Breaking Determinism: Fuzzy Modeling of Sequential Recommendation Using Discrete State Space Diffusion Model

Summary

The paper presents the DDSR model for Sequential Recommendation (SR), which better captures user interest evolution by using fuzzy sets of interaction sequences. Unlike traditional methods, DDSR effectively handles the unpredictability of user behavior and addresses cold start issues. Experiments on benchmark datasets show that DDSR outperforms existing methods, demonstrating its effectiveness in SR tasks.

Strengths

1. DDSR introduces a novel perspective in SR, effectively addressing the limitations of traditional sequential modeling methods and enhancing recommendation accuracy. 2. DDSR uses fuzzy sets of interaction sequences and diffusion transition processes in discrete state spaces, improving the model's ability to capture the randomness and unpredictability of user behavior. 3. The theoretical justification for constructing the information diffusion approximation model is sound and fundamental. 4. Quantitative experimental results demonstrate the model's effectiveness, with particularly exciting results in cold start scenarios.

Weaknesses

1. The paper lacks an algorithm to describe the entire model operation. Including this would greatly enhance readers' understanding of the model and its theoretical underpinnings. 2. All experimental results are quantitative. It is recommended to supplement with a case study or visual experiment. 3. Although the author has analyzed the time complexity, I believe most readers would also appreciate a comparison of actual running times.

Questions

See comments above.

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

This paper has discussed limitations.

Summary

This paper studies the problem of fuzzy modeling for sequential recommendation. The work proposes to leverage the fuzzy sets of interaction sequences for modeling the nature of users' evolving real interests. It also introduces the discrete diffusion modeling specifically for the discrete data. The experiment results showcase the better model performance in comparison with several included baselines.

Strengths

1. This paper studies the optimization of sequential recommender systems, which is an important application for the information retrieval domain. 2. The paper is written with fair clarity and thus it is easy to follow. 3. The paper provides careful analyses to support the claims made by this paper and the experimental results generally demonstrate the effectiveness of the proposed methods.

Weaknesses

1. Insufficient baseline comparison for the conventional sequential recommender models, as the two included models are outdated. 2. Insufficient empirical analysis in terms of the model efficiency. Firstly, the runtime cost comparison is missing, which is however very important to recommender system prototype evaluation for practical deployment consideration. Secondly, based on the complexity analysis, it is crucial to showcase the performance variations (effectiveness/efficiency) by varying the codebook length $m$.

Questions

1. How many times of performance evaluations are conducted for the average results? 2. What are the real runtime costs for the proposed model and other baselines for comparison?

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

Please see the weakness for details.

Summary

The paper presents a diffusion model-based sequential recommendation from a novel information-theoretical perspective, which operates on discrete structural state spaces along with semantic labels improving efficiency and tackling cold-start issues.

Strengths

Strengths* 1. Novelty. The paper uses a directed graph to model sequential recommendation and models the transitions of nodes with discrete diffusion. Besides, authors introduce semantic tags to replace meaningless item IDs, enhancing efficiency and improving cold start issues. 2. Soundness. The paper presents enough theoretic analysis and experimental evaluations to validate the proposed method. The significant performance improvement on three public datasets, along with a series of further ablation study and analysis, clearly demonstrates the soundness of the paper. 3. Easy-followed. The paper presents sufficient algorithmic detail. In addition, the authors provide sufficient codes to help understand and re-product the algorithm. Weaknesses* 1. Citation mixed in the text make obstacles for reading. See questions for detail. 2. Writing should be more regulated. See questions for detail.

Weaknesses

See Above

Questions

1. In text such as line 235, 236, 251, the citation text is literally repeated. 2. In text such as line 20-26, it would be better to add brackets to distinguish references from the main text. 3. In line 74 and 144, it should be explicit that “2” and “4.1” refers to “Section 2” and “Section 4.1”. 4. The usage and format of quotation marks is a bit confusing, see line 34, 219 and Table 1 caption for examples.

Rating

8

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

There is no significant limitation. Perhaps the authors could consider give a rough discussion or comment on how the computational complexity (or the coefficient constant) of the proposed method can be improved.

Summary

The paper presents a new model for sequential recommendation (SR) called DDSR, which aims to predict items of interest for users based on their past behavior. The authors critique conventional SR methods for not adequately capturing the randomness and unpredictability in user behavior. DDSR uses fuzzy sets of interaction sequences to better reflect the true evolution of user interests. Unlike common diffusion models that operate in continuous domains, DDSR uses diffusion transition processes in discrete state spaces, avoiding information loss by employing structured transitions. To tackle the inefficiency of matrix transformations in large discrete spaces, the model utilizes semantic labels derived from quantization or RQ-VAE instead of item IDs. Tests on three public datasets show that DDSR surpasses current state-of-the-art methods, proving its effectiveness in SR tasks.

Strengths

Effective combination of pre-existing ideas: discrete diffusion with semantic IDs. Empirical results in Section 5 on Scientific, Office, Online Retail datasets.

Weaknesses

I think a main weakness of the paper is a lack of clarity in writing. The method is still unclear to me, although this may be partly due to my lack of experience in recommender systems. I would condense and combine Sections 3 and 4 and move Section 2 to end. The technical novelty is also not entirely clear to me. Is it in the form of the transition matrices used for the discrete diffusion? Or is the technical novelty claimed that the items are recommended based on the output of running a few steps of discrete diffusion (adding noise to the features)?

Questions

Could the authors please elaborate on the analogy between fuzzy sets and discrete diffusion?

Rating

5

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

Yes, limitations adequately addressed.

To authors

You have addressed some of the concerns I raised in my previous review. I believe this work is indeed novel and meets the standards required for this conference. As a result, I am inclined to raise my score. Best regards,

Response to Rebuttal

I thank the authors for their thoughtful modifications to the paper - is it possible to upload this revision to OpenReview before the review period ends? I think the inclusion of the pseudo-code for training and sampling algorithms is a big step forward. Also, with the inclusion of the new experimental results I am glad to increase my score.