=============================
Reviewer Comments
======================================
* Originality: 4
* TechnicalQuality: 4
* Demonstrability: 4
* Confidence: 4
* Overall: 4
------------------------
* Comments: 1) Is the paper accessible to the general AI audience?

Yes. But the paper can do a better job in explaining the evaluation results to make it more accessible.

2) Is the paper readable and clear? 

Yes. The paper is well structured and easy to read.

3) Comments on the technical content of the paper?

Please refer to “Constructive comments to authors”.

4) Is there anything wrong with the paper? 

The paper is technically sound.

5) Should the paper be accepted to the track (please make an accept/reject recommendation)?

Accept

6) Most important: Add constructive comments to the authors of how to improve their paper.

“Scalability” argument. Because authors put the word “Scalability” in the title, there should be more detailed and systematic discussions in the paper in terms of why the proposed Trans. model is scalable. For example, can you show the recommendation performance/delay wrt. the increasing number of users/items? Otherwise, it’s better to just use the title “Translation-based recommendation”.

Related work and baselines. The paper mainly compares the proposed model to MF and MC based methods. However, recently we have seen papers leveraging RNN to conduct sequential preference modeling, such as “Recurrent Recommender Networks” in WSDM-2017. It may be unfair to ask authors to compare to the deep learning based approaches but it’s worth discussing in the related work section, and (even better) discuss the advantages of Translation-based recommendation.

Evaluation results. Evaluating recommendation systems is very different from evaluating other machine learning applications. It will be very helpful for non-Recsys people to understand the contributions if the authors can explain how to interpret the improvements, e.g., how much is considered significant, and the improvement on AUC is smaller than Hit or Recall, etc.

Qualitative examples. To help audience to get a direct sense of how the translation model works, it will be very helpful to add one or two qualitative examples from a dataset used in the paper. For example, showing how the user vector is guiding the transition of item vectors and as a result, accurately predicting the next consuming item.