ExplainedKG-Next: an explainable human mobility pattern learning framework for next location prediction

The sparsity of check-in data significantly impedes the accuracy of location prediction. To address this, various location prediction methods with knowledge graphs have been developed. Nevertheless, current approaches frequently lead to the coupling embeddings of multiple features, complicating the process of pinpointing the distinct contributions of individual elements within the knowledge graph. Inspired by explainable artificial intelligence, in this study, we proposed an explainable human mobility pattern learning framework for next location prediction. To this end, we first designed a spatio-temporal knowledge graph (ST-KG) with quadruplets to formulate the complex interactions between users and locations and employed multichannel graph attention network (GAT) to derive decoupled embeddings of point-of-interest (POI) spatio-temporal information. Next, we use the multi-head self-attention mechanism to extract multiple spatio-temporal features from check-in sequences for location prediction. Finally, we utilize the Shapley value to elucidate the significance of the elements within ST-KG. Case studies were conducted in New York City and Singapore to demonstrate the reliability and effectiveness of the proposed approach. Our model outperformed several representative baselines on all evaluation metrics. In the experiment of attribution analysis, spatial transfer and spatial proximity were critical factors influencing user mobility; however, the frequently visited time reduces the predicted value.