Abstract: To solve the problem that CF(Collaborative Filtering) recommendation highly depends on the accurate similarity measurement, a novel recommendation algorithm based on item hierarchy similarity was proposed, which was named REHIS(Recommendation Hierarchical Similarity). The framework of REHIS was described as follows. First, the mining association rules and KNN (K Nearest Neighbor) algorithm were used to complement the hierarchy structure. Afterwards, the TopK method was employed to compute the similarity between items. Finally, scores were predicted by using the framework of itembased CF algorithm. On the other hand, to solve the CF poor scalability problem, the TopK algorithm were further extended to the cosine distance and Pearson correlation coefficient, both of which were commonly used similarity measurement methods. Experimental results showed that, compared with existing algorithms, REHIS could achieve a better recommendation in term of root mean square error, and TopK could reduce the time cost for searching the most similar items, too.

Key words: recommendation system, tag, TopK, inverted index, collaborative filtering, item hierarchy