Scholar - SciOpen

Objective

Spatio-temporal correlation mining is a key technology in intelligent transportation systems and is usually applied to spatio-temporal data prediction problems such as traffic flow prediction. Accurately predicting traffic flows in urban management is extremely important for alleviating urban traffic congestion, improving traffic efficiency, and reducing traffic accident occurrences. However, it is extremely challenging to accurately predict traffic flows in large-scale traffic networks due to the high nonlinearity and complexity of the massive traffic flow data. Most existing methods usually conduct two separate components to capture the spatial and temporal correlations. A static spatial graph is constructed for each time step in the spatial dimension; furthermore, the same nodes on different time steps are connected to build a spatio-temporal graph in the temporal dimension. However, the potential correlations between the traffic flow data of different nodes at different time steps are ignored and the complex spatio-temporal correlations in the traffic flow data cannot be effectively modeled.

Methods

In this paper, we proposed a spatio-temporal combinational graph convolutional network (STCGCN) to address the issue of traffic flow prediction. STCGCN consisted of three modules: the spatio-temporal combinational graphs (STCG) construction module, the spatio-temporal combinational graph convolution (STCGC) module, and the prediction module. The STCG construction module constructed an adaptive STCG adjacency matrix across temporal slices based on spatio-temporal embedding vectors, which could automatically learn parameters during training, accommodate complex spatio-temporal correlations between nodes, and solve the problem that existing prediction methods hardly captured the potential spatio-temporal correlation between nodes. The STCGC module designed adaptive STCGC operators and adaptive STCGC layers to extract spatio-temporal features from historical traffic data of nodes and the constructed adaptive STCG. Finally, the prediction module aggregated the hidden layer representation of all historical time steps obtained using the STCGC module and outputed the prediction result via fully connected layer mapping. We evaluated STCGCN on PeMSD4 and PeMSD8, two public datasets from Caltrans performance measurement system (PeMS), by comparing it with 11 baseline methods: vector autoregressive (VAR), support vector regression (SVR), fully connected long-short term memory (FC-LSTM) neural network, diffusion convolutional recurrent neural network (DCRNN), spatio-temporal graph convolutional networks (STGCN), attention based spatial-temporal graph convolutional networks (ASTGCN), Graph WaveNet, spatial-temporal synchronous graph convolutional networks (STSGCN), adaptive graph convolutional recurrent network (AGCRN), graph multi-attention network (GMAN), and time zigzags at graph convolutional networks (Z-GCNETs). We adopted two widely used metrics for evaluation: mean absolute error and root mean squared error.

Results

The experimental results revealed that using a unified component, the proposed STCGCN model effectively modeled the dynamic temporal correlation, spatial correlation, and cross-spatio-temporal correlation in the traffic flow data. Furthermore, the model achieved the best prediction results at each moment, and its error growth was slower than other baseline methods as the prediction time increased. We also explored the effect of three hyperparameter settings in STCGCN on model performance, and the experiments demonstrated differential model performance under different hyperparameter settings. The number of parameters and training times of all models, including STCGCN and 11 baseline methods, were compared at the end of the experiment. The results showed that the STCGCN achieved the best model performance with the least number of model parameters and training time, and the algorithm efficiency was close to the best.

Conclusions

Experiments on the public datasets show that the STCGCN model outperforms 11 baseline methods in prediction accuracy.

The key challenge in processing point clouds lies in the inherent lack of ordering and irregularity of the 3D points. By relying on per-point multi-layer perceptions (MLPs), most existing point-based approaches only address the first issue yet ignore the second one. Directly convolving kernels with irregular points will result in loss of shape information. This paper introduces a novel point-based bidirectional learning network (BLNet) to analyze irregular 3D points. BLNet optimizes the learning of 3D points through two iterative operations: feature-guided point shifting and feature learning from shifted points, so as to minimise intra-class variances, leading to a more regular distribution. On the other hand, explicitly modeling point positions leads to a new feature encoding with increased structure-awareness. Then, an attention pooling unit selectively combines important features. This bidirectional learning alternately regularizes the point cloud and learns its geometric features, with these two procedures iteratively promoting each other for more effective feature learning. Experiments show that BLNet is able to learn deep point features robustly and efficiently, and outperforms the prior state-of-the-art on multiple challenging tasks.