Role-Guided Graph Neural Recommendation in User-Generated Content Scenarios
Personalized recommendation is a vital and indispensable tool at overcoming the infor-mation overload problem.It can help users to find their desired information and assist content providers to obtain fruitful profits.It has been widely deployed in various fields,such as news,e-commerce applications,location-based services,etc.In these classical scenarios,a user generally plays a single role(i.e.,customer).With the rapid development of the Internet,we have witnessed a revolution in information production and transmission manner.In this case,there is a new way to access the Internet,i.e.,user-generated content(UGC).Owing to the advantages of rapid spread and easy access,it has been an important fashion of information propagation.In contrast with traditional recommendation scenarios,a user in UGC plays dual roles:consumer and producer.When building a personalized recommendation,we consider not only the consumer-item interactions,but also the impact of a producer for user decision making.Thereby,how to sufficiently capture the relationships of consumer-item and consumer-producer is the key to an effective recommendation.In the UGC-based recommendation studies,CPRec is the most representative model.Despite its effectiveness,we argue that it still has two defects.On the one hand,in perspective of model construction,it fails to explicitly exploit high-order connectivity behind consumer-item interactions and consumer-producer relationships.As a result,it is non-trivial to learn high-quality embeddings for all users and producers.On the other hand,in perspective of model optimization,it fails to differentiate the importance of each observed instance during training,which results in the suboptimal recommendation results.In light of these two defects,we propose a Role-Guided Graph Neural Recommendation(RGNRec)for the personalized ranking task in UGC scenarios.Specifically,we first construct a consumer-item interaction graph and a consumer-producer interaction graph based on the users'historical behaviors and the item's producer information.Furthermore,in order to explicitly capture the high-order connectivity,we design a dual-channel linear propagation module upon both graphs.In this way,our solution simulates the diffusion process of user interest and producer influence.Lastly,we contribute an adaptive weight strategy for the non-sampling loss function,and view the overall training procedure as a bi-level optimization problem.The key contributions of this paper are as follows:(1)introduce the dual-channel linear propagation module,which explicitly disentangles the users'motives behind individual taste and the influence of a producer;(2)propose an adaptive weighted non-sampling loss function,which could adjust the weight coefficient of each observed instance at different training periods.We choose the classical recommenders and several recent state-of-the-art graph neural networks as baselines and perform extensive experiments under three UGC scenarios:Pinterest,Recipes,Reddit.In terms of overall performance comparison,RGNRec consistently and significantly surpasses all baselines in model effectiveness and training efficiency,especially improving Precision@10 by 4.31％-17.83％.Then,ablation studies validate the rationality and necessity of the dual-channel linear propagation module and the optimization mechanism with an adaptive weight strategy.Finally,the related experiments demonstrate the strengths of our proposed method in alleviating data sparsity and user cold-start issues.
NETL
NSTL
万方数据
