance(highest this user on action, adventure, horror, romantic and war number of votes being predicted correctly)results indicate movies are likely to be picked for the neighbourhood set that the same set of users had been selected for the As these genres are stereotypically related to gender and neighbourhood to give recommendations. Moreover, for age, for example, men prefer action movies and war other runs that did not perform as well as the best run(s), movies, the weights showed consistent description of the different users that gave the best performance had been user's preference. Another example is active user 7 whose selected. For example, for active user 2 in experiment 1, weights show strong feelings for documentary, mystery, all the runs that got the same percentage as the best, chose sci-fi and thriller genres and emphasis on age. This user is user 4 to be in the neighbourhood. The other active users a 57 year old male which may explain reduced did not select any users to give recommendations, instead significance of children and romance ger nre the mean vote was used. Data gathered during experiment From the observations above. we can see that age is 2 corroborates this view. In addition, as the number of often as or more important as rating. This shows that the users was increased, the users that were originally selected theory behind the original collaborative filtering does not for the neighbourhood in experiment I were still being al ways hold. This is hardly surprising as everyday chosen in experiment 2 as a subset of a larger experience suggests that most people listen to the neighbourhood. For example, as mentioned above, in recommendations made by their friends who are most experiment 1 user 2 picked user 4 to be in the likely to be in the same age group as ther neighbourhood, in experiment 2 this user picked users 4, 13, 18, 22, 42, 43, 49. This, however, only applies to the 4. CONCLUSIONS active users that performed better than the Pearson algorithm in experiment 1. The accuracy for active user 8 This work has shown how evolutionary search can be was worse in experiment 1, in which users 4, 5, 7 and 10 employed to fine-tune a profile-matching algorithm within were selected In experiment 2 when users 4 and 10 were a recommender system, tailoring it to the preferences of not included in the neighbourhood, the accuracy improved individual users. This was achieved by reformulating the tremendously as seen in figure 6. The trend described problem of making recommendations into a supervised not be observed when random sampling was used in learning task, enabling fitness scores to be computed by ments 3 and 4, as it was more difficult for the comparing predicted votes with actual votes. Experiments to select the same users to examine at each run demonstrated that, compared to a non-adaptive approach, the evolutionary recommender system was able to Feature weights for active user 2 successfully fine-tune the profile matching algorithm. This enabled the recommender system to make more accurate predictions, and hence better recommendations to users References [1 Schafer, J B, Konstan, J. A. and Riedl, J. January 2001.E- 123456789101112131415161718192021 ommerce Recommendation Applications. Journal of Data Figure 9: feature weights for active user 2, note that weights 5 to 22 are Mining and Knowledge Discovery lower because of the scaling factor [2] Schafer, J B, Konstan, J and Riedl, J. 1999. Recommender ystems in E-Commerce. Proc. of the ACM 1999 Confon Looking at the final feature weights obtained for each [3] Breese, J.S., Heckerman, D. and Kadie, C. 1998. Empirical active user, many interesting observations have e been analysis of predictive algorithms for collaborative filtering found. Here we focus on the first 2 experiments as they In Proc. of the 14th Conf on Uncertainty in Al, pp. 43-52 ive 10 common active users. Firstly, in experiment 2 [4] Herlocker, J.L., Konstan, J.A. Riedl,J. 2000. Explaining when more than I run came up with the best performance Collaborative Filtering Recommendations. Proc. of ACM the feature weights seem to show very similar trends. For 2000 Conf on Computer Supported Cooperative Work example, figure 9 shows the weight emphasis on the first 2experiment 3 and 50 * 30 = 1500 users in experiment 4) and hence find a better set of well matched profiles. As mentioned earlier, only the run(s) with the best feature weights for each active user were considered for this analysis. We now look into these runs in more detail to see how the feature weights obtained and users selected for the neighbourhood in these runs played a part in determining user preference. Looking at experiment 1, when more than 1 run for an active user achieved the same best performance (highest number of votes being predicted correctly) results indicate that the