Research On Image Representation For E-commercial Visual Search

Visual search,which allows users to input images to search for relevant images,is an emerging search mode.Visual search in e-commerce is often combined with online advertising,and image feature is fatal to the entire process of online advertising.Apparently,it is crucial for computing relevance for image retrieval tasks.For the other thing,as part of product information,image feature is also an important feature for click-through rate(CTR)prediction in the ranking stage.First,image representations got by image classification tasks have many shortcomings,such as coarse granularity,high cost of manual annotation,and high noise.In response to these problems,this thesis proposes visual encoding framework,first pretraining with a self-supervised method,and then fine-tuning with user interaction behavior.These two stages are both based on contrastive learning.Thus,fine-grained and more expressive visual features are obtained.Offline experiments on a real billion-scale dataset show the effects of the method.Second,this thesis finds and proposes one of problems caused by introducing user behaviour data,which is sample selection bias.In order to ease this problem,this thesis designs a debiasing network.It's based on contrastive learning,and aims to reconstruct the extracted visual features to bring similar lowimpression and high-impression image representations close.Then splice the unbiased visual features with other features in the CTR model.The debiasing network and CTR prediction model are jointly trained.This method reduces the sample selection bias,while ensuring the accuracy of CTR prediction.In this thesis,we verify the effects of the method on both real datasets and online experiments.