ARTICLE N PRESS S.K. Shinde, U. Kulkarmi/ Expert Systems with Applications xx(2011)xXx-xxx e aver ge quality of an in where Ru(t) is the rating given to the item ty by the user u, Ru(t)is Precision Size of hit set test n top- (9) the predicted value of the user u on the item t,, Tis the test set and IT] is the size of the test set. In addition to CBBCHPRS, ARS (Bedi et al 2009)is also imple We varied the number of rated items provided to the active user mented to compare the performance with our proposed system. from 3, 5, and 10. In order to compare the performance of the pro- The recommendation list is evaluated using precision and recall posed system with ARS, we empirically analyze how MAE evolves for different number of clusters as shown in the Figs. 2 and 3. with the density of rating matrix. The results obtained are shown To evaluate prediction quality metric, we have used the mean in the Fig. 4 absolute error(MAE), a statistical accuracy metric, (Herlocke In this experiment, we have randomly selected 5%, 10%, 15% et al, 2004; Zhao Bing, 2005), which is computed as 20%, 25%, 30%, 35% and 40% of whole rating data from the database to represent different degrees of density of the rating matrix. the number of nearest neighbours is set to 4 while the number of rated items for active user is set to 3. The results show that the density has a great effect on the performance of these algorithms. When the rating matrix becomes crowded all the algorithms result in higher MAE. As seen from Fig 4, the mae curve of our algorithm is below that of the other algorithm, which means that the sparse- ness has the less impact on our proposed algorithm 6. Conclusions This paper describes a novel personalized recommender system that utilizes clustering of user-item rating matrix through pro- sed CBBC and provides the recommendations for the active user with good quality rating using similarity measures. The result from various simulations using Iris data set shows that the proposed Recall of ARs clustering algorithm performs better than K-means and new Recall of CBBCHPRs K-medoid clustering, which helps to improve the quality of rating. Precision of ARs In traditional recommender system similarity is normally the Precision of CBBCHPRS only heuristic used in recommendation process where as in the clusters. This helps in the exploration of other clusters which have hilarity closer to the active user and provide him/ her with good set of recommendations emendation set ion and recall is clusters created in CBBCHPRS and ARS. It is found that the pro Fig. 3. Precision and Recall for 20 clusters. posed CBBCHPRS performs better than ARS and the sparseness has less impact on the proposed algorithm. References Adomavicius, G, Tuzhilin, A (2005) Toward the next ge on of recommender Transactions Knowledge Database Engineering. 17(6). 734-774. Al-Daoud, M. B,& Roberts, s.A(1996- w methods for the initialization User models: theory, method and practice. Intemational Joumal 0.78 V CBBCHPRS mender system based collaborative ants. Journal 0.74 Elaborative on. IEEE Transactions ont Systems, Man and Cybernetics-Part A: ystems and Humans, 34(1). 143-148. Chun Zeng. et al.(2002). Personalized services for digital library. In proce Jun(2009) A simple and fast algorithm for K- clustering Expert Systems with Applications, 3336-3341 Riedl, ]- T.(2004). filtering recommender systems. ACM Transactions on information ystems(TOIS), 22(1)5-53 Fig 4. MAE on each Item for different algorithms (A small value means a better conference on research and development in information retrieval, (pp 145-153). Please cite this article in press as: Shinde, S K.& Kulkarni, U Hybrid personalized recommender system using centering-bunching based clustering alg rithm. Expert Systems with Applications(2011). doi: 10. 1016/j eswa. 2011.08.020The precision when referring to recommender systems can be de- fined as the ratio of hit set size over the top-N set size. It gives the average quality of an individual recommendation. Precision ¼ Size of hit set Size of top-N set ¼ jtest \ top-Nj N ð9Þ In addition to CBBCHPRS, ARS (Bedi et al., 2009) is also imple￾mented to compare the performance with our proposed system. The recommendation list is evaluated using precision and recall for different number of clusters as shown in the Figs. 2 and 3. To evaluate prediction quality metric, we have used the mean absolute error (MAE), a statistical accuracy metric, (Herlocker et al., 2004; Zhao & Bing, 2005), which is computed as, MAE ¼ P u2T jRuðtjÞ e RuðtjÞj jTj ð10Þ where Ru(tj) is the rating given to the item tj by the user u; e RuðtjÞ is the predicted value of