A Click-through Rate Prediction Method Based on Interaction Features Extraction for High-dimensional Sparse Data

The dataset for Click-through rate prediction contains a large number of categorical features, and the high-dimensional sparsity presented after encoding these features will take challenges for the prediction task. Existing click-through prediction methods focus on the effective learning of interaction features of different orders, ignoring the influence of the importance of interaction features on the prediction results. This paper proposes a click-through rate prediction method based on ECANet and DeepGBM for high-dimensional sparse data, called ECA-DeepGBM. Specifically, this method improves FNFM, called ECA-FNFM, and establishes it as the base model for CatNN in DeepGBM to extract categorical features. In ECA-FNFM, we introduce ECANet attention mechanism into the feature interaction layer of FNFM, to dynamically capture the importance weight of different interaction features. Ultimately, ECA-DeepGBM model is constructed by CatNN, the improved ECA-FNFM, and GBDT2NN. Experiments were conducted on three public datasets, and the results show that the proposed method achieves better performance.