Multi-source Heterogeneous Data Fusion And Representation For User-generated Content On E-commerce Platforms

The fusion and representation of user-generated multi-source heterogeneous data have attracted wide attention in various fields.For example,user's comments and consumption behaviors in e-commerce platforms can be taken as important basis for merchants to provide personalized services.At present,the fusion representation of multi-source texts is relatively simple and seldom considers the differences between long and short texts,so the representation accuracy needs to be improved.In addition,recently proposed deep learning methods can map various structural data into the same shared space when conducting heterogeneous data fusion.However,few studies have concentrated on the fusion of user-generated contents in e-commerce platforms.Motivated by these,this paper studies the fusion and representation strategy of multisource heterogeneous data for user-generated content on e-commerce platforms.The main contents are as follows:(1)The Fusion Representation of User-generated Text Data on E-commerce Platforms: First,according to the length and characteristics of the comments,the usergenerated multi-source text is divided into long and short texts,and a Doc2 vec and the Latent Dirichlet Allocation(LDA)model based fusion strategy is proposed to extract short text features.Then,the Pearson similarity between the commodity review and description text is utilized to select the items to be fused and determine the minimum number of texts required to represent the features of commodity.Finally,the proposed algorithm is applied to the classification of multi-category items in Amazon's dataset.Experimental results show that the proposed algorithm can effectively improve the accuracy of multi-source text fusion and alleviate information overload.(2)The Fusion Representation of User-generated Multi-source Heterogeneous Data on E-commerce Platforms: Based on the research content(1),the fusion representation of user-generated heterogeneous data,including the comment texts and description images is further considered.First,the pre-trained Residual Network(Res Net)is utilized to extract the feature of commodity images.Then,according to the complementarity of images and texts,the representations of comment text obtained through the methods proposed in content(1)and features of images extracted by Res Net are combined.A discrete convolution fusion algorithm is further proposed to conduct the fusion of multi-source heterogeneous data.Finally,the proposed fusion algorithms are applied to the extended dataset of Amazon's data,and the experimental results illustrate the effectiveness of the proposed fusion algorithm.(3)Personalized Recommendation Based on User-generated Multi-source Heterogeneous Data Fusion Representation: The research contents(1)and(2)are applied to personalized recommendation to further illustrate their effectiveness.Firstly,based on the fusion representation of multi-source heterogeneous data features and commodity category attributes,a Restricted Boltzmann Machine(RBM)model is constructed to estimate the preference of users.Then,the collaborative filtering algorithm,which is jointly learned from the user's implicit and explicit preference,is adopted to complete the personalized recommendation.Finally,the proposed algorithms are applied to the personalized recommendation in multiple domains of Amazon.The experimental results illustrate that the product description based on the fusion representation of multi-source heterogeneous user-generated contents can effectively improve the accuracy of recommendation and provide more interpretability for the recommendation results.This thesis contains 18 figures,14 tables and 93 references.