Reviews: Deconvolving Feedback Loops in Recommender Systems

First of all, the problem considered in this paper is interesting and useful to some potential applications that require the true rating matrix not influenced by any recommender systems. However, the inference of the true rating matrix from the observed one is an ill-posed problem, which need the "strong" (somewhat unrealistic) assumptions. The questions about the assumptions are summarized as follows: 1. Assumption 1 is quite restricted in the sense that the popular recommendation algorithms (e.x. Bayesian matrix factorization) cannot be properly expressed in Eq. (2). If the real-world RS makes use of a complex recommendation algorithm that is not covered by the Assumption 1, it is hard to validate the quality of the true rating matrix extracted by the proposed algorithm. I think that this is also non-realistic situation.