【Debias】Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in RS（KDD‘21）

Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System (KDD’21)

chart a It is the assumption of our general recommendation model , That is, the matching degree between users and items can reflect whether there will be interaction . But in fact, it should be as shown in the figure c Shown , Interaction is not only related to the degree of matching , It is also related to the deviation between the user and the item itself ： Popularity of items 、 Whether users like popular items . Therefore, the recommended model should be modified to ：

among user module and item module It's a mlp, The output of the recommended model should be ：
$${\hat{y}}_{ui}={\hat{y}}_k\ast \sigma \left({\hat{y}}_i\right)\ast \sigma \left({\hat{y}}_u\right)$$
The loss function is designed as ：
$$\begin{array}{rl}{L}_O& =\sum _{(u,i)\in D}-y_{ui}\log \left(\sigma \left({\hat{y}}_{ui}\right)\right)-\left(1-y_{ui}\right)\log \left(1-\sigma \left({\hat{y}}_{ui}\right)\right)\\ L_U& =\sum _{(u,i)\in D}-y_{ui}\log \left(\sigma \left({\hat{y}}_u\right)\right)-\left(1-y_{ui}\right)\log \left(1-\sigma \left({\hat{y}}_u\right)\right)\\ L_I& =\sum _{(u,i)\in D}-y_{ui}\log \left(\sigma \left({\hat{y}}_i\right)\right)-\left(1-y_{ui}\right)\log \left(1-\sigma \left({\hat{y}}_i\right)\right)\\ L& =L_O+\alpha \ast L_I+\beta \ast L_U\end{array}$$

So in order to eliminate the influence of users and items , The unbiased prediction output should be ：
$${\hat{y}}_k\ast \sigma \left({\hat{y}}_i\right)\ast \sigma \left({\hat{y}}_u\right)-c\ast \sigma \left({\hat{y}}_i\right)\ast \sigma \left({\hat{y}}_u\right)$$