Research Of Unbiased Learning To Rank In E-commerce Recommendation Scenario

The product recommendation system is used by e-commerce websites to recommend personalized products for users.Among the system,the product ranking module plays a decisive role.This module often uses learning-to-rank algorithm to rank candidate products and output the products sequence.Now,industrial-level learning-to-rank algorithms require a large amount of labeled data as input for training ranking model.However,the cost of obtaining labels is very expensive,so implicit feedback such as user click records are often used instead.However,implicit feedback contains internal bias,which will affect the user's click behavior,and the biased data will lead to poor performance of the ranking model.Therefore,it is necessary to study the unbiased learning-to-rank algorithm for e-commerce recommendation scenarios.In response to the above problem,firstly,this article analyzes the user's behavior in Ecommerce recommendation scenario,and explores all the biases that may affect the user's click in the scenario.It is found that E-commerce recommendation scenario includes a new type of bias in addition to the common position bias.,This kind of bias means that the context around the target product will affect the user's click.This article names this bias as context bias;this article designs statistical experiments and independence test experiments to prove the existence of biases and the independence between the two biases.Secondly,to solve the problems of position bias and context bias in e-commerce recommendation scenarios,this paper proposes a new unbiased learning to rank based on Combinational Propensity to obtain an unbiased ranking model that eliminates the effects of two biases.,The framework consists of three modules: calculation module of biases' propensity score,bias combination module and training module;in the calculation module of biases' propensity score,this paper proposes a new type of context-aware position bias click model and expectation maximization algorithm for propensity score calculation,the former is used to model the propensity score representation of position bias and context bias,and the latter is used to calculate the values of propensity score;in the bias combination module,this paper first proposes the concept of combinational bias and designs the method to calculate combinational bias.The method is used to integrate the influence of multiple biases;the training module adopts the existing unbiased learning to rank algorithm based on inverse propensity weighting to learn the ranking model and remove the influence of bias during training the model.This paper proves through experiments that both positon bias and context bias exist in the e-commerce recommendation scenario,and the Chi-square test proves that positon bias and context bias are independent of each other.Based on the ranking-support vector machine algorithm,this paper constructs an unbiased learning to rank model for e-commerce recommendation scenarios;in the experiment compared with the industrial baseline,ULTR-CP model can achieve better results than the industrial baseline under three experimental metrics.In the comparative experiment to the reranked results of removing different bias models,it is proved that the position bias has a more obvious impact on user clicks under low browsing depth,and the context bias has more obvious impact on user clicks under high browsing depth;comparative experiments on changing model parameters,ULTR-CP model is proved to be robust.