<p>In recent years, next POI (point of interest) recommendation has emerged as a broad research area. The task involves leveraging user check-in data to recommend POIs at specific times. Despite the remarkable progress in next POI recommendation, the complex temporal correlations between POIs as well as dynamics of user preferences make it challenging to leverage user check-in data effectively. Therefore, we propose a novel user preferences learning model (UPLM) for next POI recommendation. This model employs a temporal point process to finely model the time intervals between historical POI check-ins and the user’s current time, and effectively integrate the impact of POI visit frequency on recommendations. Additionally, the model incorporates an attention mechanism to better capture both long-term and short-term user preferences information. Considering the impacts of POI categories, we utilize LSTM (long short-term memory) networks to learn users’ preferences for different categories. By comprehensively considering both POI and category preferences of users, we can obtain the probabilities of all candidate POIs and recommend the POI with the highest probability. Finally, our model is evaluated on two real-world datasets, and the results demonstrate that the proposed model is better than the state-of-the-art next POI recommendation methods. We achieved a recall@10 of 45.76% with the NYC dataset and 41.88% with the Weeplace dataset. We also achieved an MRR of 22.67% with the NYC dataset and 21.89% with the Weeplace dataset.</p>