University of Minnesota researchers Joseph A. Konstan and John Riedl report that Web sites that accept user ratings but which show users the existing ratings before the users enter their own ratings influence the ratings that later reviewers enter.

ARLINGTON, Va. – Online "recommender systems" are used to suggest highly rated selections for book buyers, movie renters or other consumers, but a new study by University of Minnesota computer science researchers shows for the first time that a system that lies about ratings can manipulate users' opinions. Over time, however, users lose trust in unscrupulous systems.

The Minnesota research group, led by professors Joseph Konstan and John Riedl, conducted experiments with MovieLens, a member-driven movie recommendation Web site, which the Minnesota team created to put their research on recommender systems into practice. The results of their latest studies will be presented at the CHI 2003 Conference on Human Factors in Computing Systems, April 5-10, 2003, in Fort Lauderdale, Fla. The work is supported by the National Science Foundation, the independent federal agency that supports basic research in all fields of science and engineering.

The results suggest that "shills," who flood recommender systems with artificially high (or low) ratings, can have an impact and that recommender systems will take longer than previously expected to self-correct. The idea behind self-correction is that misled users will come back and give the item in question a more accurate rating.

"This study isn't about defending against shill attacks, but about trust in the system," Konstan said. "Our results show that individual shilling will take longer to correct, but the larger impact is the loss of trust in the system." In the Minnesota study, users were sensitive to the quality of the predictions and expressed dissatisfaction when the system gave inaccurate ratings.

Taken together, the experiments provide several tips for the designers of interfaces to recommender systems. Interfaces should allow users to concentrate on rating while ignoring predictions. Finer-grained rating scales are preferred over simple "thumbs up, thumbs down" scales, but are not essential to good predictions. Finally, because users are sensitive to manipulation and inaccurate predictions, to keep customers happy, no recommender system at all is better than a bad recommender system.

Many online shopping sites use recommender systems to help consumers cope with information overload and steer consumers toward products that might interest them. Recommender systems based on "collaborative filtering" generate recommendations based on a user's selections or ratings and the ratings of other users with similar tastes.

The CHI 2003 paper reports the results of three experiments that attempted to shed light on questions such as "Can the system make a user rate a ‘bad’ movie 'good'?" "Do users notice when predictions are manipulated?" "Are users consistent in their ratings?" and "How do different rating scales affect users' ratings?"

"To our knowledge, no one has studied how recommendations affect users' opinions," Konstan said. "We thought it was a socially important question because many of these systems are used in commercial applications. We found that if the system lies you can gain some influence in the short term, but you lose trust in the long term. This is a lot like trust between humans."

In the first experiment, users re-rated movies while presented with some "predictions" that were different from the user's prior rating of the movie. In the second experiment, users rated movies they hadn't rated before. Some users saw "predictions" that were skewed higher or lower, while a control group saw only the system's accurate predictions. The final experiment had users re-rate movies using three different rating scales and asked users their opinions of the different scales.

In the re-rating experiment, users were generally consistent, but they did tend to change their rating in the direction of the system's skewed "prediction." In fact, users could even be influenced to change a rating from negative to positive or vice versa. When users rated previously unrated movies, they tended to rate toward the system's prediction, whether it was skewed or not.

Konstan also noted that, while this study used movie ratings, recommender systems have much wider application. Such systems can also be used for identifying highly rated research papers from a given field or for knowledge management in large enterprises. New employees, for example, can bring themselves up to speed by focusing on highly rated news items or background materials.

Some quantitative results from the study:

The results revealed that people were fairly consistent in re-rating movies when there were no other ratings on-screen. Participants gave the movies the same ratings 60 percent of the time, one star below the original rating 20 percent of the time and one star above the original rating 20 percent of the time.

The results also showed that having other ratings on screen, whether they matched the user's original rating or were one star up or down, influenced the second rating the user gave. When ratings were bumped up or down one star, participants rated nearly 30 percent of movies one star above or below the original rating, respectively.

If a few people go over to Amazon.com and put in favorable reports for a new book those initial favorable reports will then tend to skew later ratings that are entered. I'm guessing that if all the ratings for a book are skewed in one direction at least some would-be reviewers who would otherwise have entered in a rating in opposite driection will decide not to enter a review at all since their review would be so far out of step with the other ratings.

The other problem with rating systems is that one doesn't really know much about the kinds of people doing the ratings. It would be far more helpful when reviewing, say, a technical book on a programming language to know whether the reviewer has been programming for very long, knows other languages, or has other books on the same language that they like more or worse. Some reviewers include those kinds of facts but most do not. Also, most book reviewers do not explain enough to allow a reader to understand what they did or did not like about a book. There is something to be said for a rating form that would allow a person to fill in a survey about a book covering major categories such as accuracy, readability, suitability for beginners, intermediate, and advanced users, and so on.

Konstan and Reidl are also co-authors of a book entitled Word of Mouse: The Marketing Power of Collaborative Filtering.