Abstract: With the development of mobile location technology, the point-of-Interest (POI) recommendation technology has become one of the research hotspots in the field of recommendation system. Limited by the check-in ability of users, there is a serious data sparsity problem in POI recommendation, what’s more, POI recommendations with combining multiple sources of data is faced with the challenges of multiple privacy. Because data from multiple sources have privacy characteristics of diversity and pluralism, the mechanism of privacy leakage is more complex and diverse, which makes the privacy problem more challenging. To solve this problem, a multi-source POI recommendation with attention mechanism and privacy protection, MultiAM&PP_POI, is proposed in this paper. In order to achieve effective data fusion, LDA topic model to extract users’ potential features in different domains is used in this paper, and an attention mechanism is used to adaptively train the influence of potential features in different domains on POI recommendation results. Meanwhile, multi-layer perceptron method is used to realize the transfer of potential features in different domains. As for the privacy problem in the recommendation of multi-source POI, the federated learning framework is used to save the original data locally without uploading, so each participant only needs to interact with the potential features after encryption. The attention mechanism is improved so multi-layer perceptron can finish training in ciphertext state to protect the security of user privacy. Finally, through experimental verification, the proposed model can effectively improve the recommendation accuracy by 3.05% and 4.42% compared with the single-source federated model and other cross-domain models on the premise of protecting users’ privacy.

Key words: point-of-interest recommendation, multi-source fusion, attention mechanism, privacy protection, federated learning