4. Conclusions

Providing effective recommendations based on clickstream data can be important at early stages of a user's interaction (when more explicit user input such as ratings or personal perfiles are not available). Personalization at this level results in these user's engaging at a deeper level, and can help improve the conversion efficiency of a site. Standard collaborative filtering techniques such as kNN require real-time comparison of a current users record with the historical records of other users. This methodology becomes increasing unscalable as the number of users and items increase. The lack of scalability of kNN becomes amplified when dealing with the large volume of clickstream data.

In this paper we have presented a scalable framework for Web personalization based on association rule mining from clickstream data. Our framework includes an efficient data structure for storing frequent itemsets combined with a recommendation algorithm which allows for the generation of recommendations without first generating all association rules from itemsets. We have also studied the impact of using multiple support levels for different types of pageviews, as well as the use of varying-sized user histories on the precision and coverage of the generated recommendations. Our results show that the proposed framework can provide an effective alternative to standard collaborative filtering mechanism for personalization. In particular, we have shown that the association-based recommendation framework can, in fact, improve on the kNN-based collaborative filtering both in terms of the precision and coverage of recommendations, while at the same time maintaining the computational advantage over kNN attained due to the offline discovery of frequent patterns.