Rectifying Dual Bias For Recommendation

The emergence of recommendation systems is to alleviate the problem of information overload and help users quickly and accurately find some content of interest.Nowadays,recommendation systems play an important role in e-commerce,social media platforms,personalized advertising,and other fields,and become an indispensable part of modern Internet applications.Traditional recommendation systems mostly focus on how to better fit historical interaction data using machine learning models.However,the historical interaction data in recommendation systems is often observational rather than experimental.Observational user behavior data inherently contains various biases because the exposed items affect users’ behaviors,resulting in the observed data being confounded by the exposure mechanism of the recommendation systems.and free choice of user.Additionally,as a feedback loop,recommendation systems not only generate bias but also amplify them,leading to the Matthew effect.Therefore,traditional recommendation blindly fit the data without the consideration of its inherent bias will not only harm user satisfaction but also damage the interests of item providers.There are various biases in observed data,among which popularity bias attracts more attention recently.The popularity bias is moderately utilized to improve the accuracy of the model,but the gradually amplified popularity bias will cause serious damage to the recommendation systems.Therefore,the first issue is how to eliminate the popularity bias.If the biased observational data is directly used as a model input,the trained embedding vectors cannot truly represent the users and items,which leads to poor recommendation performance.Moreover,the second issue is how to leverage the popularity bias.Only removing popularity bias and overly recommending unpopular items will increase users’ antipathy.In recent years,researchers have shifted their focus to exploring the impact of bias on recommendation systems.However,currently,some models are proposed to handle popularity bias only from the perspective of items,which do not concern about the information from users.In response to the aforementioned situation,this thesis proposes a novel recommendation model to rectify the popularity bias from the perspective of items and users based on a causal graph in accordance with the data generation.The validity of the proposed method is verified on the large-scale real data sets.The main contributions of this thesis are summarized as follows:· In order to eliminate the popularity bias that has a negative impact on the recommendation system,this thesis proposes a double bias deconfounding(DBD)model.Inspired by the causal graphs,the DBD model analyzes the impact of the bias on the recommendation system from both the item and user perspectives.In the training phase,the DBD model is constructed using the backdoor adjustment technique in causal inference to remove bias from dual perspectives.The Bayesian personalized ranking loss function is adjusted to simultaneously utilize information from both item popularity and user activity to train the model.Experimental results show that the proposed method can improve the performance of the recommendation system by removing the negative impact of popularity bias during training.· A Double Bias Deconfouding and Adjusting(DBDA)model is proposed to leverage favorable popularity bias in order to eliminate the distribution shift between the training and testing sets.The DBDA model is built on top of the DBD model.In the inference phase,the model does a bidirectional intervention from items and users.Namely,the intervention probabilities are modified for each item given to a user.When recommending different items for the same user,bidirectional intervention is equivalent to popularity intervention for items.Three different formulas on the prediction for the interventional probabilities are compared to investigate how different formulas affect the final recommendation performance.The thesis explores the effects of different popularity prediction formulas on popularity intervention based on the trend of popularity changes.In the inference phase,the model adjusts the popularity bias to identify items that users are potentially interested in.Many experiments,including ablation experiments,demonstrate the effectiveness and stability of the proposed model and the necessity of each component.In summary,this thesis first constructs a causal graph to guide the model in removing popularity bias from both the item and user perspectives during the training phase.Then,from a dual-perspective approach,interventions are made to mitigate bias and identify potentially popular items.The use of popularity bias helps alleviate the problem of distribution shift between the training and testing sets,making the model more robust and interpretable.