Research on the Fusion Model of DeepFM and XGBoost for Digital Consumer Behavior Prediction

To leverage the complementary characteristics of deep models and tree models in feature interaction modeling for digital consumer behavior prediction, this paper proposes a dual channel fusion model of DeepFM and XGBoost. In this model, an optimized DeepFM branch and an enhanced XGBoost branch are constructed using a feature shunting mechanism, and the dynamic weighted fusion and attention mechanism based on sample features are introduced to realize the adaptive combination of the two branch outputs. At the same time, a feature interaction enhancement algorithm is designed, which combines the depth implicit representation with the rule features of the tree model by multiplication, and further improves the depiction ability of high-order interaction. Experiments on real e-commerce user behavior and ad click through rate data sets show that the AUC of this model reaches 0.879, LogLoss drops to 0.342, which is 5.4% higher and 11.6% lower than DeepFM, and 8.3% higher and 14.7% lower than XGBoost, respectively. Ablation experiments verify the effectiveness of the dynamic weighted fusion and feature enhancement module, and the performance degradation is 6.7% and 10.5%, respectively. The robustness test showed that the AUC remained at 0.839 and 0.851 under the proportion of 30% missing features and 1% positive samples, and the click-through rate in online simulations increased to 4.93%, which was 1.6% higher than that of the industrial reference system. The proposed model has significant advantages in prediction accuracy and stability.