This paper proposes a Deep Multi-View Spatial-Temporal Network (DMVST-Net) for taxi demand prediction. The model integrates three views: spatial, temporal, and semantic. The spatial view uses a local CNN to capture spatial dependencies among nearby regions. The temporal view employs LSTM to model sequential dependencies in demand time series. The semantic view constructs a graph of regions based on their similarity in demand patterns, using graph embedding to capture latent semantic information. The model combines these views to predict taxi demand accurately. The DMVST-Net is evaluated on a large-scale real-world taxi demand dataset from Didi Chuxing, containing taxi requests in Guangzhou over two months. The model outperforms state-of-the-art methods in terms of Mean Average Percentage Error (MAPE) and Rooted Mean Square Error (RMSE). The results show that the model achieves the lowest MAPE (0.1616) and RMSE (9.642) among all methods, demonstrating its effectiveness in predicting taxi demand. The model's performance is robust across different days of the week, with consistent results on all seven days. It also performs well on weekends, despite the less regular demand patterns. The model's performance is further validated by analyzing the influence of sequence length for LSTM and input size for local CNN. The results show that the model achieves the best performance when the sequence length is 4 hours and the input size for local CNN is 9x9. The paper also compares the proposed method with variants of the model, showing that combining all three views leads to the best performance. The model's effectiveness is attributed to its ability to capture complex spatial and temporal dependencies, as well as semantic relationships between regions. The results demonstrate that the DMVST-Net is a promising approach for taxi demand prediction in smart cities.This paper proposes a Deep Multi-View Spatial-Temporal Network (DMVST-Net) for taxi demand prediction. The model integrates three views: spatial, temporal, and semantic. The spatial view uses a local CNN to capture spatial dependencies among nearby regions. The temporal view employs LSTM to model sequential dependencies in demand time series. The semantic view constructs a graph of regions based on their similarity in demand patterns, using graph embedding to capture latent semantic information. The model combines these views to predict taxi demand accurately. The DMVST-Net is evaluated on a large-scale real-world taxi demand dataset from Didi Chuxing, containing taxi requests in Guangzhou over two months. The model outperforms state-of-the-art methods in terms of Mean Average Percentage Error (MAPE) and Rooted Mean Square Error (RMSE). The results show that the model achieves the lowest MAPE (0.1616) and RMSE (9.642) among all methods, demonstrating its effectiveness in predicting taxi demand. The model's performance is robust across different days of the week, with consistent results on all seven days. It also performs well on weekends, despite the less regular demand patterns. The model's performance is further validated by analyzing the influence of sequence length for LSTM and input size for local CNN. The results show that the model achieves the best performance when the sequence length is 4 hours and the input size for local CNN is 9x9. The paper also compares the proposed method with variants of the model, showing that combining all three views leads to the best performance. The model's effectiveness is attributed to its ability to capture complex spatial and temporal dependencies, as well as semantic relationships between regions. The results demonstrate that the DMVST-Net is a promising approach for taxi demand prediction in smart cities.