Intelligent Vehicle‑Cargo Matching, Driver Tagging, and Freight Price Prediction in a Logistics Big Data Platform

Shiqiao Group has built an intelligent logistics platform that digitizes all logistics processes, collects massive business data, and applies machine‑learning algorithms to create a high‑efficiency ecosystem for shippers, carriers, and drivers.

Logistics industry background : China’s road‑transport market reached RMB 6.2 trillion in 2020, with most volume in full‑truck‑load (FTL) and less‑than‑truck‑load (LTL) shipments. Traditional operations rely heavily on manual order passing, phone dispatch, and paperwork, limiting responsiveness.

Platform architecture : The platform consists of four layers – an application layer (sales, smart dispatch, cargo recommendation, image verification), an algorithm platform (providing models), a data warehouse (storing all business data), and a big‑data infrastructure (CDH/Hadoop, Spark, Doris, monitoring) that enables batch and near‑real‑time analytics.

Application cases :

Vehicle‑in‑transit tracking: GPS devices upload location every 30 seconds, allowing the platform to monitor routes, detect deviations, and identify overspeed events.

Real‑time dispatch center: aggregates live order volume, active drivers, shippers, and transaction amount to support rapid business decisions.

Intelligent vehicle‑cargo matching :

Problem formulation: match drivers (users) with cargo (items) in a given context, modeled as a click‑through‑rate (CTR) prediction problem y = F(X_i, X_u, X_c).

Model: DeepFM combines a Factorization Machine (low‑order feature interactions) with a deep neural network (high‑order interactions) to predict the probability of a driver clicking a cargo listing.

Evaluation: offline AUC, Top‑10 CTR, conversion rate (CVR), and order volume are used to select the best model before online A/B testing.

Driver tag system :

Cold‑start: assign initial tags based on simple rules such as copying cargo tags when a driver clicks a listing; weights decay over time using a formula inspired by Newton’s cooling law.

LSTM multi‑label model: after sufficient behavior data is collected, an LSTM processes variable‑length driver actions (clicks, calls, etc.) to predict multiple tags (origin, destination, vehicle type, length, cargo type). The model outputs a binary decision for each tag and is evaluated with precision and recall.

Freight price prediction :

Feature engineering extracts city, month, distance, vehicle length, and other attributes.

Cost decomposition: linear cost (distance, tolls) modeled with linear regression; periodic cost (weather, season) handled with appropriate models; time‑varying cost (driver labor) predicted by an LSTM.

Evaluation metric: SMAPE, achieving around 85 % accuracy within ±10 % price deviation.

Conclusion : By leveraging massive business data, DeepFM for cargo recommendation, LSTM for driver tagging, and specialized models for price forecasting, the platform creates a smart, data‑driven logistics ecosystem that improves matching efficiency, reduces manual effort, and lowers operational costs.