Paper address: https://www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/download/16069/15978
Thesis title: Deep Multi-View Spatial-Temporal Network for Taxi Demand Prediction
Summary
1. The model integrates spatiotemporal data and semantic relationships
2. The local CNN model captures spatial proximity relationships
3. POIs that are far away but have similar semantics are vectorized through graph embedding.
4. LSTM models time series, including CNN features and other features of the current time (such as weather, etc.)
Model
a: spatial view. First, the entire map is rasterized and divided into 7*7 pixels. After k layers of convolution, it is flattened and then connected to a fully connected layer to obtain the spatial characteristics of each moment.
b: Temporal view, put the spatial features and other features of the spatial view into the lstm model to obtain the spatio-temporal features
c: semantic view. Semantic features are obtained through graph embedding. The similarity definition uses Dynamic Time Warping. The embedding method uses Tang Jian’s 15-year paper: Line: Large-scale information network embedding.
Paper introduction: https://blog.csdn.net/guoyuhaoaaa/article/details/78743023
Paper code: https://github.com/snowkylin/line or https://github.com/VahidooX/LINE
d: The fusion of spatiotemporal features and semantic features is followed by a fully connected layer to obtain the prediction result. The loss function includes mean square error and mean square percentage error.
Model effect
