Research On Link Prediction In Co-travel Networks

With the development of modern passenger transportation industry and the improving of people’s living standard, people travel more and more. And sometimes people travel with other people, such as friends, family, people in the same group etc. These co-travel relationships could build the passengers’ co-travel networks. Predicting the hidden or future links of the co-travel networks can lay a solid foundation for personalizing passenger transport services. And it also can support the transport industry market decision.In fact, the problem of predicting co-travel relationships between the passengers is also called link prediction in complex networks. In recent years, link prediction has been concern many scholars, and it has a wide application in some fields. In this paper, we use the passenger travel data to construct the passenger co-travel networks, and then build a co-travel relationship prediction model based on supervised learning. The model construct three types of predict characteristics, which are the individual information and historical behavior features, the microscopic network features and the medium network features. Among them, the individual information and historical behavior features are extracted from the individual information and the history travel records of the passengers. The microscopic network features are focus on the local network topology around the pair of passengers. And they include features based on common neighbor and the features based on the subnet of the all neighbors of the pair of passengers. Then we use community detection algorithm to detect the hierarchy communities of the co-travel networks, and construct a series of medium network features. Finally we use the classification method to build the prediction model for predicting the future co-travel links.We do experiments in a real passenger travel dataset. We selected two type sample sets, which are called the random sets and the two hop sets. The experimental results show that the features we designed are very effective for the link prediction in co-travel networks. Specially, the precision of random sets reached 92%, which show the prediction model we built is excellent for the link prediction in co-travel networks.