Representation Learning With Implicit Feedback Data For Personalized Recommendation

Recommendation based on implicit feedback is a classic task in recommendation system.It refers to the use of users’ implicit feedback behavior,such as click and purchase,to provide users with a list of recommended items.The existing algorithms used to solve this task usually have two problems:(1)the use of potential negative examples of errors,resulting in a decline in the recommendation effect;(2)The implicit feedback data used is sparse,leading to the lack of supervision signals of the model,the difficulty of optimization and the decline of effect.Therefore,this paper proposes two new recommendation frameworks to solve the above problems.This paper proposes an Interaction Domain Aware Affinity and Uniqueness Learning for Recommendation(IAUL).IAUL uses a learning objective based on synergy and specificity to avoid introducing potential negative examples of errors,and uses a graph convolutional network that can capture fine-grained features inside the interaction domain to help the model learn better using this objective.In addition,IAUL designed an auxiliary self-monitoring task for this scenario to provide additional monitoring signals to alleviate the data sparsity problem.The experimental results show that IAUL achieves better performance than the baseline method in three real scenario data sets,and effectively alleviates the data sparsity problem.This paper proposes Multi-behavior based Graph Contrastive Learning for Recommendation(MGCL).MGCL introduces a variety of user behaviors to assist target behavior learning to alleviate the problem of data sparsity.This framework adopts a negative example free optimization framework based on contrast learning,which can simultaneously avoid the problem of effect decline caused by improper negative examples and the problem of model learning collapse caused by only positive examples.In order to better capture and utilize the features of multi-behavior data in this multi-behavior no-negative example optimization framework,MGCL adopts graph convolution network which can simultaneously capture the importance of multi-behavior to user preference,thus better solving the problem of data sparsity and improving the recommendation effect of target behavior.The experimental results show that MGCL’s recommendation effect in the three public data sets is better than the existing multi-behavior and single-behavior methods.In addition,MGCL has significant advantages in data sparsity.