Gossip Learning of Personalized Models for Vehicle Trajectory Prediction

Abstract

Gossip Learning (GL) is a peer-to-peer machine learning protocol based on direct, opportunistic exchange of models among nodes via wireless D2D communications, and on collaborative model training, which has recently proven to scale efficiently to large numbers of nodes, and to offer better privacy guarantees than traditional centralized learning architectures. Existing approaches to GL are however limited to scenarios in which nodes are static, or in which the node connectivity graph is fully connected, and they are fragile to node churn as well as to any change in network configuration. To overcome this limitation, we present a new decentralized architecture for GL suitable for setups with dynamic nodes, which benefits from node mobility instead of being hampered by it. In our approach, nodes improve their personalized model instance by sharing it with neighbors, and by weighting neighbors’ contributions according to an estimate of their marginal utility. We apply our GL algorithm to short-term vehicular trajectory estimation in realistic urban scenarios. We propose a new strategy for the estimation of the neighbors’ instances marginal utility, which yields satisfactory trajectory estimation accuracy for nodes with long enough sojourn times.