Effect of suppression on number of recommend a small amount of suppression may indeed be beneficial to AIs performance, in particular recommendation. It is interesting to note that the increase in recommendation quality occurs with a relatively constant overlap size. At too high levels of suppression, it is harder to fill the neighbourhood, with consequent lack of diversity and hence ecommendation accuracy We believe that these initial results show two things Firstly, population effects can be beneficial for CI 002040608 algorithms, particularly for recommendation; secondly, that CF is a promising new application area for artificial immune Figure 4: Effect of suppression rate on prediction and recommendation systems. In fact, we can widen the context, since the process of neighbourhood selection described in this paper can easily Again, the graphs show averaged results over five runs at be generalized to the task of ad-hoc community formation. each suppression rate. The bars show standard deviations (similar size bars for rates 0.2 and 0.5 have been omitted in the interests of clarity). At low levels of stimulation, REFERENCES prediction accuracy is not significantly affected. However [1] Aggarwal C and Yu P, On Text Mining for Personalization recommendation accuracy is improved significantly (95% Lecture Notes in Artificial Intelligence, ve pp.12-18,1999 Wilcoxon). For instance, for 0.3 stimulation, rates from 0.0 Billsus, D. and Pazzani, M. J, " Leamin to 0.2 gave a significantly improved performance. In actual Filters, Proceedings of the Fifteenth onal Conference terms, the Kendall measure rises from 0.5 to nearly 0.6. This means that the chance of any two randomly sampled pairs [Breese JS, Heckerman D and Kadie C, Empirical Analysis of predictive algorithms for collaborative filtering, Proceedings of the being correctly ranked has risen from 60%to 80%. Too much suppression had a detrimental effect on all measures. [5] Compaq Systems Research Centre. Ea collaborative filtering lataset(http://www.research.compaq.com/src/eachmOvi [6 Delgado J, Ishii N and Tomoki U. Content-based Collaborative 4. CONCLUSIONS Information Filtering: Actively Learning to Classify and Recommend Documents. In: Cooperative Information Agents Il Learning, Mobility erce for Information Discovery on the Internet, It is not particularly surprising that the simple AlS performs similarly to the SP predictor. This is because they are, at their [7 Delgado J. and Ishii, Multi-agent Learning in Recommender System ore based around the same algorithm. The stimulation rate Information Systems, vol. 10, Pp. 81-100, 2001 (in absence of any idiotypic effect)is effectively setting a [ 8] Farmer JD, Packard NH and Perelson AS,The threshold for correlation. This has both strengths and adaptation, and machine learning Physica, vol. 22, pp. 187-204, 1986 weaknesses. It has been shown that a threshold is useful in 1 Fisher D, Hildrum K, Hong J, Newman M and Vuduc R, SWAMI: a discarding the potentially misleading predictions of poorly framework for collaborative filtering algorithm development and evaluation1999.http://guir.berkeley.edu/projects/swami/ correlated reviewers [10]. On the other hand, a rigid [10] Gokhale A, Improvements to Collaborative Filtering Algorithms 1999 threshold means that one has to 'prejudge' the appropriate Worcester Polytech Ite.http://www.cs.wpi.edu/-claypool/ms level to avoid both premature convergence communities. Indeed, detailed examination of the individual [11] Hightower RR, Forrest S and Perelson AS "The evolution of emergent neighbourhood either early or not at all. The setting of a四m,cmCe)P3/的 rganization in immune system gene libraries, " Proceedings of the 6th mational Confere threshold also means that sufficiently good antibodies are 1996 taken on a first come, first served basis. It is interesting to [13 Jene nK, Towards a network theory of the immune system Annals of observe that such a strategy nevertheless seems (in these Immunology, vol. 125, no C, pp 373-389, 1973 experiments) to provide a more constant level of overlaps, [14] Kirkwood E and Lewis C Understanding Medical Immunology, John and better recommendation quality wiley Sons, Chichester, 1989 The richness of our AIS model comes when we allow [15] Perelson AS and Weisbuch G, Immunology for physicists Reviews of Modern Physics, voL. 69, pp. 1219-1267, 1997 nteractions betwe antibodies. Early ualitative [16 Resnick P and Varian HR, Recommender systems Communications experimentation with the idiotypic network showed antibody the ACM, vol. 40, pp. 56-58, 1997. concentration rising and falling dynamically as the population varied. For instance, in the simple AlS, the concentration of an antibody will monotonically increase to saturation, or decrease to elimination, unaffected by the other antibodies. however there is a delicate balance to be struck between stimulation and suppression. An imbalance may lead to a loss in population size or diversity. The graphs show thatEffect of suppression on number of recommendations 0.0% 20.0% 40.0% 60.0% 80.0% 100.0% 120.0% 0 0.2 0.4 0.6 0.8 1 Suppression rate Number of recommendations (relative to baseline) Rate 0.2 Rate 0.3 Rate 0.5 Figure 4: Effect of suppression rate on prediction and recommendation. Again, the graphs show averaged results over five runs at each suppression rate. The bars show standard deviations (similar size bars for rates 0.2 and 0.5 have been omitted in the interests of clarity). At low levels of stimulation, prediction accuracy is not significantly affected. However recommendation accuracy is improved significantly (95% Wilcoxon). For instance, for 0.3 stimulation, rates from 0.05 to 0.2 gave a significantly improved performance. In actual terms, the Kendall measure rises from 0.5 to nearly 0.6. This means that the chance of any two randomly sampled pairs being correctly ranked has risen from 60% to 80%. Too much suppression had a detrimental effect on all measures. 