Personalized Recommendation for Vehicle-cargo Matching: A Study of YMM Freight Platform Data

Vehicle-cargo matching(VCM)in the network freight platform is responsible for connecting drivers and cargo owners, improving the supply and demand matchingefficiency of vehiclesources and cargo sources, and reducing information asymmetry in the market. However, this task is quite complex due to the involvement of multiple parties' information such as drivers, vehicles, cargo sources, cargo owners, and context, as well as specific application scenarios and optimization goals. The existing research on vehicle-cargo matching mainly focuses on credit evaluation, route optimization, freight rate prediction, and other aspects. There are relatively few studies on personalized recommendations for vehicle-cargo matching, and there are limitations. This paper studies the general process of vehicle-cargomatching based on big data analysis and proposes a personalized recommendation model for vehicle-cargomatching in the Full TruckLoad(FTL) transportation mode. Specifically, to better integrate the multi-party features involved in vehicle-cargomatching, a specialfeature fusion module is designed to capture the importance of different features. The feature fusion module coordinates the multi-head attention network and the SENet network to fine-grained extract the significant information beneficial to the current task, and then introduces the bilinear interaction layer for low-order feature interaction learning. To learn the patterns behind the historical transaction records of vehicle-cargo matching, a high-orderfeature interaction learning module is designed to discover complex feature combination patterns. The high-order feature interaction learning module coordinates the CrossNetand MLP to explicitly and implicitly fit the nonlinear correlations among multiple features such as drivers, driver behaviors, cargos, cargo owners, and context, respectively. To evaluate the validity of the proposed model, it is verified using the data from the YMM freight platform. The experimental results show that the proposed model achieves the best effects of 0.8149 and 0.1970 respectively in the AUC and LogLoss metrics, outperforming all the baseline models. This study can not only improve the precisionof vehicle-cargo matching and the loading efficiency of freight platforms, but also establish new methodological boundaries for technology enabled management practices to a certain extent.