Sociologists have long tried to characterize the in uence of a person in a social network of many people [1]. Identifying the in uential people can bring twin advantages to those who study group dynamics: (1) The in uential people can be directly studied, yielding insight since their choices may be predictive of group choices; or (2) The in uential people may be in uenced to change the behavior of the group. Many social networks are formed and maintained through informal, qualitative, and unobserved interactions. Capturing data about these interactions is difficult, and the act of capturing those data may change the social interactions themselves.
Collaborative Filtering (CF) recommender systems [2, 3, 4] base their decisions on the opinions of users. In contrast to other social networks, recommender systems capture interactions that are formal, quantitative, and observed. The social network can be analyzed directly through data already captured in the computer system. 数据挖掘研究院
Past research has demonstrated that analyzing the social network can provide leverage in in uencing the group [5]. The analysis performed in these studies is based on a deep investigation of the characteristics of one particular recommender algorithm, the wellknown user-user nearest neighbor algorithm [2]. Careful analysis of this type has many advantages, but one key disadvantage: it is tied closely to the details of the algorithm. In principle, similar techniques could be applied to other algorithms, but doing so would be laborious, and the resulting in uence measure only applies to algorithms that work precisely according to the details of the analysis. Since many commercial operators tweak the operation of the recommender in many ways to t the needs of their business, this analysis may not apply in practice. Further, the resulting measures of in uence would be unlikely to be comparable between di erent algorithms, since they have been produced through very di erent techniques. 数据挖掘研究院
A key goal of the present research is to identify a measure of in uence for recommender systems that is applicable to any ratings-based recommender system, independent of the particulars of the algorithm. Such a measure would allow for consistent, black-box analysis of in uence. 数据挖掘研究院
