Reviewer #2
Questions
1. Summary and contributions: Briefly summarize the paper and its contributions.
The paper presents to use linear recurrent units to do sequential recommendation. The linear unit allows the model to be trained in parallel (in terms of sequence of items). The experimental results show the method improves the recommendation performance.
2. Strong points. List three to five strong points about the paper. Please be precise and explicit. Clearly explain the value and nature of the contribution.
The paper is well written in general, while there are multiple sections are unclear to me about the motivation.
The proposed work shows better experimental performance than existing works.
3. Weak points. List three to five weak points about the paper. Please clearly indicate whether the paper has any mistakes, missing related work, or results that cannot be considered as a contribution. Please be polite, specific, and constructive.
Many definitions/math description in this paper is sloppy as there are multiple errors and mismatchings in the paper. E.g. \bar{C} = CP^{-1}. How can you do P^{-1}\Lambda P^{-1}? E.g. \bar{h} = P^{-1}h, then \bar{h} = \Lambda... in equation 4.

The paper borrows multiple equations and concepts from other papers ,Resurrecting Recurrent Neural Networks for Long Sequences, without proper acknowledgement. Putting other works into the methodology section without proper citation and clear statement are not good.
[1] Efficiently Modeling Long Sequences with Structured State Spaces
[2] Resurrecting Recurrent Neural Networks for Long Sequences

Multiple settings of this work are not clearly stated. E.g. "we doubled the size of H", why? E.g. "Representation in complex space", why?
4. Detailed Evaluation. Please provide detailed feedback about the strengths and the weaknesses of the paper and support your overall rating. You may talk about significance, technical depth, novelty, reproducibility, relevance to the community, and potential ethical considerations. Please be polite, specific, and constructive.
The paper is fairly well written and interesting in terms of methodology. I liked this paper when i read it for the first time. However, later, I notice the majority of the algorithm description has been described in recent DeepMind paper that described the linear recurrent units. I start concerning its novelty if it is simply borrowing idea from other works and applied it to recsys. Indeed, borrowing idea from literature is fine and common, but presenting work without properly stating that the algorithms is from other work is not quite good.
5. Overall rating.
Weak Reject (Probable reject, unless convinced otherwise)
8. Reviewer confidence.
Expert: I have published papers on this topic.
Reviewer #3
Questions
1. Summary and contributions: Briefly summarize the paper and its contributions.
The authors propose a novel sequential recommender called LRURec. The proposed model introduces: (1) linear recurrence with matrix diagonalization to efficiently capture user transition patterns; and (2) a recursive parallelization framework that significantly accelerates training and inference. Moreover, LRURec is designed with a series of improvements for optimal recommendation performance and inference efficiency.
2. Strong points. List three to five strong points about the paper. Please be precise and explicit. Clearly explain the value and nature of the contribution.
+ Novelty
+ Performances of the model
+ Technical quality
3. Weak points. List three to five weak points about the paper. Please clearly indicate whether the paper has any mistakes, missing related work, or results that cannot be considered as a contribution. Please be polite, specific, and constructive.
- Replicability
- Statistical validation
- Resource Consumption Analysis Missing
4. Detailed Evaluation. Please provide detailed feedback about the strengths and the weaknesses of the paper and support your overall rating. You may talk about significance, technical depth, novelty, reproducibility, relevance to the community, and potential ethical considerations. Please be polite, specific, and constructive.
The authors introduce fresh ideas and approaches to face the problem of efficiency of Session-Based RSs, contributing to the advancement of knowledge. The paper proposes an innovative approach, providing a unique perspective that distinguishes it from existing research. This novelty adds significant value to the work and makes it relevant and exciting for the conference audience.

The performance of the model showcased in the paper is encouraging. The authors present compelling evidence of the model's effectiveness, achieving reasonably good results on benchmark datasets. The performance metrics and evaluation methodologies employed are appropriate and well-described.

In terms of technical quality, the paper exhibits a strong foundation in terms of methodology, experimental design, and implementation. The authors demonstrate a deep understanding of the underlying AI techniques and algorithms employed in their research. The experimental setup is well described. The paper also presents a thorough analysis of the experimental findings, supported by sound reasoning and logical coherence.

