| Popis: |
As the extensive use of e-commerce through web-site increases, the need for other marketing approach is also increasing more than ever before. Increased concern by on-line company and academia has led to the development of numerous method and techniques that improve the performance of recommender system and promote customers’ interests. In this work we presented our research results in the area of collaborative filtering algorithm and other techniques to improve the performance of recommender systems which are one of the most important tools for the on-line marketing. From our experimental results, it can be summarized as two main parts. One is algorithmic improvements for prediction accuracy and the other is possibilities of pre-evaluation methods using each user’s rating pattern which is already collected in the system. In the view point of algorithmic improvements, the followings are the results of this study. First, the prediction performance of CMA on the view of accuracy is superior to that of NBCFA compared to all the results of user-based and item-based approaches. Second, the significance weight which makes up for overestimated preference relationships between target user and his or her neighbours, where the number of co-ratings is so small, contributes greatly to the accuracy of prediction. Also it is necessary to set the extended weighted range rather than existing N/50 ratings. Third, under the extending scale of recommender system, it is more efficient to run the recommender system controlling the increasing numbers of items than to control the increasing numbers of customers. Itembased approach which controls the numbers of items has the more accurate prediction results than those of user-based approach, but our another research which isn’t presented on this work shows that the rank correlations between predicted values and real values of userbased approach are more accurate than those of item-based approach. This means that it would be needed to decide one of the two approaches between accurate prediction for rating and customer’s preference rank for trade-off. It will be needed that the further research on this topic follows. In the view point of pre-evaluation, the followings are the results of this study. This work presents experimental results about setting the error bound for classifying the users who have lower prediction performance before prediction process using collaborative filtering algorithm in the recommender system. Through the statistical analysis, we have |
