Customer Purchase Prediction Based on Improved Gradient Boosting Decision Tree Algorithm

With the fast expansion in the number of Internet transactions, predicting customers' purchase behavior has become a significant auxiliary for e-commerce companies to develop accurate marketing strategies and improve the return on investment. In general, the data related to customers' purchasing behaviors have characteristics of large magnitude and high dimension. Companies often need to spend a high cost on implementing feature engineering for these data, but the quality of feature engineering cannot be guaranteed. Therefore, an improved gradient boosting decision tree algorithm, GBDT-NN, is proposed in this paper. Through the training of GBDT, features and feature combinations are automatically screened, and the artificial neural network is used to further process the screening results, so that more complex and nonlinear features can be learned adaptively to obtain better model performance. In this research, the effectiveness of the algorithm is verified on a real customer purchase behavior dataset. Experimental results demonstrates that the accuracy of GBDT-NN in the test set is 96.3%, which is improved by 10.6% compared with GBDT algorithm. The test AUC of GBDT-NN is 0.99, which is improved by 15.9% compared with the GBDT algorithm.