Mining the Typical Preference of Collaborative User Group Su-Jeong Ko and Jiawei Han   View Paper (PDF)   Return to Data Mining Abstract Collaborative filtering systems have the problems of sparsity– providing a recommendation by correlation between only two customers preferences, and being unable to recommend a unique item owing to the recommendation based on preference rather than on the content of the item. The native feature space consists of unique words with single dimension when it occurs in documents as items, which can be tens or hundreds of thousands of words for even a moderate-sized text collection. This is prohibitively high for many learning algorithms. Since the feature extraction method using association word mining does not use the profile, it needs not update the profile, and it automatically generates noun phrases by using confidence and support of the Apriori algorithm without calculating the probability for index. However, in case that the feature extraction method is based on a set of association words, it makes an error of judging different documents identically. This paper proposes an association word mining method with weighted word, which reflects not only the preference rating of items but also information on the items. The proposed method is capable of creating the profile of the collaborative users, in which users are grouped according to the vector space model and Kmeans algorithm. Thus, the new method eliminates the existing collaborative filtering systems problems of sparsity and of recommendations based on the degree of correlation of user preferences. Entropy is used in order to address the said systems shortcoming whereby items are recommended according to the degree of correlation of the two most similar users within a group. Thus, the typical preference of the group is extracted. Since user preferences cannot be automatically regarded as accurate data, users within the group who have entropies beyond the threshold are selected as typical users. After this selection, the typical preference can be extracted by assigning typical user preferences in the form of weights. By using the typical preference of the group, the method also reduces the time required for retrieving the most similar users within the group. ©2005 Association for Computing Machinery