Cross-Domain Age Inference Framework Based on Common-Topics

Privacy leakage is one of the raising serious concerns in public. Some online service systems cope with this problem bypassing gathering user privacy information, age, for instance. Without collecting and studying these information, is it possible to infer user age or not in such systems? In this paper, we proposed a cross-domain age inference framework based on common-topics to show the possibility of age disclosure with aid of transferred knowledge from an auxiliary domain. Specifically, we take the online book and movie systems, i.e., BookCrossing and Movielens systems, as target and auxiliary domains respectively. First, we extract common-topics from item titles and descriptions to aggregate user behavior. Such aggregation behaviors of Movielens system are regarded as transferred knowledge, and bridge BookCrossing system to build age inference module. Moreover, various classical algorithms are compared. And targeting age imbalance distribution, balance evaluation metrics are proposed to evaluate our algorithm. Experiment results show that for BookCrossing system, age inference performance of our proposed cross-domain age inference framework based on common-topics is even better than that of system with available age information, which means that the privacy risk even exist in a supposedly safe environment.