A new framework of data assessment and prioritization for real-time prediction of spatial information is presented. The real-time prediction of spatial information is promising for next-generation mobile networks. Recent developments in machine learning technology have enabled prediction of spatial information, which will be quite useful for smart mobility services including navigation, driving assistance, and self-driving. Other key enablers for forming spatial information are image sensors in mobile devices like smartphones and tablets and in vehicles such as cars and drones and real-time cognitive computing like automatic number/license plate recognition systems and object recognition systems. However, since image data collected by mobile devices and vehicles need to be delivered to the server in real time to extract input data for real-time prediction, the uplink transmission speed of mobile networks is a major impediment. This paper proposes a framework of data assessment and prioritization that reduces the uplink traffic volume while maintaining the prediction accuracy of spatial information. In our framework, machine learning is used to estimate the importance of each data element and to predict spatial information under the limitation of available data. A numerical evaluation using an actual vehicle mobility dataset demonstrated the validity of the proposed framework. Two extension schemes in our framework, which use the ensemble of importance scores obtained from multiple feature selection methods, are also presented to improve its robustness against various machine learning and feature selection methods. We discuss the performance of those schemes through numerical evaluation.

提出了一种用于空间信息实时预测的数据评估和优先排序的新框架。空间信息的实时预测对于下一代移动网络很有希望。机器学习技术的最新发展已经实现了空间信息的预测，这对于包括导航，驾驶辅助和自动驾驶在内的智能移动服务将非常有用。形成空间信息的其他关键要素是移动设备（如智能手机和平板电脑）以及车辆（如汽车和无人机）中的图像传感器以及实时认知计算（如自动车牌/车牌识别系统和对象识别系统）。然而，由于需要将移动设备和车辆收集的图像数据实时传送到服务器以提取输入数据以进行实时预测，因此移动网络的上行链路传输速度是一个主要障碍。本文提出了一种数据评估和优先级排序的框架，该框架可在保持空间信息预测精度的同时，减少上行流量。在我们的框架中，机器学习用于估计每个数据元素的重要性，并在可用数据的限制下预测空间信息。使用实际车辆机动性数据集的数值评估证明了所提出框架的有效性。在我们的框架中，有两种扩展方案，它们使用从多种特征选择方法中获得的重要性得分集合，还提出了针对各种机器学习和特征选择方法来提高其鲁棒性的方法。我们通过数值评估来讨论这些方案的性能。