Research On Cross-domain Recommendation Algorithm Based On Graph Neural Networ

The rapid development of the Internet and wide-area applications had resulted in explosive growth in the amount of information,which greatly inhibited the efficiency of users’ processing and utilization of information.In the face of massive and intricate information,recommendation algorithms could mine user-related content preferences based on historical interaction data and selectively display them.The cross-domain recommendation algorithm aimed to use user interaction information in multiple fields to assist in mining user preferences,outline a relatively complete user portrait,and then improve the quality of recommendations.However,compared with traditional recommendation algorithms,the design of cross-domain recommendation algorithms faced more challenges.First,noise data under the problem of data sparsity further reduced the accuracy of recommendation performance;second,the insufficient utilization of data attributes made it difficult for algorithms to capture cold-start user preferences and correlations between different fields;and when the conventional heterogeneous graph network architecture was applied to the cross-domain recommendation model,the scalability was poor and the model complexity was high,which led to high computational cost and poor interpretability.In addition,compared with traditional machine learning models,graph neural networks could represent non-Euclidean spatial data in a topological form,and could demonstrate a more powerful ability to capture complex relationships between data by virtue of their special graph structure and message passing aggregation method.Therefore,this article was based on the requirements of the big data user platform,based on GNN,to complete the construction of a cross-domain recommendation algorithm,so as to achieve personalized and precise recommendations.Aiming at the data sparsity problem,this paper proposed Cross-Domain Recommendation based on Graph Representation Optimization,CDRGO.Considering the constraints of noisy data on recommendation performance,CDRGO fully considered the constraint relationship between data sparsity and noise data deletion,and used graph structure characteristics to balance the impact of weakening noise and the improvement of recommendation performance.CDRGO adapted the graph neural network to build interactive data models for user items in the source domain and target domain,projected the embedded feature representations to the potential space,and performed outlier detection and deletion on the source domain item graph representations,and finally the cross-domain information aggregation and prediction of the neural network was carried out through joint training.In addition,for the cold-start user problem,this paper proposed Transfer Learning method based on Inter-Domain Feature Transformer of Visual Elements,TTVE.TTVE used the visual elements of the project as user attributes to further mine user preferences,and designed an inter-domain feature converter based on visual element features to implement transfer learning.In the process of model training,TTVE first captured the cover image style characteristics of the items through the convolutional neural network,and then input them as user attributes into the graph neural network in the source domain and the target domain respectively,and modelled the user embedding features.Finally,the source domain user embedding features and visual element features were used as input,and the target domain user embedding features were used as output to train the inter-domain feature transformer to capture the correlation between different domains.The experimental results showed that,compared with other typical cross-domain recommendation algorithms,the cross-domain recommendation algorithm proposed in this paper had higher reliability and effectiveness.