the user u on the item tj, T is the test set and jTj is the size of the test set. We varied the number of rated items provided to the active user from 3, 5, and 10. In order to compare the performance of the pro￾posed system with ARS, we empirically analyze how MAE evolves with the density of rating matrix. The results obtained are shown in the Fig. 4. In this experiment, we have randomly selected 5%, 10%, 15%, 20%, 25%, 30%, 35% and 40% of whole rating data from the database to represent different degrees of density of the rating matrix. The number of nearest neighbours is set to 4 while the number of rated items for active user is set to 3. The results show that the density has a great effect on the performance of these algorithms. When the rating matrix becomes crowded, all the algorithms result in higher MAE. As seen from Fig. 4, the MAE curve of our algorithm is below that of the other algorithm, which means that the sparse￾ness has the less impact on our proposed algorithm. 6. Conclusions This paper describes a novel personalized recommender system that utilizes clustering of user-item rating matrix through pro￾posed CBBC and provides the recommendations for the active user with good quality rating using similarity measures. The result from various simulations using Iris data set shows that the proposed clustering algorithm performs better than K-means and new K-medoid clustering, which helps to improve the quality of rating. In traditional recommender system similarity is normally the only heuristic used in recommendation process where as in the proposed CBBCHPRS, similarity is combined with density of the clusters. This helps in the exploration of other clusters which have similarity closer to the active user and provide him/her with good set of recommendations. The precision and recall is compared by varying the number of clusters created in CBBCHPRS and ARS. It is found that the pro￾posed CBBCHPRS performs better than ARS and the sparseness has less impact on the proposed algorithm. References Adomavicius, G., & Tuzhilin, A. (2005). Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE Transactions Knowledge Database Engineering, 17(6), 734–774. AI-Daoud, M. B., & Roberts, S. A. (1996). New methods for the initialization of clusters. Pattern Recognition Letters, 17, 451–455. Allen, R. B. (1990). User models: theory, method and practice. International Journal of Man-Machine Studies, 43(11), 27–52. Balabanovic, M., & Sholam, Y. (1997). Combining content-based and collaborative recommendation. Communications of ACM, 40(3), 46–61. Basu. C. H., et al. (1998). Recommendation as classification: Using social and content–based information in recommendation. In Proceedings 15th international conference on artificial intelligence, (pp. 714–720). Bedi, P. et al. (2009). Recommender system based collaborative ants. Journal on Artificial Intelligence, 2(2), 40–55. Billsus, D., & Pazzani, M., (1998). Learning collaborative information filter. In Proceedings 5th international conference on machine learning, (pp. 46–54). Cheung, K. W., & Tsui, K. Ch. (2004). Extended latent class models for collaborative recommendation. IEEE Transactions on Systems, Man and Cybernetics-Part A: Systems and Humans, 34(1), 143–148. Chun Zeng, et al. (2002). Personalized services for digital library. In Proceedings of 5th international conference on asian digital libraries, (pp. 252–253). Hae-Sang & Chi-Hyuck Jun (2009). A simple and fast algorithm for K-medoids clustering. Expert Systems with Applications, 3336–3341. Herlocker, J. L., Konstan, J. A., Terveen, L. G., & Riedl, J. T. (2004). Evaluating collaborative filtering recommender systems. ACM Transactions on Information Systems (TOIS), 22(1), 5–53. Hofmann, T., (2003). Collaborative filtering via Gaussian probabilistic latent semantic analysis. In Proceedings of the 26th annual international ACM SIGIR conference on research and development in information retrieval, (pp. 145–153). 5 10 15 20 25 30 10 20 30 40 50 60 70 80 Precision, Recall Recommendation set Recall of ARS Recall of CBBCHPRS Precision of ARS Precision of CBBCHPRS Fig. 3. Precision and Recall for 20 clusters. 10 20 30 40 50 60 0.7 0.72 0.74 0.76 0.78 0.8 0.82 MAE Size of Neighbour set ARS CBBCHPRS Fig. 4. MAE on each Item for different algorithms. (A small value means a better performance). 6 S.K. Shinde, U. Kulkarni / Expert Systems with Applications xxx (2011) xxx–xxx Please cite this article in press as: Shinde, S. K., & Kulkarni, U. Hybrid personalized recommender system using centering-bunching based clustering algo￾rithm. Expert Systems with Applications (2011), doi:10.1016/j.eswa.2011.08.020