SUBMISSION: 1870 TITLE: Coarse-to-Fine Sparse Sequential Recommendation ------------------------- METAREVIEW ------------------------ Important and relevant research area, a sound method, and thorough experimentation. Some modelling assumptions should be better explained and/or discussed. The paper should briefly elaborate how their coarse-to-fine approach leads to a different model than other coarse-to-fine methods. ----------------------- REVIEW 1 --------------------- SUBMISSION: 1870 TITLE: Coarse-to-Fine Sparse Sequential Recommendation AUTHORS: Jiacheng Li, Tong Zhao, Jin Li, Jim Chan, Christos Faloutsos, George Karypis, Soo-Min Pantel and Julian McAuley ----------- Relevance to SIGIR ----------- SCORE: 4 (good) ----------- Technical soundness ----------- SCORE: 4 (good) ----------- Quality of presentation ----------- SCORE: 4 (good) ----------- Adequacy of citations ----------- SCORE: 4 (good) ----------- Reproducibility of methods ----------- SCORE: 4 (good) ----------- Strengths ----------- 1. The research problem itself is important. 2. The motivation is reasonable 3. The proposed method makes sense. 4. Experiments are roughly conducted ----------- Weaknesses ----------- 1. couldn't see its performance when applied to real traffic. ----------- Overall recommendation ----------- SCORE: 2 (accept) ----------- Detailed comments to authors ----------- In this paper, the authors tried to address a problem in the previous sequential recommendation problem. Specifically, the previous methods would suffer from the data sparsity issue, especially for the items that rarely appear in the training. To solve this problem, the authors proposed a method CAFE which first learns the high level coarse grained intent sequences, which would help incorporate dense intent data, and then fuse intent representations into item encoder outputs. The way of leveraging high level coarse intent data to address the data sparsity issue makes sense. The method that the authors proposed seems to be reasonable and could incorporate the information from both dense and sparse data together. The authors conduct experiments on two public datasets, Amazon and Tmall. The comparison results against others such as item only methods and intent aware methods demonstrated the advantages of the proposed method and the results seems to be solid. The only issue is that it is hard to tell what the impact would be when it got applied to real traffic instead of the toy situation here. But this is a minor concern for me for now. ----------- Nominate for Best Paper ----------- SELECTION: no ----------------------- REVIEW 2 --------------------- SUBMISSION: 1870 TITLE: Coarse-to-Fine Sparse Sequential Recommendation AUTHORS: Jiacheng Li, Tong Zhao, Jin Li, Jim Chan, Christos Faloutsos, George Karypis, Soo-Min Pantel and Julian McAuley ----------- Relevance to SIGIR ----------- SCORE: 4 (good) ----------- Technical soundness ----------- SCORE: 4 (good) ----------- Quality of presentation ----------- SCORE: 5 (excellent) ----------- Adequacy of citations ----------- SCORE: 4 (good) ----------- Reproducibility of methods ----------- SCORE: 3 (fair) ----------- Strengths ----------- 1. Clarity of writing and presentation. 2. Motivation behind the approach is clearly presented and backed by experiments. 3. The novel contribution is clearly described in terms of intuition and mathematical expressions. ----------- Weaknesses ----------- 1. The "trained embedding" part of the motivating experiment was hard to follow from section 2.3.2. However, its significance becomes clear later in the method section. 2. Ablation study is carried out only on the Tmall dataset. Does the finding correlate with the Amazon dataset? ----------- Overall recommendation ----------- SCORE: 2 (accept) ----------- Detailed comments to authors ----------- 1. The authors propose an enhancement to self-attentive methods for the sequential recommendation task. 2. The shortcomings of the existing approach are described and motivate the modification proposed by the authors. 3. I had a hard time following the second motivating experiment described in section 2.3.2. It was not clear to me what the authors wanted to observe here and why it motivates the proposed change from the text. However, after going through the method section, it became clear to me that the embeddings of infrequent items will be improved and after that figure 2.d will look more like 2.c. Maybe it could be more clear in section 2.3.2. 4. The mathematical expressions of the existing approach and the modifications are sound. It is accompanied by justifications and intuition which really helps the reader understand the proposed method. I think the paper addresses an interesting and relevant problem for the community. Overall their approach is well justified and supported by empirical results. I recommend accepting the paper for SIGIR. ----------- Nominate for Best Paper ----------- SELECTION: no ----------------------- REVIEW 3 --------------------- SUBMISSION: 1870 TITLE: Coarse-to-Fine Sparse Sequential Recommendation AUTHORS: Jiacheng Li, Tong Zhao, Jin Li, Jim Chan, Christos Faloutsos, George Karypis, Soo-Min Pantel and Julian McAuley ----------- Relevance to SIGIR ----------- SCORE: 4 (good) ----------- Technical soundness ----------- SCORE: 4 (good) ----------- Quality of presentation ----------- SCORE: 4 (good) ----------- Adequacy of citations ----------- SCORE: 4 (good) ----------- Reproducibility of methods ----------- SCORE: 4 (good) ----------- Strengths ----------- 1.This paper has a clear idea which explicitly considers user intent, and the performance is obviously improved. 2.This paper uses sufficient illustrations to enhance the persuasion and help readers to understand. ----------- Weaknesses ----------- 1.The paper showed the limitations of sparse data on model learning, but lacked a intuitive explanation that "considering user intents can alleviate the problem of data sparsity", which caused some confusion for me. 2.Some more well-known sequential recommendation models can be considered in experiments, not just backbone model-based comparisons. ----------- Overall recommendation ----------- SCORE: 1 (weak accept) ----------- Detailed comments to authors ----------- This paper proposes an approach to learn high-quality item representation from sparse data by modeling intent sequences and item sequences simultaneously. The highlight is to consider user intent explicitly. And the comparative experiments in this paper prove the effectiveness of the model. Especially, the analyses emphasize the advantages of the model in sparse data, which is also the focus of this paper. It would be better if several different better baselines (such as STAMP, NARM...) could be added. ----------- Nominate for Best Paper ----------- SELECTION: no