Sie sind hier: Startseite Daten Lehrstühle Datenbanken und Informationssysteme Next Generation Recommender Systems

Next Generation Recommender Systems

Recommender systems are personalization tools that intend to provide people with lists of suggestions that best reflect their individual taste. In order to create profiles of the users' behavioral patterns, explicit ratings (e.g., user a rates book b with 4 of 5 stars) and purchase data are collected.

However, in the past, recommender systems research has primarily concentrated on accuracy metrics only, i.e., increasing the performance along precision/recall and mean absolute error (MAE) benchmarks.
However, some issues have remained unnoticed, for instance the lack of diversity in recommendation lists. Overfitting of personal user profiles may lead to situations where most recommendations come from one single field of interest only, as can be seen in the image below, which shows C.-N. Ziegler's top-4 recommendations on Amazon.com.
All 4 recommendations are from the Tolkien universe, though the user's interest also span other fields of interest, e.g., travel, medieval romance, and science. Unfortunately, these are not contained in the recommended products list. Particularly commercial recommender systems suffer from lack of diversity and overfitting.

Overfitting effect in recommendation lists

In order to overcome the diversity issue and have recommendation lists reflect all major fields of interest in a proportial weighting, we have developed the novel Topic Diversification method, in cooperation with GroupLens Research labs in Minneapolis, MN, USA. In order to demonstrate that our approach bears some actual benefit for the user, increasing his level of satisfaction with recommendations made, we collected extensive user data from BookCrossing (see below) and computed personalized surveys that were completed by more than 2,100 BookCrossing members. The results confirmed our hypothesis. Users like to have diversified recommendation lists spanning their entire spectrum of interests rather than merely one or two topics.
Hereby, tests have indicated that diversification levels around 30-40% are best for the item-based recommender algorithm that commercial systems (e.g., Amazon.com, TiVo) are typically using (see below).

Effect of increasing diversification on user satisfaction


Benutzerspezifische Werkzeuge