Abstract:

Many e-commerce sites use a recommendation system to filter the specific information that a user wants out of an overload of information. Currently, the usefulness of the recommendation is defined by its accuracy. However, findings that users are not satisfied only with accuracy have been reported. We consider that a recommendation having only accuracy is unsatisfactory. For this reason, we define the usefulness of a recommendation as its ability to recommend an item that the user does not know, but may like. To improve user satisfaction levels with recommendation lists, we propose an alternative recom-mendation algorithm that increases the diversity of the recommended items. We examined items that appeal to several different taste tendencies to create a list and achieved diversity in that list. First, we created a similarity network of items by using item rating data. Second, we clustered the items in the network and identified the topics that appealed to the same preference tendency. Our proposed algorithm was able to include items covering several topics in the recommendation list. To evaluate the effect on user satisfaction levels, we used our algorithm to make a recommendation list for DVD items carried by Amazon.co.jp and conducted a questionnaire survey. The results showed higher levels of user satisfaction with our list than a list created using Collaborative Filtering (CF).

Kyewords: recommender systems, diversity, collaborative filtering, topic, network, cluster-ing.

---

proceedings