same set of users had been selected for the neighbourhood to give recommendations. Moreover, for other runs that did not perform as well as the best run(s), different users that gave the best performance had been selected. For example, for active user 2 in experiment 1, all the runs that got the same percentage as the best, chose user 4 to be in the neighbourhood. The other active users did not select any users to give recommendations, instead the mean vote was used. Data gathered during experiment 2 corroborates this view. In addition, as the number of users was increased, the users that were originally selected for the neighbourhood in experiment 1 were still being chosen in experiment 2 as a subset of a larger neighbourhood. For example, as mentioned above, in experiment 1 user 2 picked user 4 to be in the neighbourhood, in experiment 2 this user picked users 4,13,18,22,42,43,49. This, however, only applies to the active users that performed better than the Pearson algorithm in experiment 1. The accuracy for active user 8 was worse in experiment 1, in which users 4, 5, 7 and 10 were selected. In experiment 2 when users 4 and 10 were not included in the neighbourhood, the accuracy improved tremendously as seen in figure 6. The trend described could not be observed when random sampling was used in experiments 3 and 4, as it was more difficult for the system to select the same users to examine at each run. Figure 9: feature weights for active user 2, note that weights 5 to 22 are lower because of the scaling factor Looking at the final feature weights obtained for each active user, many interesting observations have been found. Here we focus on the first 2 experiments as they have 10 common active users. Firstly, in experiment 2 when more than 1 run came up with the best performance, the feature weights seem to show very similar trends. For example, figure 9 shows the weight emphasis on the first 2 features: rating and age. It is also clear that this user does not show any interest in the 3rd feature which is gender. So as long as the people that are giving him recommendations have similar opinions and are in the same age group as him, he does not care whether they are male or female. The feature weights obtained for active user 8 were also interesting. They show that for this user, age and gender are more significant. By looking further at the movie genres, we found that people who have similar opinions as this user on action, adventure, horror, romantic and war movies are likely to be picked for the neighbourhood set. As these genres are stereotypically related to gender and age, for example, men prefer action movies and war movies, the weights showed consistent description of the user’s preference. Another example is active user 7 whose weights show strong feelings for documentary, mystery, sci-fi and thriller genres and emphasis on age. This user is a 57 year old male which may explain reduced significance of children and romance genres. From the observations above, we can see that age is often as or more important as rating. This shows that the theory behind the original collaborative filtering does not always hold. This is hardly surprising as everyday experience suggests that most people listen to the recommendations made by their friends who are most likely to be in the same age group as them. 4. CONCLUSIONS This work has shown how evolutionary search can be employed to fine-tune a profile-matching algorithm within a recommender system, tailoring it to the preferences of individual users. This was achieved by reformulating the problem of making recommendations into a supervised learning task, enabling fitness scores to be computed by comparing predicted votes with actual votes. Experiments demonstrated that, compared to a non-adaptive approach, the evolutionary recommender system was able to successfully fine-tune the profile matching algorithm. This enabled the recommender system to make more accurate predictions, and hence better recommendations to users. References [1] Schafer, J.B., Konstan, J. A. and Riedl, J. January 2001. E￾Commerce Recommendation Applications. Journal of Data Mining and Knowledge Discovery. [2] Schafer, J.B., Konstan, J. and Riedl, J. 1999. Recommender Systems in E-Commerce. Proc. of the ACM 1999 Conf. on Electronic Commerce. [3] Breese, J.S., Heckerman, D. and Kadie, C. 1998. Empirical analysis of predictive algorithms for collaborative filtering. In Proc. of the 14th Conf. on Uncertainty in AI, pp. 43-52. [4] Herlocker, J.L., Konstan, J. A. & Riedl, J. 2000. Explaining Collaborative Filtering Recommendations. Proc. of ACM 2000 Conf. on Computer Supported Cooperative Work. Feature weights for active user 2 0 0.05 0.1 0.15 0.2 0.25 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 feature run1 run2 run3 run4 run5 run6 run7 run8 run9