The Research Of Click-Through Rate Prediction Based On Feature Interaction And Behavior Sequences

As a key technology in online advertising and recommender systems,click-through rate prediction is used to predict the probabilities of users clicking on items.Its performance growth at 0.001-level can also bring considerable benefits to enterprises,so it has received extensive attention in academia and industry.Deep learning has been successfully applied to computer vision and natural language processing recently.Moreover,due to its great power of feature representation learning,it has been introduced into click-through rate prediction tasks.Clickthrough rate prediction based on deep learning has gradually become the main research method.By analyzing the relevant work of click-through rate prediction models and methods,we have found that mining useful information through features is crucial to click-through rate prediction.From the perspective of feature usage,deep click-through rate prediction models can be classified into two categories,i.e.,feature interaction modeling-based methods and behavior sequences modeling-based methods.The existing modeling methods based on feature interaction do not consider the semantic diversity of features and ignore the high-level semantic relationship.The existing modeling methods based on behavior sequences pay more attention to the item relevance and user representation,resulting in a single way to mine user interest.Therefore,we have put forward two models to solve the two problems.The proposed models can be used in such links as product recommendation,ad bidding,and payment based on CPC,which not only affects the revenue of businesses but also affects the satisfaction of users.They play an essential role in production and life.The main work of this paper is summarized as follows:Firstly,to capture the interaction relationship more deeply,we propose a click-through rate prediction model based on multi-semantic feature interaction,Me Fi Net.This model constructs multiple semantic spaces based on convolution operation,mines rich semantic information of different feature interactions in multiple spaces,and provides interpretability for feature interactions from a new perspective of semantic diversity.Moreover,in order to further improve the performance,the two-dimensional squeeze & excitation module based on SENet is proposed to learn the importance of different feature interactions in different semantic spaces.This module is used to dynamically learn the weight of the semantic space through the global semantic descriptor,then selectively emphasize important semantics and suppress less useful semantic features.We have conducted many experiments on three public datasets of Movie Lens-1M,Criteo,and Avazu and compared the proposed model with the relevant SOTA model.The multisemantic feature interaction solves the oneness of interaction modeling and improves the model performance by introducing more value information through rich semantic relations.Secondly,to mine users’ interests more comprehensively,we propose a click-through rate prediction model based on a multi-dimensional interest network,MIN.The model applies the idea of collaborative filtering to the sorting stage and innovatively summarizes three kinds of relationships between users and items: user-to-user,user-to-item,and item-to-item.It captures user interest patterns from group interests and individual interests.In addition,user behavior is modeled to learn user representation using Transformer,in which the global dependency among items is learned through the self-attention mechanism to obtain a deeper user representation vector.We have compared the proposed model MIN with the relevant SOTA model in the three datasets of Electronics,Movies & TV,and Grocery,then analyzed the importance of different interest sub-networks in the model through ablation experiments and discussed the impact of each critical parameter in the model.The multi-dimensional interest network solves the problem of the singleness of interest mining methods,and the modeling of group interest and individual interest makes the user profile more accurate,which is helpful for the model to judge the user’s click behavior.