Sequential POI Recommend Based on Personalized Federated Learning

Qian Dong, Baisong Liu, Xueyuan Zhang, Jiangcheng Qin, Bingyuan Wang

Abstract: Point-of-Interest recommendation system (POI-RS) aims at mining users’ potential preferred venues. Many works introduce Federated Learning (FL) into POI-RS for privacy-protecting. However, the severe data sparsity in POI-RS and data Non-IID in FL make it difficult for them to guarantee recommendation performance. And geographic factors in check-ins easily make model training ineffective in FL. For example, geographical cultural differences will aggravate the Non-IID nature of data. To cope with the problems, we propose a new framework FedSR for POI-RS based on sequential information. In the FedSR, we construct a multi-task framework through Contrastive Learning (CL). In this multi-task, Bayesian Personalized Ranking (BPR) optimization is used for the recommendation task, and a data augmentation method is applied to CL based on geographical correlations between POIs. In addition, to effectively train the multi-task model, we adopt a personalized federation method, which includes similar user grouping based on Locality Sensitive Hashing. The way consumes less computational resources as well as does not expose any private information. We validate the effectiveness of FedSR on two real datasets. Further analysis reveals the performance improvement by FedSR.

External IDs:dblp:journals/npl/DongLZQW23