ARTICLE IN PRESS H.-F. Wang C-T. Wu /Computers 8 Operations Research I(m)Il Difference Rate: Recall 4.2. Summary and remarks of experiments 0.70.8-09 We have conducted three experiments in the case study for the % proposed strategy-oriented operation module of a 3C retailer in Taiwan. In the first experiment, we compared the performances of CECF-C, CECF-NC, and Cf by three measures of Recall, Precision, nd Fl In Table 7 and Fig. 4, it showed that the proposed CECF-NC and CECF-C perform better than conventional CF except for 0=0 0.00 0.80100 which was the situation in which new user recommendations was simulated. it has been mentioned that cf could not recommend Difference Rate: Precision while the target user is without basket data. Besides, the results in Experiment 1 also showed slightly better performance of CECF-NC 0.8-09 as compared to CECF-C The reason is that in the relatively sparse users basket data. the non -common item set would show additional information for recommendation In Table 7, we could observe that when 0=0.7, 0.8, 0.9, the performances are almost 5% the same: when 0=0.6, the recommendation performance reaches the highest level. This phenomenon, which would be data-specific, tells that the effects from non-neighbor groups would not enhance but maintain the performance while 820.6. Difference Rate: F1 In the second experiment, while introducing the profit parameter B and user' s satisfaction level(b), we set up the recommendation environment by 0=0.6 for better and more stable performance. It is very important to note that while we introduce the profit parameter B in the recommendation process. the recommendation performance with BE(0, 1] would probably 5% decrease since the goal of recommendation was no longer to emphasize users benefits with B=0 only. Then the emphasis 0.40 0. should be on whether the service level decreases while the profit gain increases In Figs. 5 and 6, the results showed that while we Fig. 5. The difference rates while introducing profit consideration. ncreased B, the profit gained increases without losing recommendation performance. This phenomenon could be attributed to the analytical model since the recommended items are aligned with the users satisfaction level. Therefore, even if we look forward to increasing the profit gains of the supplier, we Profit increased proportion would still maintain the recommendation performance for Qol- Satisfaction LeveL:==7 =08-09 n the third experiment, we tested the effect of neighborhood size on fl measure. the three schemes all showed consistent results that the F1 values were positively related to the neighborhood sizes: and the CECF-NC outperformed CF in F1 measure. In particular, while the neighborhood sizes were small i.e. 3, 5, 7, the CECF-NC still reached higher F1 values than CF. 0.00.10.20304 0.60.7080.91.0 which shows the advantage of using the clique effects to Fig. 6. Profit gained under different B-values 5. Conclusion and further research In the field of RSs, there have been numerous studies proposed in Comparison of FI values order to find the best recommendation to users. Among those studies, CF has been regarded as the most effective method for its 0.35 recommendation accuracy and flexibility. However in practice, it is confronted with th olem that target users with rare information 0.25 could not get recommendations from the system. Although many approaches based on Cf have been proposed to pursue better 0.15 performance by increasing service levels and solving the problem of sparse data, the excessive emphasis on recommendation perfor mance would lead to overlooking the profit consideration, which 3571020 also an essential concern for an EC company In addition, a systematic and comprehensive module for an rS is + CECF-NC01240135016002420319 still lacking. In this regard, we have proposed a strategy-oriented +-CECF-NC with profit 0.12020.1168 21020.2774 operation module that could be comprehensively applied to EC Web 一CF 00250.0680.0900.101 sites as a decision support module so that the choice of various marketing strategies combining profit consideration for suppliers Fig. 7. Comparison of FI values with respect to the neighborhood sizes. and users can be developed. Consequently, under the framework of Please cite this article as: Wang H-F, Wu C-T. A strategy-oriented operation module for recommender systems in E-commerce. Computers and Operations Research(2010). doi: 10. 