[2411.13057] Learning Multi-Branch Cooperation for Enhanced Click-Through Rate Prediction at Taobao

View a PDF file from the paper entitled Multiple Cooperation between researchers to predict rate

PDF HTML (experimental) view

a summary:The current clicking rate (CTR) has studied the role of feature interaction through a variety of technologies. Each technique of interaction of its own strength appears, and the use of one type usually leads to the restrictions of the model’s ability to capture complex features, especially for industrial data with massive input feature fields. Recent research indicates that effective stars models often combine the MLP network with a dedicated interaction network in a parallel structure. However, interactive dynamics and cooperation between the different flows or branches are still under discussion. In this work, we offer a new MBCNET network that enables multiple branches networks to cooperate with each other to better interact of complex features. Specifically, MBCNET consists of three branches: the expansion and transit features branch (EFGC) that enhances the ability to save the model for specific features, and a net net net rank to enhance cross -explicit features to improve generalization. Among these branches, a new cooperation plan is proposed based on two principles: the joint teaching of the branch and moderate differentiation. Joint teaching in the branch branch, which has been well learned, encourages the support of those learned poorly on specific training samples. It calls for moderate differentiation to the branches to maintain a reasonable level of difference in their advantages on the same inputs. This cooperation strategy improves learning by sharing mutual knowledge and enhances the discovery of various features of features across branches. Experiences on large -scale industrial data sets and A/B test via the Internet in the TAobao application are the superior MBCNET, which provides an increase of 0.09 points in the neurology, 1.49 % growth in deals, and a height of 1.62 % in GMV. The basic symbols are available on the Internet.