uTransfer: Unified Transferability Metric Incorporating Heterogeneous User Data in Social Network

Nuo Li, Bin Guo, Yao Jing, Zhiwen Yu

Abstract: Numerous users in social networks exhibit few-shot behaviors, and identifying appropriate neighbors has emerged as a promising solution. However, traditional similarity metrics often yield redundant neighbor information and fail to adequately consider the scarcity of user behaviors, thereby diminishing their effectiveness. This study endeavors to delve into the transferability between users by analyzing their heterogeneous data, to identify the most suitable users for knowledge exchange and reduce the impact of negative transfers. Existing transferability metrics mainly target homogeneous data, without considering the inherent characteristics and complementarity of heterogeneous data. To solve this, this paper proposes a novel metric, uTransfer, measuring the transferability between users with heterogeneous data in a more fine-grained and accurate way. Specifically, uTransfer first unifies user heterogeneous data into the behavior space to facilitate the fusion of heterogeneous knowledge. Then, uTransfer innovatively considers the specificity of heterogeneous data, proposes static and dynamic transfer modes, and models them separately to obtain finer-grained transferability results. Moreover, uTransfer uniquely models the complementarity between heterogeneous data to obtain more accurate transferability results. Finally, we integrate the complementarity and transferability results to measure the transferability between users. Extensive experiments demonstrate that uTransfer can effectively measure user transferability.