Posts

In summary, we proposed a Popularity-bias Deconfounding framework based on backdoor adjustment to deconfound popularity bias for better recommendation. The new paradigm was demonstrated on the latent factor model and empirical studies on three real-world datasets.

In the original article 1, the authors focus on the method for identification and estimation, while do not put much effort in discussing the assumptions. There is no sensitivity analysis part in the article.

Estimation method Last time, we have identified the followiing causal effect. $$ \mathbb{E}\left[ C(i)\middle| U=u \right]=\mathbb{E}\left\{\mathbb{E}\left[ C\middle|I=i,U=u,Z \right]\right\}. $$ To estimate the statistical estimand on the RHS, we first estimate $\mathbb{E}\left[ C\middle|I=i,U=u,Z \right]$.

In a recent work 1, Zhang et al. (2021) characterized the item popularity as a confounder, and then developed a method, Popularity-bias Deconfounding(PD), to eliminate popularity bias in recommendation. The following causal graph shows the assumptions they made on how the training data is generated.

In the era of the information overwhelming, recommender systems play a crucial role in information filtering and further affect people’s cognition and decision-making. Existing recommender systems, like behavioral correlation in collaborative filtering, are mostly based on learning the correlation in data.

My name is Qilin Li. I am a second-year PhD student in Statitstics. With great interest in the intersection of statistics and computer science, I have been exploring the areas across high-dimensional statistics, machine learning, optimization and network analysis.