Sunday: UCERSTI2, Monday: Poster session, Tuesday: Human Factors session, Thursday: Decisions@RecSys workshop
Hi there! Chances are you were attending the RecSys 2011 conference and I gave you this link, or you saw it on a slide. Below are all the things I am presenting at the conference. If you have more questions, feel free to shoot me an email or send me a tweet. Enjoy the conference!
Joe Konstan and I are hosting a panel discussion in the UCERSTI::2 workshop on Sunday. We want to engage the audience in a discussion on how to turn the evaluation of recommender systems into a unified, cumulative science. More information about this presentation, and about the UCERSTI workshop in general can be found on the UCERSTI website.
A Pragmatic Procedure to Support the User-Centric Evaluation of Recommender Systems
I believe that you cannot test a recommender system offline and conclude that it is "better" for the user; In order to make that claim, you will have to do online user testing. However, user testing is a difficult endeavor! Is there a way to make it easy?
Yes, there is! Our poster / short paper presents the 4 essential steps to user testing for recommender systems. Our pragmatic procedure can help you find out how different aspects of your system influence the user's experience. This involves:
Our poster will give you tips on how to do this efficiently. The explained procedure is based on our upcoming UMUAI paper. A short introduction to this work can be found here.
Each to His Own: How Different Users Call for Different Interaction Methods in Recommender Systems
Recommender systems help you to make decisions by taking your personal preferences into account. But people also differ in how they make decisions. We explore user characteristics that lead to different decision strategies, and tailor the interface of a recommender system to these user characteristics.
Our results show that in general it is best to employ a hybrid recommender system which implicitly measures user preferences, but also allows the user to make adjustments to these preferences. However, domain novices lack the knowledge to succesfully use a personalized recommender system. They are better off with a non-personalized list of most popular items. The same seems to be true for maximizers, who experience post-decision regret when they are allowed to actively control the recommendation process.
In the paper, we discuss how you can combine these two interfaces to provide a good user experience for all users.
Using latent features diversification to reduce choice difficulty in recommendation lists
Choice overload describes the situation where you cannot decide between too many options. This problem seems to be even more prominent in recommender systems, where all recommended options fit your personal preferences. In a previous study we have shown that choice overload causes a lower choice satisfaction, especially when there are many recommendations. Is there a way to overcome this problem?
In the current study, we use a new technique to diversify recommendations based on the latent features of a Matrix Factorization algorithm. We subsequently show that more diversifying the recommendations can reduce tradeoff difficulty and choice difficulty, thereby effectively eliminating choice overload.