------------------------- METAREVIEW ------------------------ RECOMMENDATION: no recommendation TEXT: This paper presents an intent contrastive learning model for sequential recommendation. All the reviewers think the research question is important and the proposed method is interesting. However, they also have some concerns. 1) User level clustering is expensive, and efficiency may be a problem. 2) Cluster number selection is not clearly explained. 3) The paper writing can be improved. 4) How to combine contrastive learning with the recommendation model is not well explained. 5) The idea of using contrastive learning for training user representation is not novel, and there are some existing works. 6) The high complexity of the proposed method may hinder its application on large dataset and industrial scenarios. 7) Whether modeling intent as latent clustering of sequences can actually capture user intent is questionable. I hope the authors can refine their work based on the comments from the reviewers. ----------------------- REVIEW 1 --------------------- SUBMISSION: 1167 TITLE: Intent Contrastive Learning for Sequential Recommendation AUTHORS: Yongjun Chen, Zhiwei Liu, Jia Li, Julian McAuley and Caiming Xiong ----------- Summary ----------- The authors proposed an intent contrastive learning model for sequential recommendation. The idea to contrast user intent with its behavior sequence to denoise the data is very interesting. The authors also did very clear illustrations and analyses on model performance against state-of-art baselines, model efficiency, robustness, ablation study and hyperparameter impact are all covered. Easy to understand and follow. ----------- Originality ----------- SCORE: 1 (Creative: An original research question and/or approach that distringuishes the work from the state of the art.) ----------- Impact ----------- SCORE: 1 (Broad: Could help ongoing research in a broader research community.) ----------- Technical quality ----------- SCORE: 2 (Excellent: The approach is well-justified and all the claims are convincingly supported.) ----------- Related work quality ----------- SCORE: 1 (Comprehensive: Cannot think of any important paper that is missed.) ----------- Presentation quality ----------- SCORE: 1 (Lucid: Very well written in every aspect, a pleasure to read, easy to follow.) ----------- Strengths ----------- The authors proposed to contrast latent intention with sequence recommendation is an interesting idea that has not been explored in the past. Mostly contrastive loss are used on the augmented sequence views only. It is insightful to see borrowing user similarities can improve the sequential recommendation quality. The approach to use EM for estimation is also a good connection made with traditional modeling approaches in the neural model era, give some guarantee for convergence. The writing of the paper is very easy to follow and precise. The performance illustrations give necessary details and conclusive learnings. It can convey solid insights to the audience. ----------- Weaknesses ----------- Not a weakness in the paper. But good future work to improve upon. The user level clustering is a little expensive during training. The time complexity O(|U|mkd) is pretty high when hundreds of millions of users are likely at the industrial level. Would love to see if there are good ways to make the learning more efficient. Some minor typos in the paper, good to proofread again. 4.1.2 “aggregation layer” to denote the the -- dup Equation (14), underscript of sum should be u = 1 not v = 1 Figure 9 is the same as Figure 3. ----------- Relevance to the Web Conference ----------- SCORE: 1 (In scope) ----------- Ethics ----------- no concern ----------- Overall evaluation ----------- SCORE: 2 (Definitely accept) ----------- Reviewer's confidence ----------- SCORE: 4 ((high)) ----------------------- REVIEW 2 --------------------- SUBMISSION: 1167 TITLE: Intent Contrastive Learning for Sequential Recommendation AUTHORS: Yongjun Chen, Zhiwei Liu, Jia Li, Julian McAuley and Caiming Xiong ----------- Summary ----------- This paper studies the sequential recommender systems by proposing intent contrastive learning. It proposes to learn users’ intent distribution functions from unlabeled user behavior sequences, then optimizes sequential recommendation models with contrastive self-supervised learning by considering the learned intents. Overall, the studied problem is interesting and the proposed method is promising. ----------- Originality ----------- SCORE: 1 (Creative: An original research question and/or approach that distringuishes the work from the state of the art.) ----------- Impact ----------- SCORE: 1 (Broad: Could help ongoing research in a broader research community.) ----------- Technical quality ----------- SCORE: -1 (Poor: Potentially reasonable approach, but certain core claims lack justification.) ----------- Related work quality ----------- SCORE: 0 (Reasonable: Coverage of past work is acceptable, but a few papers are missing.) ----------- Presentation quality ----------- SCORE: 0 (Reasonable: Understandable to a large extent, but parts of the paper need more work.) ----------- Strengths ----------- This paper studies an interesting problem. The proposed method is promising. The experiments are extensive. ----------- Weaknesses ----------- How contrastive learning is utilized in the recommendation model is not well explained. And using intent for recommendation is not a novel idea. The authors need to proofread the paper thoroughly. The full text should be provided when first mentioning an abbreviation, e.g., SSL in the abstract. How to select the cluster number in K-means seems not clear. ----------- Relevance to the Web Conference ----------- SCORE: 1 (In scope) ----------- Ethics ----------- N/A ----------- Overall evaluation ----------- SCORE: 1 (Accept) ----------- Reviewer's confidence ----------- SCORE: 3 ((medium)) ----------------------- REVIEW 3 --------------------- SUBMISSION: 1167 TITLE: Intent Contrastive Learning for Sequential Recommendation AUTHORS: Yongjun Chen, Zhiwei Liu, Jia Li, Julian McAuley and Caiming Xiong ----------- Summary ----------- this paper proposed a new model to understand latent user intent in the form of sequence patterns and leverage it for sequential recommendation. ----------- Originality ----------- SCORE: 1 (Creative: An original research question and/or approach that distringuishes the work from the state of the art.) ----------- Impact ----------- SCORE: 1 (Broad: Could help ongoing research in a broader research community.) ----------- Technical quality ----------- SCORE: 1 (Good: Generally solid work, and any claims that might need adjustment are unlikely to impact the central contributions of the paper.) ----------- Related work quality ----------- SCORE: 0 (Reasonable: Coverage of past work is acceptable, but a few papers are missing.) ----------- Presentation quality ----------- SCORE: 0 (Reasonable: Understandable to a large extent, but parts of the paper need more work.) ----------- Strengths ----------- - understanding user intent and sequential recommendation is an important research topic with real-world applications. - this paper proposed a new approach by jointly learning latent user intent (by clustering sequence patterns) and using it for next item recommendation. The proposed approach is interesting and novel. - the experiments compared with multiple SOTA methods on sequential recommendation using public datasets; and showed the proposed method significantly outperformed previous methods ----------- Weaknesses ----------- - the proposed method is very complex and expensive, so it may not be able to scale on large datasets in real world industry settings to make this work applicable. - using contrastive learning for training user representation is not novel (CL4Rec) - it is unclear if modeling intent as latent clustering of sequences can actually capture user intent, it would be good to show some case studies / visualization to understand that exactly the model learns. ----------- Relevance to the Web Conference ----------- SCORE: 1 (In scope) ----------- Ethics ----------- N/A ----------- Overall evaluation ----------- SCORE: 1 (Accept) ----------- Reviewer's confidence ----------- SCORE: 4 ((high))