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.

Traffic flow prediction is an important functional component of Intelligent Transportation Systems (ITS). In this paper, we propose a hybrid deep learning approach, called graph and attention-based long short-term memory network (GLA), to efficiently capture the spatial-temporal features in traffic flow. Firstly, we apply graph convolutional network (GCN) to mine the spatial relationships of traffic flow over multiple observation stations, in which the adjacent matrix is determined by a data-driven approach. Then, we feed the output of the GCN model to the long short-term memory (LSTM) model which extracts temporal features embedded in traffic flow. Further, we implement a soft attention mechanism on the extracted spatial-temporal traffic features to make final prediction. We test the proposed method over the PeMS data sets. Experimental results show that the proposed model performs better than the competing methods.

Reinforcement learning (RL) can automatically learn a better policy through a trial-and-error paradigm and has been adopted to revolutionize and optimize traditional traffic signal control systems that are usually based on handcrafted methods. However, most existing RL-based models are either based on a single scenario or multiple independent scenarios, where each scenario has a separate simulation environment with predefined road network topology and traffic signal settings. These models implement training and testing in the same scenario, thus being strictly tied up with the specific setting and sacrificing model generalization heavily. While a few recent models could be trained by multiple scenarios, they require a huge amount of manual labor to label the intersection structure, hindering the model’s generalization. In this work, we aim at a general framework that could eliminate heavy labeling and model a variety of scenarios simultaneously . To this end, we propose a general Scenario-Agnostic (GESA) reinforcement learning framework for traffic signal control with: (1) A general plug-in module to map all different intersections into a unified structure, freeing us from the heavy manual labor to specify the structure of intersections; (2) A unified state and action space design to keep the model input and output consistently structured; (3) A large-scale co-training with multiple scenarios, leading to a generic traffic signal control algorithm. GESA can automatically handle various structured intersections from various cities without human labeling, and it co-trains a generalist agent to control traffic signals for multiple cities together, which also demonstrates superior transferability in zero-shot settings. In experiments, we demonstrate our algorithm as the first one that can be co-trained with seven different scenarios without manual annotation and gets 13.27% higher rewards than baselines. When dealing with a new scenario, our model can still achieve 9.39% higher rewards. The code, scenarios, and demos are available https://github.com/bonaldli/GESA.

With the widespread application of mobile phones, it has become possible to study human mobility and travel behaviors based on cellular network data. Contrary to call detail records, the data is triggered by mobile cellular signaling and can provide fine-grained information about users’ daily routines. However, it does not explicitly provide semantic details about traveling traces, e.g., trip purposes. In this paper, we propose a methodological framework to handle large-scale cellular network data and discover the underlying trip purposes in an unsupervised way. We first devise heuristic rules to identify home/work purposes. Then, a flexible latent Dirichlet allocation (LDA) model is presented to discover the activities for remaining trips, in which each trip is depicted by four attributes, i.e. arrival time, age group, stay duration, and the point of interest tag for the destination. Experimental results show that the proposed method can identify diverse trip purposes by explaining their structures over trip attributes and outperform baselines in terms of log-likelihood and perplexity. We also analyze the difference between the automatically discovered trip purposes and those estimated from household census, and the analyzed results demonstrate the feasibility of our proposed method.

As an essential component of Internet of Things, GPS-enabled devices record tremendous digital traces, which provide a great convenience for understanding human mobility. How to discover transportation modes efficiently from such valuable sources has come into the spotlight. In this article, the transportation mode detection is treated as a dense classification task, and a similarity entropy-based encoder-decoder (SEED) model is proposed. We first design an encoder-decoder backbone for end-to-end mode detection. Then, a semi-supervised learning module based on similarity entropy is proposed to exploit numerous unlabeled data. Specifically, we stack several convolutional layers as an encoder to capture hierarchical features from fixed-length trajectories, and then adopt transposed convolutional layers as a decoder. For a semi-supervised module, inspired by entropy regularization, we use the ${K}$ -Means algorithm to cluster prototype vectors from the encoder’s predictions. We then fine-tune the encoder by sharpening the similarity distribution between unlabeled predictions and prototypes, aiming to make the former close to one prototype only while staying away from others. A majority-voting post-processing method is used to alleviate jitter impact when inferring. The Experimental results show that SEED significantly outperforms segmentation-then-inference methods. Furthermore, the similarity entropy-based module can improve the generalization performance of the model, and the metrics such as intersection over union can be increased by 5% over baselines. All of these verify the superiority of our method.