Network Representation Learning And Its Application In Transportation

The improvement of Internet infrastructure has accelerated the process of information digitization.The available information in real life is no longer simply stacked,but represented as a graphical information network,such as transportation network,multimedia network and social network.Network data is generally heterogeneous,nonlinear and dynamic,which brings many challenges to advanced data mining tasks.Network representation learning is an effective method to solve these problems.It transforms high-dimensional sparse network information into low dimensional dense real-valued representation,thereby improving the efficiency of network analysis tasks.In real scenarios,network nodes often have rich attributes,which can establish semantic connections for nodes.However,most of the current sampling methods for attribute networks are based on attribute coding or traditional random walk,which will lead to problems such as large amounts of computation in high-dimensional environment and difficulty in capturing the node distribution characteristics.In addition,most network representation learning methods learn the low-dimensional representation of networks in Euclidean space,and it is difficult to retain the hierarchical information in complex attribute networks.Based on the above problems,this paper proposes a representation learning method named Adaptive Hyperbolic Attributed Network Embedding with self-adaptive Random Walks.It adopts the bias adaptive random walk strategy to efficiently sample the attribute network,and generate sampling sequences that can preserve the node distribution characteristics.Then,the hyperbolic skip-gram model is used to learn the node representation in hyperbolic space which can capture the hierarchical structure of complex network.The learned node representation is applied to the subsequent tasks,such as node classification,link prediction and so on.The experimental results show that compared with other comparison algorithms,the experimental results of this algorithm on the two real data sets are improved,which verifies the effectiveness of the algorithm.Using the network representation learning method to analyze and predict the traffic network is a feasible method to solve the real traffic problems.However,most of these studies lack the analysis and effective utilization of the characteristics of the traffic network,and their performance in the traffic analysis task also needs to be improved.To solve this problem,this paper proposes a traffic link attribute prediction method based on network representation learning,which uses the network topology and attribute importance score to learn the section representation,so as to improve the accuracy of traffic prediction.The experimental results show that,compared with other methods,the Macro-F1 score of this method in the prediction task of link function level and link speed limit level is improved by 17.99% and 9.9% on average.