Eigentaste 5.0: Constant-Time Adaptability in a Recommender System Using Item Clustering
Tavi Nathanson, Ephrat Bitton and Ken Goldberg
Our goal is to create a highly adaptive collaborative filtering system that runs in constant online time.
Recommender systems strive to recommend items that users will appreciate and rate highly, often presenting items in order of highest predicted ratings first. Our current research consists of the development of Eigentaste 5.0, a constant-time recommender system that dynamically adapts the order that items are recommended by integrating user clustering with item clustering. This extends our Eigentaste 2.0 algorithm, which uses principal component analysis to cluster users offline. In preliminary experiments we backtested Eigentaste 5.0 on data collected from Jester (http://eigentaste.berkeley.edu), our online joke recommender system. Results suggest that it will perform better than Eigentaste 2.0. The new algorithm also uses item clusters to address the cold-start problem for introducing new items.
We are seeking a more robust method for determining similarity between items when data is very sparse. We will also be experimenting with generalizing Eigentaste 5.0 in order to introduce more diversity among recommendations and to give users more control over how much item similarity they desire.
Figure 1: Jester: The Online Joke Recommender
Figure 2: Items are recommended to a user from his top-rated item cluster in the order determined by the aggregate item rankings of users from the same user cluster.
Figure 3: Average difference (across 7,000 users) between actual ratings for Eigentaste 5.0 and 2.0 for the i-th recommended item
- K. Goldberg, T. Roeder, D. Gupta, and C. Perkins, "Eigentaste: A Constant Time Collaborative Filtering Algorithm," Information Retrieval, Vol. 4, No. 2, pp. 133-151, 2001.