毕业生知识库

As the management of transportation transitioning from industrial technology to information technology, Intelligent Transportation System (ITS) has made transportation safer and more efficient. However ITS has encountered technology bottlenecks with the development of societies and transportation systems. Therefore, management of transportation systems needs parallel intelligence, which will lead a transition from information technology to intelligence technology. More specifically, the application of parallel intelligence in transportation is called parallel transportation, which is based on software-defined transportation systems, O2O (online to offline and vice versa) computational transportation experiments, and parallel transportation with knowledge automation for closed-loop control and management with society-wide feedback.

Intelligent processing and analyzing of data is an elemental function of parallel transportation, which is of great theoretical significance and practical value to accelerate the researches and applications. Previous work in ITS focus on the entities of physical space, parallel transportation gains intelligence by extending the exploration to artificial space and social space. More specifically, structural data collected from physical detectors are main entities to process in ITS. It is very expensive and even impossible to collect big and accurate data. Therefore, we investigate the alternative artificial data to be used in data processing and data-driven modeling. Further, we investigate the processing and analyzing of the unstructured data generated by social sensors. The major works and achievements consist of four aspects:

1. A method of traffic data generation based on generative adversarial networks (GANs) is proposed. As it is difficult for the standard GAN algorithm to generate time-dependent traffic flow data, we made twofold modifications: (i) Using the real data or the corrupted ones instead of random vectors as latent codes to generator within GANs, and (ii) Introducing a representation loss to measure discrepancy between synthetic data and real data. Experimental results on a real traffic dataset demonstrate that our method can significantly improve the performance of traffic data imputation.

2. An approach of traffic data imputation based on parallel data paradigm is proposed. To develop traffic data imputation models with high accuracy, traffic data must be large and diverse, which is costly. An alternative is to use synthetic traffic data, which is cheap and easy-access. We train GANs model to generate artificial traffic data and design imputation models for the tasks consisting of discrete corrupted points and continuous corrupted points, respectively. We perform extensive experiments on real traffic dataset to investigate the effectiveness of the proposed method. Experimental results show that our method leads to significant improvements.

3. An approach of traffic prediction based on parallel data paradigm is proposed. Taking the dynamic temporal behavior for a traffic data sequence into consideration, we propose to apply long short-term memory (LSTM) model to predict traffic data.  As developing traffic prediction models with high accuracy, traffic data must be large and diverse, which is costly, therefore, we alternatively use synthetic traffic data to train traffic prediction models. For live traffic condition prediction, we compared the performance of the proposed Stacked LSTM model with multilayer perceptron model, decision tree model and support vector machine model. Experiments show the proposed method is superior to the competing methods. For traffic flow prediction, we perform extensive experiments to investigate the parallel data method, and experimental results show that our method leads to significant improvements.

 4. An approach of traffic relevant information extraction is proposed. We focus on a specific problem in social media mining which is to extract traffic relevant microblogs from Sina Weibo, a Chinese microblogging platform. It is transformed into a machine learning problem of short text classification. Firstly, we apply the continuous bag-of-word (CBOW) model to learn word embedding representations based on a dataset of three billion microblogs. Compared to traditional one-hot vector representation of words, word embedding can capture semantic similarity between words and has been proved effective in natural language processing tasks. Next we propose using convolutional neural networks (CNNs), long short-term memory (LSTM) models and their combination LSTM-CNN to extract traffic relevant microblogs with the learned word embeddings as inputs. We compare the proposed methods with competitive approaches, including the support vector machine (SVM) model based on bag of n-gram features, SVM model based on word vector features, and multi-layer perceptron model based on word vector features. Experimental results show the effectiveness of the proposed deep learning approaches.