GAFM: Learning the Weights of Feature Interaction via Graph Attentional Factorization Machine

Click-through rate prediction, which refers to predicting the probability of a user clicking on an ad or item based on input features, is critical in the development of recommendation systems. Traditional machine learning algorithms are less effective, while deep learning-based models that capture higher-order feature interactions can yield better results. However, these models ignore the capture of feature dependencies. In this work, we propose a new model, GAFM, which is based on a graph attention network, trained to obtain attention scores between features and other features. Then updates the features to capture feature dependencies flexibly on the one hand, and to filter out features with higher reusability for input into the Factorization Machines for feature interaction on the other. Our method undergoes rigorous evaluation on four real-world datasets, conclusively demonstrating its ability to boost CTR prediction accuracy. By considering feature dependencies and selecting pertinent features for interaction modeling, GAFM emerges as a promising approach to enhance recommendation system performance in CTR prediction. Additionally, this research advances deep learning techniques in recommendation systems and sheds light on the significance of feature dependencies and selection for precise predictive outcomes.