| The rise of the Internet age has brought forth an array of convenient services for people,while also producing a vast amount of data.As a result,individuals are now faced with the challenge of efficiently searching for content of interest from this massive information.Fortunately,the development of recommendation systems has provided an effective solution to address this issue.One essential task within recommendation systems is click-through rate prediction,which has become increasingly important in modern e-commerce and online advertising.This is because predicting clickthrough rates is crucial in increasing business revenue and enhancing user experience.In e-commerce systems,users engage in various behaviors,including clicking,favoriting,and purchasing,with clicking serving as the foundation for all of these actions.In online advertising,ad displays must be ranked based on cost per click,which is a product of ad bidding and the predicted click-through rate.As a result,click-through rate prediction research has garnered significant interest from both academia and industry.Various click-through rate prediction algorithms based on deep learning have been studied,focusing primarily on two aspects.First,click-through rate prediction algorithms based on feature interactions have been developed to capture higher-order interaction information of features,leading to more accurate predictions.However,such models often do not distinguish between interaction and independent information in features and use a feature vector for modeling,which limits the quality of the learned feature representation vector and ultimately restricts the model’s performance.Second,click-through rate prediction based on behavioral sequences models the click-through rate prediction task by mining the user interests embedded in historical sequences.However,such models do not always fully exploit the behavioral sequence data and may capture relatively homogeneous user interests.In reality,users often have multiple different interests at the same moment,necessitating the need to explore various interest patterns in historical behaviors to improve user experience and prediction accuracy.To address the above issues,the work in this thesis is summarized as follows:First,to address the issue of low-quality feature vector representations in feature interactionbased click-through prediction algorithms,which ultimately affects model performance.this thesis proposes a novel approach called the Feature Interaction Refinement Network(FIRN).FIRN addresses this problem by first designing a Mixture-of-Expert network to extract interaction information and independent information from each feature.Next,it employs a higher-order combination of adaptive explicit modeling of interaction information through a self-attentive mechanism,and a fusion gating unit to adaptively weight the interaction and independent information,ultimately resulting in a transformed feature vector representation.To extend the feature semantics as much as possible,FIRN introduces an extended gating unit to adaptively aggregate the original feature vector and the transformed feature vector representation,thereby solving the issue of adaptive feature semantic assignment and achieving high-quality learning of the feature representation vector.This provides high-quality feature representation input for the subsequent model prediction layer,ultimately resulting in better prediction performance.Second,to address the issue of insufficiently mining behavioral sequence data and capturing users’ interests in a relatively single way in click-through rate prediction algorithms based on behavioral sequences,this thesis proposes the User Behavioral Interest Network(UBIN).UBIN focuses on two different interest patterns in user behavior sequences.Specifically,UBIN employs the time-aware attention network and GRU to respectively capture the target user’s comprehensive interest in candidate items and their own dynamically evolving current interest.Furthermore,UBIN introduces triplet auxiliary loss to supervise the learning of current interests,improving the quality of its representation learning.Besides,based on UBIN,in order to capture the user interest patterns embedded in item behavior,this thesis also introduces the joint learning of Item Behavior Interest Network(IBIN)and UBIN,and proposes a Deep Multi-Interest Network(DMN).In DMN,the comprehensive interest and current interest output from UBIN are connected as a query and inputs to IBIN using a self-attention mechanism to model the user-user interest between the target user interest and the users who have clicked the candidate item,capturing the collaborative filtering information among users.Meanwhile,the popularity of candidate item at current moment is portrayed by aggregating the clicked time of candidate item through the attention mechanism,which reflects users’ unique preferences for popular items.Finally,the DMN output layer utilizes these multiple different interest representations as inputs,which greatly enriches the model expression capability and improves the model prediction performance.Finally,using a publicly available real-world click-through prediction dataset,this thesis analyzes the efficacy of the proposed approach and conducts experiments comparing it with various state-ofthe-art algorithms.The experimental results demonstrate the effectiveness and superiority of the proposed method for click-through prediction tasks. |
