H-N. Kim et aL/ Expert Systems with Applications 38(2011 )8488-8496 8495 UTU -HITI-HVSM-O-CUM With respect to the ranking accuracy, Fig. 5 confirms that CUM ontinues outperforming the other methods on all variations of N value. For example, when N is 10, CUM obtains an RK value of 0.86 whereas uru. tl. and vsm demonstrates an rK value of 0.78, 0.71, and 0.37, respectively. The result can indicate that our 503 model provides more suitable items with a higher rank in the rec ommended set than the other methods, and consequently makes ompelling recommendations for all users. More interestingly, the simple vector model did not perform well, compared to the other methods. This result might be caused by the different degree 量量 of the density of matrices. In general, it is known that CF produces locker et al., 2004). As mentioned previously, the density of the user-item matrix R we used was 13.4% whereas that of the tag- the number of recommended items(N) item matrix Q was 3.8%. We further examined the recommendation performance for Fig 4. Comparisons of precision as the number of recommended items N increases sers who had few ratings, namely cold start users, and had lots of ratings, namely active users, in the training set. Recommender systems are generally unable to make high quality recommenda- 0.9 tions, compared to the case of active users, which is pointed out as one of the limitations. We selectively considered two groups of users who have less than 10 ratings and greater than 250 ratings. 8 And for two groups we calculated precision and ranking at top-10 i.e P@10 and rK@10)obtained by the four algorithms in order to analyze whether the differences between the two groups are sta- tistically significant or not. Fig. 6 shows the results for the cold 0.2}-是..,鲁二 start group and active group. we can see from the graphs, the results of the baseline algo- rithms demonstrated that the values of the two metrics for the cold start group were considerably low. However, our method provided quite consistent performance no matter whether they have suffi ent ratings or not. Such results were caused by the fact that the baseline methods were hard to analyze the users' propensity the number of recommended items(N) items because they did not have enough information(ratings or Fig. 5. Comparisons of ranking accuracy as the number of recommended items tags). In contrast, our collaborative user model could represent suitable preference of the users by enriching both relevant topics To experimentally evaluate the performance of top-N recom- achieved by CUM and the baseline algorithms, for the cold start mendation, we calculated PoN and RKoN of each method no mat- users, precision and ranking values of the former was found to ter whether users have sufficient ratings or not. We selectively be superior to those of the other methods. For example, CUM ob- aried the number of returned items N from 1 to 10. First, we mea- tains 10%, 15%, and 13% improvement for precision compared to ured P@N obtained by UTU. IT, VSM, and CUM UTU, ITI, and VSM, respectively. with respect to the ranking accu o Fig. 4 shows the results of precision showing how CUM outper- racy, it is clear that CUM significantly outperforms the three meth- sion values tend to decrease as the number of recommended items alleviating the problem of the cold start users and thus in p0i3 ods. This result indicates that our user model N increases. However, in the case of VSM, there is very little differ- ing the quality of item recommendations ence for different values of N. Comparing the results achieved by With respect to the active group. it can be observed that CF ap- CUM and the baseline methods, the precision value of the former proaches(i.e, UTU and ITI) provide better performance than our was found to be superior to that of the benchmark methods in method The result indicates that collaborative filtering relatively all cases. When compared to VSM, CUM is significantly more accu- works well when users have abundant rating information Compar- rate on precision. On average, on all occasions, CUM outperforms ing results for the cold start group and the active group obtained by UTU, ITI and vSM by 5%, 6. 2% and 18. 1%, respectively the Cf approaches, it is apparent that the two groups have Precision at Top 10 Ranking at Top 10 aUTU aIm DVSM O CUM 12 Iaru aIm DVSMDCUM 1.4 0.8 0.6 0.4 cold start users cold start user active users Fig. 6. Comparisons of precision and ranking accuracy for cold start users and active users.To experimentally evaluate the performance of top-N recom￾mendation, we calculated P@N and RK@N of each method no mat￾ter whether users have sufficient ratings or not. We selectively varied the number of returned items N from 1 to 10. First, we mea￾sured P@N obtained by UTU, ITI, VSM, and CUM. Fig. 4 shows the results of precision showing how CUM outper￾forms the baseline methods. The graph curves show that the preci￾sion values tend to decrease as the number of recommended items N increases. However, in the case of VSM, there is very little differ￾ence for different values of N. Comparing the results achieved by CUM and the baseline methods, the precision value of the former was found to be superior to that of the benchmark methods in all cases. When compared to VSM, CUM is significantly more accu￾rate on precision. On average, on all occasions, CUM outperforms UTU, ITI and VSM by 5%, 6.2% and 18.1%, respectively. With respect to the ranking accuracy, Fig. 5 confirms that CUM continues outperforming the other methods on all variations of N value. For example, when N is 10, CUM obtains an RK value of 0.86 whereas UTU, ITI, and VSM demonstrates an RK value of 0.78, 0.71, and 0.37, respectively. The result can indicate that our model provides more suitable items with a higher rank in the rec￾ommended set than the other methods, and consequently makes compelling recommendations for all users. More interestingly, the simple vector model did not perform well, compared to the other methods. This result might be caused by the different degree of the density of matrices. In general, it is known that CF produces good performance in situation where a rating matrix is dense (Her￾locker et al., 2004). As mentioned previously, the density of the user-item matrix R we used was 13.4% whereas that of the tag￾item matrix Q was 3.8%. We further examined the recommendation performance for users who had few ratings, namely cold start users, and had lots of ratings, namely active users, in the training set. Recommender systems are generally unable to make high quality recommenda￾tions, compared to the case of active users, which is pointed out as one of the limitations. We selectively considered two groups of users who have less than 10 ratings and greater than 250 ratings. And for two groups we calculated precision and ranking at top-10 (i.e., P@10 and RK@10) obtained by the four algorithms in order to analyze whether the differences between the two groups are sta￾tistically significant or not. Fig. 6 shows the results for the cold start group and active group. As we can see from the graphs, the results of the baseline algo￾rithms demonstrated that the values of the two metrics for the cold start group were considerably low. However, our method provided quite consistent performance no matter whether they have suffi- cient ratings or not. Such results were caused by the fact that the baseline methods were hard to analyze the users’ propensity for items because they did not have enough information (ratings or tags). In contrast, our collaborative user model could represent suitable preference of the users by enriching both relevant topics and irrelevant topics from similar neighbors. Comparing the results achieved by CUM and the baseline algorithms, for the cold start users, precision and ranking values of the former was found to be superior to those of the other methods. For example, CUM ob￾tains 10%, 15%, and 13% improvement for precision compared to UTU, ITI, and VSM, respectively. With respect to the ranking accu￾racy, it is clear that CUM significantly outperforms the three meth￾ods. This result indicates that our user model can help, indeed, in alleviating the problem of the cold start users and thus in improv￾ing the quality of item recommendations. With respect to the active group, it can be observed that CF ap￾proaches (i.e., UTU and ITI) provide better performance than our method. The result indicates that collaborative filtering relatively works well when users have abundant rating information. Compar￾ing results for the cold start group and the active group obtained by the CF approaches, it is apparent that the two groups have the number of recommended items (N) precision Fig. 4. Comparisons of precision as the number of recommended items N increases. ranking accuracy the number of recommended items (N) Fig. 5. Comparisons of ranking accuracy as the number of recommended items N increases. Precision at Top 10 Ranking at Top 10 Fig. 6. Comparisons of precision and ranking accuracy for cold start users and active users. H.-N. Kim et al. / Expert Systems with Applications 38 (2011) 8488–8496 8495