However, there are a few weaknesses that should be addressed. Firstly, the replicability of the experiments could be improved. The authors should provide more detailed information regarding the datasets, model architecture, hyperparameters, and any other necessary resources to allow other researchers to replicate their experiments. It could be nice if they decide to release source code, datasets and any other resource for ensure replicability. This would strengthen the confidence in the reported results.
Additionally, the paper lacks sufficient statistical validation. While the model's performance is presented, statistical tests or confidence intervals should be utilized to assess the significance of the reported results. Furthermore, the authors should include an analysis of resource consumption, such as computational requirements, memory usage, or energy consumption, associated with the proposed model. This analysis would provide valuable insights into the practical feasibility and scalability of the approach, allowing for a more comprehensive evaluation of its potential impact.
5. Overall rating.
Weak Accept (Probable accept, unless convinced otherwise)
8. Reviewer confidence.
Familiar: I have passing knowledge on this topic.
Reviewer #4
Questions
1. Summary and contributions: Briefly summarize the paper and its contributions.
This paper proposes a new sequential recommendation model called Linear Recurrent Units for Sequential Recommendation (LRURec). The key ideas are:

Use linear recurrent units (LRUs) to model user sequences. LRUs are simplified recurrent units that use matrix diagonalization and exponential parameterization for efficient computation.
Introduce recursive parallelization to accelerate training and inference by splitting sequences into smaller subsequences for parallel processing.
Design improvements like layer normalization, residual connections, and position-wise feedforward networks to compensate for the lack of non-linearity in LRUs.
Conduct experiments on real-world datasets, showing LRURec outperforms state-of-the-art methods like GRU4Rec and SASRec in terms of accuracy, training speed, and inference efficiency.
2. Strong points. List three to five strong points about the paper. Please be precise and explicit. Clearly explain the value and nature of the contribution.
LRURec achieves recommendation improvement over related baselines across multiple datasets.
The linear design enables highly efficient parallelized training. Experiments show much faster convergence.
Recursive parallelization reduces inference complexity to O(logL) for sequence length L, supporting real-time recommendation.
3. Weak points. List three to five weak points about the paper. Please clearly indicate whether the paper has any mistakes, missing related work, or results that cannot be considered as a contribution. Please be polite, specific, and constructive.
The paper does not sufficiently compare against some recent related work in efficient sequence modeling, such as linear transformer variants and Retentive Networks.
The paper focuses heavily on accuracy and efficiency results. More in-depth analysis on the linear modeling approach itself could provide useful insights into why and how LRURec achieves the reported gains.
There seems to be limited analysis around the impact of sequence ordering and potential noise. Additional experiments testing the robustness of LRURec could reveal useful insights.
4. Detailed Evaluation. Please provide detailed feedback about the strengths and the weaknesses of the paper and support your overall rating. You may talk about significance, technical depth, novelty, reproducibility, relevance to the community, and potential ethical considerations. Please be polite, specific, and constructive.
Conduct experiments on larger and more complex datasets. The current results are mostly limited to relatively small recommendation datasets. Testing on larger corpora with more complex transitions and longer sequences could reveal limitations.
Evaluate the interpretability of LRURec compared to simpler sequential models through inspection of learnt parameters. This could reveal insights into the recommendation logic.
Provide theoretical analysis of the proposed recursive parallelization algorithm. Deriving time and space complexity bounds can quantify efficiency gains.
Explore different matrix decomposition techniques beyond eigendecomposition to further improve stability and efficiency.
Test the robustness of LRURec to changes in sequence ordering and noise through additional experiments. This can analyze the sensitivity of the approach.
Release source code and pretrained models to facilitate reproducibility and enable applications by others.
5. Overall rating.
Weak Reject (Probable reject, unless convinced otherwise)
8. Reviewer confidence.
Knowledgeable: I have read papers on this topic.
Reviewer #5
Questions
1. Summary and contributions: Briefly summarize the paper and its contributions.
In the paper, the linear recurrent unit idea of [23] is deployed in a SAR-type neural recommender to improve efficiency without sacrificing recommendation quality.

While the application of the linear recurrent unit from [23] is relative straightforward, the idea to use for recommendation and the adaptation is novel and the proposed algorithm performs well.

The reproducibility of the result must be improved by releasing the source code and using standardized recommender evalutation frameworks.
2. Strong points. List three to five strong points about the paper. Please be precise and explicit. Clearly explain the value and nature of the contribution.
1 Novel idea of using linear recurrent units for recommendation
2 Works well in experiments
3. Weak points. List three to five weak points about the paper. Please clearly indicate whether the paper has any mistakes, missing related work, or results that cannot be considered as a contribution. Please be polite, specific, and constructive.
1 The technique to adapt linear recurrent units in a SAR-type recommender is relative straightforward
2 Limited reproducibility. No sorce code is provided, no evaluation framework is used.
4. Detailed Evaluation. Please provide detailed feedback about the strengths and the weaknesses of the paper and support your overall rating. You may talk about significance, technical depth, novelty, reproducibility, relevance to the community, and potential ethical considerations. Please be polite, specific, and constructive.
Please release your source code, even for anonymous review there is the option https://anonymous.4open.science/. Please use an evaluation framework such as RecBole or https://github.com/MaurizioFD/RecSys2019_DeepLearning_Evaluation
5. Overall rating.
Weak Accept (Probable accept, unless convinced otherwise)
8. Reviewer confidence.
Knowledgeable: I have read papers on this topic.