Research On Multi-behavior Product Recommender System For Sparse Interactive Data

In the modern society with the rapid development of Internet technology,users are mainly concerned about how to find the part they are interested in from a large amount of information,while platforms need to make the information produced by themselves stand out and attract the attention of the majority of users.For e-commerce platforms,the purchase behavior of users directly determines their overall revenue.These two reasons promote the development of recommender system.In order to improve user retention and the number of merchants settled in,the platform can establish the relationship between users and information through the recommendation system,and use users’ historical interaction data as a medium to strategically allow users to actively or passively choose products that are valuable to them,which greatly promotes the upgrade of platforms and deep consumption of users.In real scenarios,cold start and sparsity of the sample data will have a negative impact on the effect of the recommendation system.In order to solve this problem,the main work of this thesis is:ⅰ)Relying on the background of e-commerce recommendation,we design and build a multi-behavior recommendation model named Integration-Augmented Transformer Graph Network(IATG).By incorporating other types of behavioral features between users and products except purchase behaviors to alleviate the negative impact of the sparsity of the target behavioral data,IATG makes predictions for the user’s interactive products in the target behavioral mode and obtains a list of recommendation.At the same time,we compare IATG with several benchmark models,explaining the reasonability of structure of the model and design concept from various aspects.Experiment shows that IATG achieves an average performance improvement of 4.5%on both evaluation metrics compared to the suboptimal baseline.ⅱ)we implement a multi-behavior product recommendation system based on hierarchical collaboration,which takes IATG as the core and combines the original information and characteristic corpus of each object element.We design offline training and online inference from the system level,and test the performance of each module.After validation,this system realizes the overall process from the client initiating the request to the server returning the product recommendation list,and has good recommendation effect and applicability in the environment of sparse interactive data.