Advertisement click classification method based on multi-scale stacking network

Abstract

The invention discloses an advertisement click classification method based on a multi-scale stacking network. According to the advertisement click classification method, combined features are automatically constructed through an MSSP structure for constructing multi-scale features based on different receptive fields; by constructing a plurality of observers with different angles and different visual fields, multi-scale features are stacked bidirectionally from two angles of depth and width, and high-order and low-order features in different local visual fields are mined, and the diversity of extracted features is ensured; in addition, the structure learns parameters through factorization, thus guaranteeing that high-order features can be effectively learned in sparse data. According to theadvertisement click classification method, the defect that LR, Wide & Deep excessively depend on manual construction of combined features is overcome; meanwhile, compared with traditional Poly2 and FM models, characteristics of different scales can be mined from multiple angles to guarantee the diversity of information learned by the model; and compared with the characteristic that the time complexity of models such as FFM is too high, the time complexity can be kept at the linear level, and the high requirement of online advertisements for time response can be met.

Description

technical field [0001] The invention specifically relates to a method for classifying advertisement clicks based on a multi-scale stacking network. Background technique [0002] The task of display ad click classification refers to predicting whether users will click on certain advertisements under a given user, product, and scenario. Accurate ad click classification can reduce the invalid delivery of advertisements, which is directly related to the revenue of the advertising platform and user experience. [0003] As the most classic classifier, LR (Logistic Regression) has the advantages of simple form, good model interpretability, and fast training speed, but it does not have the ability to automatically construct features and relies too much on artificially constructed features. The Poly2 model considers second-order combination features, however, if a feature combination does not appear in the training set, the weight of the corresponding item will not be fully learned,...

AI Technical Summary

Problems solved by technology

[0007] FFM (Field-aware Factorization Machines) introduces the concept of feature domain (Field) on the basis of FM model, and proposes a factorization machine for feature domain. Each feature will learn different hidden vectors for different feature domains. Model learning is more refined, but the problem is that the model is too complex and the time complexity is too high

Images