Click-through rate prediction model based on LightGBM and DeepFM

Abstract

Aiming at the problems of information overload and insufficient personalized service in multi-service system, a click-through rate prediction model based on LightGBM and DeepFM (LGDF) for multi-service systems is proposed. The LGDF model is based on the framework of LightGBM and DeepFM model. Firstly, LightGBM gradient lifting decision tree is added to the model to perform high-order combination feature transformation and fusion extraction on the features in the original dataset to obtain effective integer result vectors. Then, the integer result vector generated by LightGBM tree prediction is spliced with the original data set to form the new dataset. Finally, the new dataset is used as the input of the DeepFM model to learn the combination relationship of high-order and low-order features between the data. The proposed model is verified on the public dataset Criteo, and the experimental results show that the proposed model LGDF has higher accuracy than other classical models.