People face a bewildering number of choices when looking for items that they are interested in. A seemingly never-ending food of content is available, but certainly not enough time exists to evaluate all possible choices. This, in a nutshell, is the problem of information overload. In recent years, recommender systems have emerged as one tool that can help people overcome this problem and quickly locate items to consume. These systems use opinions about items in some information domain in order to make recommendations to a user regarding which items she may nd interesting.
One instance of a recommender system is MovieLens (http://www.movielens.org). GroupLens, our research group, operates this recommender system, which makes personalized recommendations suggesting movies that a user might like based on movies that she has seen and has expressed an opinion about. While recommender systems are clearly benecial to users, they can also be a valuable asset to retail companies in helping their customers nd things that they might want to buy and, in effect, increasing not only sales, but perhaps also cross-sales and customer retention. This is particularly true in the realm of e-commerce. For example, Amazon.com has made many recommender systems available to their customers. These range from manually operated recommenders where users can recommend items to other users by writing reviews or creating lists, to automated systems where the site generates a list of recommended items based on what the user has looked at recently or has purchased in the past.
Producers of items (manufacturers, authors, etc.) would like their products to sell well in the marketplace. With recommender systems, there is a natural motivation to want one′s own products to be recommended more often than those of a competitor. Of course, one way to accomplish this is to produce quality goods that people like and regard highly. However, unscrupulous producers may opt to take a more deceitful route; they may try to inuence recommender systems in such a way that their items are recommended to users more often, whether or not they are of high quality. An instance of a company generating false ?recommendations? to consumers arose in June 2001 when Sony Pictures admitted that it had used fake quotes from non-existent movie critics to promote a number of newly released lms.1 The online retailer Amazon.com has found that their recommenders are prone to some level of abuse on at least two different occasions.2;3 Also, eBay, which uses a recommender system as a reputation mechanism, has found itself continually dealing with users who subvert the system in various ways, including purchasing good ratings (feedback) from other members in order to bolster their own reputations.4 One way to inuence a recommender system is to arrange to have a group of users (human or agent) enter the system and vouch for the items in question. These users become shills, whose false opinions are intended to mislead other users. Shills pose a serious threat to users and operators of recommender systems. They may cost users time and money by recommending bad items. They may cost operators by degrading the user′s level of trust in the recommender system and the retailer behind it. 数据挖掘实验室
This paper focuses on recommender systems that use automated collaborative ltering (ACF) to generate recommendations. ACF is a class of algorithms commonly used in the implementation of recommender systems. These algorithms operate on the basis that similar users have similar tastes; thus, if people similar to you can be located, then the items they enjoy are likely to be ones you will also enjoy. These algorithms normally have two modes of operation: prediction and recommendation. In the prediction mode, the algorithm simply predicts how much a user will like some item or set of items. The items may have been selected by the user through browsing or searching. In the recommendation mode, the algorithm produces an ordered list of items that it believes the user is most likely to enjoy. This distinction will become important as we explore the practical effects of attacks on recommender systems. The rest of the introduction comprises related work and the statement of hypotheses, setting the stage for the experimental work in the rest of the paper. 数据挖掘研究院
