<p>Road type prediction, as one of the solutions to improve semantic quality, plays a crucial role in addressing the quality issues of OpenStreetMap (OSM) data. However, due to issues of data sparsity and imbalance, most existing algorithms lack accuracy or require external data support. This paper proposes an OSM road prediction method based on the eXtreme Deep Factorization Machine (xDeepFM) model, which predicts the types of geographical objects by integrating the semantic and spatial features of roads, without relying on external data. The data contributed by volunteers is first classified according to OSM classification rules, after which features are extracted and encoded from the classified data. Subsequently. The xDeepFM model is then deployed to facilitate the interaction learning of both low-order and high-order features. Road type prediction involves a multi-class prediction problem. To address the issue of data imbalance during the prediction process, we design a multi-level classifier based on Binary Classification Feedforward Neural Network (BCFNN). Experiments conducted using the OSM data indicate that our method achieves high accuracy in predicting road types for the selected area, outperforming existing methods. Our approach does not rely on external data, but is entirely based on the semantic and spatial features inherent to OSM data. This enhances its scalability and compatibility, providing valuable insights and approaches for related research.</p>