Research On Recommendation Algorithms In Complex Data Scenarios For Industrial Internet

Recommender systems have become the foundation of e-commerce scenarios,and session-based recommendations have played an increasingly important role in recent years.Although previous works have achieved relatively good performance,these methods are still insufficient to achieve excellent recommendation performance due to the limited and probabilistically noisy information involved in the next click in each session.This paper studies the misleading accuracy of model learning caused by hard labels and the model training in the presence of data noise.Only considering the hard labels generated by sequential links in a session loses associations between items in a session and makes the model learn noisy information.To address this challenge,this paper introduces the concept of the soft label and proposes a recommendation model based on Self-Distillation Graph Neural Networks(SD-GNN).Specifically,in order to obtain reliable soft labels,this paper adopts the well-performing and reliable deep ensemble as the teacher model,which consists of multiple randomly initialized GNNs with good flexibility and scalability.Furthermore,this paper utilizes the soft label distribution produced by the teacher model to train each GNN in the ensemble model to implement the proposed self-distillation technique.Experimentally,extensive experiments on two public datasets show that the proposed method(SDGNN)outperforms other baseline methods proposed in the past three years,achieving a 10%improvement in recommendation accuracy.In addition,this paper also proposes a novel dual sparse attention network model to solve the problems of conversational recommendation systems in practical industrial Internet scenarios.Concretely,users’ current preferences are simulated by exploring the learned target item embeddings,and an adaptive sparse transformation function is used to remove the influence of irrelevant items.Empirically,compared with other baselines,the proposed method achieves 20.1%and 5.34%improvement in the MRR@20 and P@20 indicators in the actual session data set obtained in the Industrial Internet Identification Resolution search engine Project.