4. CONCLUSIONS It is not particularly surprising that the simple AIS performs similarly to the SP predictor. This is because they are, at their core, based around the same algorithm. The stimulation rate (in absence of any idiotypic effect) is effectively setting a threshold for correlation. This has both strengths and weaknesses. It has been shown that a threshold is useful in discarding the potentially misleading predictions of poorly correlated reviewers [10]. On the other hand, a rigid threshold means that one has to ‘prejudge’ the appropriate level to avoid both premature convergence and empty communities. Indeed, detailed examination of the individual runs showed that the AIS had a tendency to fill its neighbourhood either early or not at all. The setting of a threshold also means that sufficiently good antibodies are taken on a first come, first served basis. It is interesting to observe that such a strategy nevertheless seems (in these experiments) to provide a more constant level of overlaps, and better recommendation quality. The richness of our AIS model comes when we allow interactions between antibodies. Early, qualitative experimentation with the idiotypic network showed antibody concentration rising and falling dynamically as the population varied. For instance, in the simple AIS, the concentration of an antibody will monotonically increase to saturation, or decrease to elimination, unaffected by the other antibodies. However, there is a delicate balance to be struck between stimulation and suppression. An imbalance may lead to a loss in population size or diversity. The graphs show that a small amount of suppression may indeed be beneficial to AIS performance, in particular recommendation. It is interesting to note that the increase in recommendation quality occurs with a relatively constant overlap size. At too high levels of suppression, it is harder to fill the neighbourhood, with consequent lack of diversity and hence recommendation accuracy. We believe that these initial results show two things. Firstly, population effects can be beneficial for CF algorithms, particularly for recommendation; secondly, that CF is a promising new application area for artificial immune systems. In fact, we can widen the context, since the process of neighbourhood selection described in this paper can easily be generalized to the task of ad-hoc community formation. REFERENCES [1] Aggarwal C and Yu P, On Text Mining Techniques for Personalization Lecture Notes in Artificial Intelligence, vol. 1711, pp. 12-18, 1999. [2] Amazon.com Recommendations (http://www.amazon.com/ /). [3] Billsus, D. and Pazzani, M. J., "Learning Collaborative Information Filters," Proceedings of the Fifteenth International Conference on Machine Learning. pp. 46-54, 1998. [4] Breese JS, Heckerman D and Kadie C, Empirical Analysis of predictive algorithms for collaborative filtering, Proceedings of the 14 Conference on Uncertainty in Reasoning, pp. 43-52, 1998. [5] Compaq Systems Research Centre. EachMovie collaborative filtering data set (http://www.research.compaq.com/SRC/eachmovie/). [6] Delgado J, Ishii N and Tomoki U. Content-based Collaborative Information Filtering: Actively Learning to Classify and Recommend Documents. In: Cooperative Information Agents II. Learning, Mobility and Electronic Commerce for Information Discovery on the Internet, ed. M. Klusch, G. W. E. Springer-Verlag, 1998. [7] Delgado J. and Ishii, Multi-agent Learning in Recommender Systems For Information Filtering on the Internet Journal of Co-operative Information Systems, vol. 10, pp. 81-100, 2001. [8] Farmer JD, Packard NH and Perelson AS, The immune system, adaptation, and machine learning Physica, vol. 22, pp. 187-204, 1986. [9] Fisher D, Hildrum K, Hong J, Newman M and Vuduc R, SWAMI: a framework for collaborative filtering algorithm development and evaluation 1999. http://guir.berkeley.edu/projects/swami/. [10] Gokhale A, Improvements to Collaborative Filtering Algorithms 1999. Worcester Polytechnic Institute. http://www.cs.wpi.edu/~claypool/ms /cf-improve/. [11] Hightower RR, Forrest S and Perelson AS. "The evolution of emergent organization in immune system gene libraries," Proceedings of the 6th International Conference on Genetic Algorithms, pp. 344--350, 1995. [12] Hunt J, King C and Cooke D, Immunizing against fraud, IEEE Colloquium on Knowledge Discovery and Data Mining, vol. 4, pp. 1-4, 1996. [13] Jerne NK, Towards a network theory of the immune system Annals of Immunology, vol. 125, no. C, pp. 373-389, 1973. [14] Kirkwood E and Lewis C. Understanding Medical Immunology, John Wiley & Sons, Chichester, 1989. [15] Perelson AS and Weisbuch G, Immunology for physicists Reviews of Modern Physics, vol. 69, pp. 1219-1267, 1997. [16] Resnick P and Varian HR, Recommender systems Communications of the ACM, vol. 40, pp. 56-58, 1997