Generalization-Aware Optimal Mixture Models of Inter-Contact Times for the Internet of Vehicles

The inter-contact time (ICT) is one of the important indicators that characterize the Internet of Vehicles (IoV), and its distribution is of great significance to the formulation of the network protocol and the improvement of network performance. It is more valuable and appropriate to reveal the distribution by mixture distribution models. However, choosing the optimal and generalizable mixture models remains an unsolved problem. So a framework for finding generalization-aware mixture models with the suitable number of branches and model patterns is proposed in this work, which mainly uses the Expectation Maximization (EM) algorithm for parameter estimation. Based on the big data of the real vehicle trajectory of Beijing taxis recorded by Global Positioning System (GPS), we conduct modeling experiments on the empirical ICT dataset, considering various combinations of the four distributions of normal, lognormal, exponential and power-law in the previous conclusions. The result shows that the 3-branch mixtures can best balance accuracy and simplicity, in which the 3-branch lognormal mixture model and 2-branch lognormal 1-branch exponential mixture model fit well with different spatial, different temporal scales as well as different vehicle density conditions. The theoretical relationship between such optimal models and traffic flow remains to be further explored.