Collaborative filtering has been widely used in recommender systems. Existing work has primarily focused on improving the prediction accuracy mainly via either building refined models or incorporating additional side information, yet has largely ignored the inherent distribution of the input rating data. In this paper, we propose a data debugging framework to identify overly personalized ratings whose existence degrades the performance of a given collaborative filtering model. The key idea of the proposed approach is to search for a small set of ratings whose editing (e.g., modification or deletion) would near-optimally improve the recommendation accuracy of a validation set. Experimental results demonstrate that the proposed approach can significantly improve the recommendation accuracy. Furthermore, we observe that the identified ratings significantly deviate from the average ratings of the corresponding items, and the proposed approach tends to modify them towards the average.

We then search them from {0.01, 0.05, 0.1, 0.5, 1} via cross-validation. We will show the experimental results later for different values of fold number. For the results reported in the paper, we set it to 4. An additional In this document, we will also provide sensitivity experiments about this parameter. We then output a ranked list of each user based the predicted rating. The HR and nDCG are defined as follows.

The proposed algorithm is limited to matrix factorization model, and can be hardly extended to more state-of-art neural network-based latent factor models proposed in recent years. Because the derivate of model parameters with respect to training labels in Equation (7) needs to be a closed form solution as in matrix factorization. This may restrict a broader impact of the proposed solution. I'm concerned about the authors' claim on the trade-off between personalization and accuracy. As emphasized in the title, the authors consider the performance gain of the proposed algorithm as trading personalization for accuracy, but there is no direct empirical evaluation evidence to support this claim.

The paper received overall very positive scores (after communication through author response). All the reviewers agree that the paper made a very interesting contribution from a novel angle to understand the tradeoff between "over-personalization" and accuracy. The empirical results provide convincing support for the claim. I suggest the authors incorporate the feedback from the reviewers in the revision.

Collaborative filtering has been widely used in recommender systems. Existing work has primarily focused on improving the prediction accuracy mainly via either building refined models or incorporating additional side information, yet has largely ignored the inherent distribution of the input rating data. In this paper, we propose a data debugging framework to identify overly personalized ratings whose existence degrades the performance of a given collaborative filtering model. The key idea of the proposed approach is to search for a small set of ratings whose editing (e.g., modification or deletion) would near-optimally improve the recommendation accuracy of a validation set. Experimental results demonstrate that the proposed approach can significantly improve the recommendation accuracy.