Sampling-based epoch differentiation calibrated graph convolution network for point-of-interest recommendation

Abstract

In location-based social networks, calibrating a point-of-interest (POI) recommendation system is as important as its accuracy for improving user satisfaction. POI recommendation calibration is primarily classified as categorical or geographical calibration. Categorical calibration ensures that the recommended items are distributed proportionally among the past interest categories of the target user. When a target user checks 80 Chinese, 10 Japanese, and 10 French restaurants, a recommendation list with a ratio of 8:1:1 for Chinese, Japanese, and French restaurants can be reasonably expected. In addition to categorical calibration, geographical calibration has been proposed to increase user interest in the recommended results. Users have a high probability of revisiting locations in their subareas. Therefore, the POIs recommended in multiple subareas of interest are more suitable than those from one small and frequently visited subarea. However, improving the calibration and accuracy are conflicting tasks. To achieve high calibration while maintaining accuracy, previous studies proposed reranking-based techniques to rerank the candidate list and return POIs with high calibration. However, optimizing the calibration by reranking is independent of the basic-candidate-item generation model, resulting in a suboptimal system. To tackle the problem, we propose a novel sampling-based differentiation technique to merge the task of improving calibration into the GCN model training process and directly generate the final recommendation list. The model is flexible and can be applied to different domains, where a domain can be a subarea or category. In a three-layer GCN, the layer one represents the historical check-ins of the user, whereas layer three includes the candidate POIs from which the target user aggregates information. We trained the model to make the distribution of the POI domains at layer three approximated the distribution at layer one. Experimental results on Philadelphia and Tucson datasets confirmed that the proposed method outperforms all state-of-the-art GCN+ geo-reranking and GCN+ MCF baselines, improving Recall@ 5 from 0.0394 to 0.0412 (4.57%) and Jensen–Shannon measure (JS)@ 5 from 0.5931 to 0.6734 (13.54%) on the Philadelphia dataset and improving Recall@ 5 from 0.0495 to 0.0517 (4.40%) and JS@ 5 from 0.5869 to 0.6598 (12.42%) on the Tucson dataset for categorical calibration. The model was also tested in the geographical domain and a similar trend was observed.