Posted by Admin: System Admin
Accurate and timely traffic flow prediction is crucial for intelligent transportation systems (ITS). Recent advances in graph-based neural networks have achieved promising prediction results. However, some challenges remain, especially regarding graph construction and the time complexity of models. In this paper, we propose a multi-stream feature fusion approach to extract and integrate rich features from traffic data and leverage a data-driven adjacent matrix instead of the distance-based matrix to construct graphs. We calculate the Spearman rank correlation coefficient between monitor stations to obtain the initial adjacent matrix and fine-tune it while training. As to the model, we construct a multi-stream feature fusion block (MFFB) module, which includes a three-channel network and the soft attention mechanism. The three-channel networks are graph convolutional neural network (GCN), gated recurrent unit (GRU) and fully connected neural network (FNN), which are used to extract spatial, temporal and other features, respectively. The soft-attention mechanism is utilized to integrate the obtained features. The MFFB modules are stacked, and a fully connected layer and a convolutional layer are used to make predictions.We conduct experiments on two real-world traffic prediction tasks and verify that our proposed approach outperforms the state-of-the-art methods within an acceptable time complexity. Machine learning is an important component of the growing field of data science. Through the use of statistical methods, different type of algorithms is trained to make classifications or predictions, and to uncover key insights in this project. These insights subsequently drive decision making within applications and businesses, ideally impacting key growth metrics. Machine learning algorithms build a model based on this project data, known as training data, in order to make predictions or decisions without being explicitly programmed to do so. Machine learning algorithms are used in a wide variety of datasets, where it is difficult or unfeasible to develop conventional algorithms to perform the needed tasks.
Recently, several researchers apply the graph-based deep learning approaches for traffic prediction. Thanks to the powerful expression of graphs for non-Euclidian structures, learning from graphs based on road sensor networks has achieved more accurate results [26]–[28]. In this kind of method, the road sensor network is regarded as a graph, where nodes represent monitor stations and contain traffic information, and an adjacent matrix is used to describe the correlation between stations. The construction of an adjacent matrix affects the expressive power of the graph directly. The graphs can be divided into directed and undirected graphs. The adjacent matrix for undirected graphs is symmetric, such as the connection between social networks [29] and quantum chemistry [30]. It is not the same case in directed graphs, such as paper citation networks and road sensor networks [7]. As to the implementation of GCN, there are two alternative approaches including spectral methods and non-spectral methods. Based on spectral methods, the convolution operation is mapped to the frequency domain, so the convolution in the time domain is replaced by the product operation in the frequency domain. To reduce the computing complexity, localized spectral graph convolution [31] and polynomials approximate expansion [32] are proposed. Yu et al. constructed the ST-block which is composed of graph convolution layers and sequence convolution layers. It can capture spatiotemporal correlation by applying a convolution operation [26]. Based on non-spectral methods, the convolution operation of the adjacent matrix is carried out directly and the pooling operation is replaced by sparsing the adjacent matrix [33]. Later, the graph attention neural network (GAT) is proposed to use the attention mechanism to update the information of nodes [34]. The graph diffusion neural network implemented by random walk also achieves the same functions [35]. To better extract spatio-temporal information, researchers have integrated temporal models with graph convolution neural networks. Seo et al. proposed a temporal sequence model based on convolution spatial information termed GCGRU. The gated product in GRU is changed to a graph convolution operation to extract spatio-temporal features simultaneously [36]. Zhao et al. proposed a T-GCN model, in which GCN and GRU are stacked to extract spatial and temporal features respectively [27]. Graph models combined with other frameworks are also developed. Li et al. proposed a model to capture the spatial dependency using bidirectional random walks on the graph and the temporal dependency using the encoderdecoder architecture with scheduled sampling [37]. Liao et al. proposed a hybrid model in which spatial features extracted by GCN and the original features are integrated and fed into the sequence to sequence (seq2seq) structure. Disadvantages ? The system is not implemented The Hybrid Multi-Stream Feature Fusion Network. ? The system is not implemented data-driven adjacent matrix.
The system highlights how the proposed model tackles the challenges: • The system harness the power of GCN, GRU and FNN in a joint model that captures the complex nonlinear relations of the traffic dynamics observed from the road sensor network, which improves the model’s ability to express traffic features. • The architecture for feature extraction is parallelized instead of in cascade, which is helpful for accelerating the training and inferring process of the model. The main contributions of this paper are three-fold: • The system proposes a data-driven adjacent matrix instead of a distance-based matrix to map the road sensor network as a graph, which reduces manual design burden and achieves comparable performance than a distance-based approach. • The system constructs a multi-stream feature fusion module, in which a three-channel network is used to extract spatial-temporal and other features effectively, and the soft-attention mechanism is applied to integrate them. • The system balances the performance and complexity of the prediction model. Compared to the state-of-the-art methods in two real-world prediction tasks, our model can achieve comparable even better results within acceptable time complexity. Advantages ? In the proposed system, Attention-Based Multi-Stream Feature Fusion in which prediction accuracy is more. ? The proposed system developed an Effect of Graph Construction of Road Sensor Network in which datasets are accurate for predictions using classifiers.