Spatial–temporal restricted supervised learning for collaboration recommendation

Collaboration recommendation from scholarly big data is an important but challenging problem as it might suffer the difficulty of accurate recommendation from three aspects: how to efficiently integrate the available author-related information, how to precisely describe the characteristics of the scholarly data samples, and how to extract the intrinsic features that are more suitable for collaboration recommendation. Facing these challenges, we incorporate the temporal and academic-influence information of the publications with the spatial information of the researchers to present a spatial–temporal restricted supervised learning (STSL) model for collaboration recommendation. We first present a topic clustering model to determine the topic distribution vector of each researcher, where a temporal parameter is introduced to exponentially weight each topic distribution vector and an academic-influence parameter is further introduced to linearly combine all the topic distribution vectors of the publications. Then, inspired by the geographical-advantage phenomena in collaboration, spatial labels are generated by using the personal information of the researchers. Furthermore, considering that the publication data enhanced by spatial–temporal and academic-influence descriptions usually exhibit multimodal or mixmodal properties, we propose a data-driven supervised learning model to extract the intrinsic features inhered in data, which determines a low-dimensional recommendation subspace. A number of experiments are conducted to test the impact of the topic-clustering number, the temporal parameter, the academic-influence parameter, and the number of extracted features. Besides, several widely-used models are adopted to compare with the proposed STSL model for collaboration recommendation, with results verifying its feasibility and effectiveness.