A Hybrid ANN–KNN Framework for Conversion Prediction in Digital Marketing

Abstract

With the continuous increase in the volume of user interaction data, it becomes essential to develop effective data-driven models to identify potential customers. This study proposes a hybrid modeling framework combining ANN and KNN for customer conversion prediction. Firstly, the original regression problem based on conversion rate was rephrased as a binary classification task and a threshold-based labeling strategy was adopted. Then, ANN was used to learn nonlinear feature representations from the input data. These learned features were further used as the input of the KNN classifier for the final prediction. To evaluate the proposed method, several models were compared on a digital marketing dataset, including an independent KNN model, an independent ANN model, and a hybrid ANN- KNN model. The experimental results show that the independent ANN model is significantly superior to the baseline KNN model, indicating the importance of nonlinear feature learning. Moreover, the hybrid ANN-KNN model achieved the best overall performance in multiple evaluation metrics. Particularly, the ANN[64, 32] + KNN model obtained the highest F1 score, demonstrating a good balance between precision and recall. Furthermore, this paper also analyzed the impact of ANN depth on model performance. The results showed that an appropriate network structure could achieve the best results. The feature importance analysis based on the leave-one-out feature cross-validation strategy further revealed that both user features and interaction behaviors contribute to the conversion prediction.