The point-of-interest (POI) recommendation is a key function of location-based social networks that can help users exploit unfamiliar areas. Due to the massive check-in records accumulated in these location-based applications, the sequential POI recommendation has evolved quickly in the research community. Although the existing sequential POI recommendation models have reached an encouraging performance in predicting the next-N POIs to users, the data sparsity problem is still severe in the sequential POI recommendation task. It is challenging to learn the users’ preferences of POIs under the highly sparse dataset. Meanwhile, many sequential models focus on exploiting users’ interests from the entire dataset, ignoring the effect of collaborative information from similar users. To this end, we propose a user-based clustering deep model (UCDM) for the sequential POI recommendation to deal with these issues. UCDM extracts collaborative information via a user-based intent clustering module and uses a binary self-attention layer to both learn the general preference from the entire dataset and the local preference from the collaborative information. In addition, our proposed model uses a POI candidate filter to control the size of the POI candidate set to reduce the sparsity of the dataset. In the model learning phase, we adopt the Bayesian personalized ranking to train our model. The experiment verifies that our proposed UCDM outperforms the selected baseline models for the sequential POI recommendation on two real-world check-in datasets.