Research On Collaborative Filtering By Mining User Interaction Behavior

With the development of online services,more and more applications rely on recommendation systems to provide personalized recommendation services.The recommender system has produced huge economic benefits and has become a hot research point.As one of the most widely used recommendation algorithms,collaborative filtering algorithms are widely studied in industry and academia.Collaborative filtering models users’ preferences or demands based on interactive behaviors,and produces a recommendation list for each user.Due to data sparsity in interaction behaviors and the complex relationship between behaviors,the collaborative filtering algorithm still faces problems such as inaccurate user interest modeling and unfair results.To this end,we conduct research from three aspects:representation modeling,optimization strategy,and decision objective.First,we design a representation modeling method for mining user-item behavior graph structure.Second,we propose an optimization strategy that fully uses the relationship between behaviors.Third,we consider the relevance of user behavior and design a decision objective to enhance fairness.The main research contents mine user interaction behaviors from different levels.1)To solve inaccurate user interest modeling,we propose a representation modeling method based on linear graph convolution and residual structure.We treat the interaction behavior data as a graph structure,and use a graph convolution network to model user and item representation.The existing recommendation model based on graph convolution network suffers from training difficulty and the over smoothing problem.We solve the above two problems from two aspects.On the one hand,most recommendation models based on graph convolution networks have non-linear structures.We empirically show that removing the non-linear structures would enhance the recommendation performance with less complexity.On the other hand,we use residual structure to fuse multiple layer graph convolution networks,and the residual structure could help to alleviate the over smoothing problem.In this work,a linear and deeper graph convolution model(LRGCCF)is proposed by combining the linear graph convolution and residual structure.The experimental results show that LR-GCCF can effectively mine user-item graph structure information,and LR-GCCF is easier to train.2)To solve inaccurate user interest modeling,we propose an optimization strategy based on set to set ranking.The relationship between user-item interaction behavior is complex.We leverage the relationship between interaction behaviors from the set level,and propose a novel optimization strategy(Set2set Rank).Set2 set Rank samples observed set and unobserved set from implicit feedback data.Then Set2 set Rank explores how to fully use the relationship in each set and the relationship between the two sets.We introduce a two-level comparison of Set2 set Rank.The first-level comparison is item to set comparison,which encourages each observed item from the observed set to be ranked higher than any unobserved item from the unobserved set.The second-level comparison is the set to set comparison,which encourages a margin between the distance summarized from the observed set and the most “hard” sample from the unobserved set.In addition,we design an adaptive sampling method to select the observed set size flexibly,and further enhance the generalization ability of Set2 set Rank.Set2 set Rank method exploits the relationship between interaction behaviors to improve the optimization strategy.Set2 set Rank method is model-agnostic and can be easily applied to most collaborative filtering algorithms.Extensive experiments on three real-world datasets demonstrate the effectiveness of Set2 set Rank method.3)To alleviate unfair results,we propose a fairness-enhanced recommendation algorithm.Most recommendation algorithms take recommendation accuracy as the decision goal,ignoring the unfairness issues in the recommendation results.Most of the existing fairness algorithms assume the independence of entities,but there is a correlation between user behaviors in the recommendation task.We consider the correlation between user behaviors from a graph-based perspective,and propose a recommendation algorithm based on fair representation learning(Fair Go).Fair Go focuses on the decision process and is model-agnostic.Fair Go method learns fairness-aware representation of any pre-trained recommendation models.We design multiple filters and discriminators.Given the original representation from any recommendation algorithm,each filter transforms the original representation into a filtered representation space based on the sensitive attribute set,and each discriminator is a classifier to guess the sensitive attribute.The filter and discriminator are learned under the adversarial training process.The filtered representations do not leak sensitive attributes,so Fair Go achieves fair representation.Extensive experiments on two real data sets clearly show the effectiveness of Fair Go.