1016 j cor. 2010.03.011ARTICLE IN PRESS 4.2. Summary and remarks of experiments We have conducted three experiments in the case study for the proposed strategy-oriented operation module of a 3C retailer in Taiwan. In the first experiment, we compared the performances of CECF-C, CECF-NC, and CF by three measures of Recall, Precision, and F1. In Table 7 and Fig. 4, it showed that the proposed CECF-NC and CECF-C perform better than conventional CF except for y¼0, which was the situation in which new user recommendations was simulated. It has been mentioned that CF could not recommend while the target user is without basket data. Besides, the results in Experiment 1 also showed slightly better performance of CECF-NC as compared to CECF-C. The reason is that in the relatively sparse user’s basket data, the non-common item set would show additional information for recommendation. In Table 7, we could observe that when y¼0.7, 0.8, 0.9, the performances are almost the same; when y¼0.6, the recommendation performance reaches the highest level. This phenomenon, which would be data-specific, tells that the effects from non-neighbor groups would not enhance but maintain the performance while yZ0.6. In the second experiment, while introducing the profit parameter b and user’s satisfaction level ðbfj Þ, we set up the recommendation environment by y¼0.6 for better and more stable performance. It is very important to note that while we introduce the profit parameter b in the recommendation process, the recommendation performance with bAð0,1 would probably decrease since the goal of recommendation was no longer to emphasize user’s benefits with b¼0 only. Then the emphasis should be on whether the service level decreases while the profit gain increases. In Figs. 5 and 6, the results showed that while we increased b, the profit gained increases without losing recommendation performance. This phenomenon could be attributed to the analytical model since the recommended items are aligned with the user’s satisfaction level. Therefore, even if we look forward to increasing the profit gains of the supplier, we would still maintain the recommendation performance for services. In the third experiment, we tested the effect of neighborhood size on F1 measure. The three schemes all showed consistent results that the F1 values were positively related to the neighborhood sizes; and the CECF-NC outperformed CF in F1 measure. In particular, while the neighborhood sizes were small i.e. 3, 5, 7, the CECF-NC still reached higher F1 values than CF, which shows the advantage of using the clique effects to compensate for rare information. 5. Conclusion and further research In the field of RSs, there have been numerous studies proposed in order to find the best recommendation to users. Among those studies, CF has been regarded as the most effective method for its recommendation accuracy and flexibility. However, in practice, it is confronted with the problem that target users with rare information could not get recommendations from the system. Although many approaches based on CF have been proposed to pursue better performance by increasing service levels and solving the problem of sparse data, the excessive emphasis on recommendation perfor￾mance would lead to overlooking the profit consideration, which is also an essential concern for an EC company. In addition, a systematic and comprehensive module for an RS is still lacking. In this regard, we have proposed a strategy-oriented operation module that could be comprehensively applied to EC Web sites as a decision support module so that the choice of various marketing strategies combining profit consideration for suppliers and users can be developed. Consequently, under the framework of 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0.0 0.5 Profit increased proportion 0.7 0.8 0.9 -Value Satisfaction Level: 0.1 0.2 0.3 0.4 0.6 0.7 0.8 0.9 1.0 Fig. 6. Profit gained under different b-values. 3 5 7 10 20 CECF-NC 0.124 CECF-NC with profit 0.1202 CF 0.025 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 F1 values Comparison of F1 values 0.135 0.160 0.242 0.319 0.1168 0.1329 0.2102 0.2774 0.068 0.090 0.101 0.120 Fig. 7. Comparison of F1 values with respect to the neighborhood sizes. 0% 5% 10% 15% 20% 25% 30% 0.00 Difference Rate: Recall 0.7 0.8 0.9 0% 5% 10% 15% 20% 25% 30% Difference Rate: Precision 0.7 0.8 0.9 0% 5% 10% 15% 20% 25% 30% Difference Rate: F1 0.7 0.8 0.9 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00 Fig. 5. The difference rates while introducing profit consideration. H.-F. Wang, C.-T. Wu / Computers & Operations Research ] (]]]]) ]]]–]]] 11 Please cite this article as: Wang H-F, Wu C-T. A strategy-oriented operation module for recommender systems in E-commerce. Computers and Operations Research (2010), doi:10.1016/j.cor.2010.03.011