Research On Key Technologies Of Recommendation,Matching,Linking And Recognition For Video E-Commerce

With the development of e-commerce and online video platforms,video e-commerce has become a core business.It can not only expand the business scope of e-commerce companies but also help online video platform to convert the video viewing numbers into commercial rev-enue.In this paper,deep learning,statistical learning,and reinforcement learning are exploited to explore the key technologies of video e-commerce.The main innovation of this paper is summarised as:An online video advertising system named Video eCommerce++is proposed to exhibit appropriate product ads to particular users at proper timestamps of videos.First,an Incremental Co-Relation Regression(ICRR)model is novelly proposed to construct the semantic associa-tion between videos and products.ICRR is implemented in an incremental way to reduce time complexity.User Preference Diffusion(UPD)is employed to construct user-product associa-tions,which alleviates the problems of data sparsity and cold start.A Video Scene Importance Model(VSIM)is proposed to model the scene importance,so that ads can be embedded in the most attractive positions in the video stream.To combine the outputs of ICRR,UPD,and VSIM,a unified Distributed Heterogeneous Relation Matrix Factorization(DHRMF)is applied for online video advertising.Extensive experiments conducted on a variety of online videos from Tmall MagicBox demonstrate that Video eCommerce++significantly outperforms the state-of-the-art advertising methods,and can handle large-scale data in real-timeA novel deep neural network,called AsymNet,is proposed to explore a new cross-domain task,Video2Shop,targeting for matching clothes that appeared in videos to the same items in online shops.For the image side,well-established methods are used to detect and extract fea-tures for clothing patches with arbitrary sizes.For the video side,deep visual features are extracted from detected object regions in each frame,and further fed into a Long Short-Term Memory(LSTM)framework for sequence modeling,which captures the temporal dynamics in videos.To conduct exact matching between videos and online shopping images,a recon-figurable deep tree structure is utilized to model the similarity.Moreover,an approximate training method is proposed to achieve efficiency when training.Extensive experiments con-ducted on a large cross-domain dataset have demonstrated the effectiveness and efficiency of the proposed AsymNetAs the best of our knowledge,this paper is first to solve video hyperlinking from the sta-tistical perspective.The main contribution consists of two parts.First,to fully verify statistical properties are effective for video hyperlinking,the comprehensive simulations are conducted on the Blip 1000 dataset under various experimental settings.Through the qualitative and quan-titative analysis on the ground-truth of TRECVid 2016 and 2017,several insights are explored to satisfy the demand of popular,certainty and diversity for video hyperlinking.Second,a novel statistical framework is proposed to simultaneously exploit these statistical properties for low-risk automatic selection of anchors and targets.To fully evaluate our method,we con-duct comprehensive experiments.For the target selection,the experiment on the ground-truth of TRECVid 2016 and 2017 demonstrates our method significantly improves the performance of all the state-of-the-art baselines.Importantly,it also obtains the best Precision@K and Mean Average Precision(MAP)at TRECVid 2017A Generalizable Attribute Learning Model(GALM)is proposed to automatically design the neural networks for generalizable attribute learning.The main novelty of GALM is that it fully exploits the Multi-Task Learning and Reinforcement Learning to speed up the search procedure.With the help of parameter sharing,GALM can transfer the pre-searched archi-tecture to different attribute domains.In experiments,we comprehensively evaluate GALM on 251 attributes from three domains:animals,objects,and scenes.Extensive experimental results demonstrate that GALM significantly outperforms the state-of-the-art attribute learn-ing approaches and previous neural architecture search methods on two generalizable attribute learning scenarios.