| With the rapid development of the Internet,we are in an era of information overload.Faced with massive data,it is difficult for us to quickly gain insight and obtain valuable information.Recommendation system plays an important role in improving the efficiency of data utilization.Based on machine learning,the recommendation system analyzes users’ historical behavior records to understand users’ preferences,so as to actively recommend information that users are interested in.By understanding users’ preferences,it helps them dig potential demands from massive information and realize personalized recommendation function.Although the research on recommendation systems has been growing rapidly in recent years,most papers focus on developing machine learning models to fit users’ behavior better.However,the user behavior data that can be collected in reality are observational rather than experimental,and user behavior is influenced by factors such as self-selection,which makes a wide variety of biases in the data.Recommendation models capture these biases and even expand them in a loop,so blindly fitting the data without considering its inherent biases can lead to many serious problems that affect users’ satisfaction and trust in the recommendation.This paper mainly studies the debias method for recommendation system bias and applies the debias method to rating prediction task.In this paper,two sets of propensity scores are estimated using two model estimation methods,and three sets of propensity score are calculated from the observational data.The samples are weighted with all groups of propensity scores separately,and the de-biasing effect is improved through experimental demonstration.Since the inverse propensity score weighting method is sensitive to the propensity values,the results obtained by different estimators vary widely.Considering that the biased sample set is nonrandomly missing,filling the biased data set with missing values is also a common method of debiasing,and the missing value filling method requires high accuracy of filling values.In order to improve the stability and prediction accuracy of the model,an asymmetric tri-learning framework(AT)is introduced in this paper.The AT framework firstly pretrains three base models using the propensity score,and the classical matrix decomposition model is selected as the base model,and then two learners are used to estimate the missing values,and the filling is performed only when the difference between the estimation results of the two learners is not large,and the filling values can be obtained more reliably by this way.Finally,the prediction model is trained on the filled dataset.The experiments show that the AT framework has a significant improvement effect under the partial propensity score estimation method.In order to pay more attention to the items at the end of the longtail distribution,this paper proposes to improve the form of the propensity score by increasing the weight of the unpopular items so that the model can pay more attention to this part of the items.The experimental results show that this method can obtain a better debias effect. |
