pith. sign in

arxiv: 1907.00590 · v1 · pith:TDCWKE6Cnew · submitted 2019-07-01 · 💻 cs.IR

A Review-Driven Neural Model for Sequential Recommendation

Chenliang Li , Xichuan Niu , Xiangyang Luo , Zhenzhong Chen , Cong Quan This is my paper

Pith reviewed 2026-05-25 11:49 UTC · model grok-4.3

classification 💻 cs.IR
keywords sequential recommendationreview-driven modelneural modelattention mechanismaspect-aware representationscollaborative filtering
0
0 comments X

The pith

Aspect-aware review representations plus hierarchical attention improve sequential item recommendations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that user and item representations built from aspects mentioned in reviews, when fed into a hierarchical attention-over-attention layer, can capture both long-term preferences and short-term purchase sequences for the next-item prediction task. A sympathetic reader would care because reviews already exist on e-commerce sites and contain semantic signals that pure rating-based collaborative filtering often misses. The claim is that this combination yields better accuracy than prior sequential models on real purchase data. If the claim holds, recommendation engines could make fuller use of existing review text to predict what a user will buy next.

Core claim

RNS encodes each user or item with aspect-aware representations extracted from the reviews. Given a sequence of historical purchased items for a user, a novel hierarchical attention over attention mechanism captures sequential patterns at both union-level and individual-level. Extensive experiments on three real-world datasets of different domains demonstrate that RNS obtains significant performance improvement over uptodate state-of-the-art sequential recommendation models.

What carries the argument

Aspect-aware representations extracted from reviews, processed by a hierarchical attention-over-attention mechanism that isolates union-level and individual-level sequential patterns.

If this is right

  • The model integrates long-term preference signals from reviews with short-term sequential behavior from purchase histories.
  • Review semantics supply complementary information to rating-based collaborative filtering.
  • The approach produces measurable accuracy gains across datasets from different product domains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If aspect extraction works reliably, recommendation systems could reduce dependence on explicit ratings.
  • The same multi-level attention structure might apply to other user sequence data such as clicks or searches.
  • Domain-specific review language could require separate tuning of the aspect extraction step for new product categories.

Load-bearing premise

Aspect-aware representations extracted from reviews reliably encode users' intrinsic long-term preferences rather than transient review text patterns.

What would settle it

Retrain the model after replacing all review text with random or permuted words while keeping the same architecture and check whether the performance advantage over baselines disappears.

Figures

Figures reproduced from arXiv: 1907.00590 by Chenliang Li, Cong Quan, Xiangyang Luo, Xichuan Niu, Zhenzhong Chen.

Figure 1
Figure 1. Figure 1: The network architecture of RNS. sequential recommendation with implicit feedback. Conse￾quently, we merge the set of reviews written by user u and the set of reviews written for item i to form user document Du and item document Di respectively. Given user u, her recently purchased L items and their corresponding review documents, the goal is to rank candidate items in terms of the likelihood that the user… view at source ↗
Figure 2
Figure 2. Figure 2: Parameter sensitivity of RNS: L and α Due to space limitation, we only list the results with respect to Recall and HR. Similar performance patterns are also ob￾served for Precision and NDCG. RNS-u means that the short￾term user preference is modeled by excluding union-level, i.e., p s u = p s2 u , and RNS-i for excluding individual level, i.e., p s u = p s1 u . RNS-pe means that position embeddings are ex￾… view at source ↗
read the original abstract

Writing review for a purchased item is a unique channel to express a user's opinion in E-Commerce. Recently, many deep learning based solutions have been proposed by exploiting user reviews for rating prediction. In contrast, there has been few attempt to enlist the semantic signals covered by user reviews for the task of collaborative filtering. In this paper, we propose a novel review-driven neural sequential recommendation model (named RNS) by considering users' intrinsic preference (long-term) and sequential patterns (short-term). In detail, RNS is devised to encode each user or item with the aspect-aware representations extracted from the reviews. Given a sequence of historical purchased items for a user, we devise a novel hierarchical attention over attention mechanism to capture sequential patterns at both union-level and individual-level. Extensive experiments on three real-world datasets of different domains demonstrate that RNS obtains significant performance improvement over uptodate state-of-the-art sequential recommendation models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes a review-driven neural sequential recommendation model (RNS) that encodes each user or item with aspect-aware representations extracted from reviews to capture long-term intrinsic preferences. For a sequence of historical items, it applies a novel hierarchical attention-over-attention mechanism to model short-term sequential patterns at both the union level and the individual level. The central claim is that extensive experiments on three real-world datasets from different domains demonstrate significant performance improvements over up-to-date state-of-the-art sequential recommendation models.

Significance. If the claimed gains prove robust under proper controls, the work could advance review-augmented collaborative filtering by showing how semantic signals in reviews can be combined with sequential modeling to improve recommendation accuracy beyond purely ID-based approaches.

major comments (2)
  1. [Abstract] Abstract: the claim that RNS 'obtains significant performance improvement' is unsupported by any quantitative results, ablation studies, statistical tests, dataset statistics, or error analysis, which is load-bearing for the central claim and prevents evaluation of whether post-hoc modeling choices drove the headline metric.
  2. [Abstract] Abstract: no equations, fitting procedures, or architectural details are supplied for the aspect-aware representations or the hierarchical attention-over-attention mechanism, so it is impossible to assess whether these components isolate union-level and individual-level patterns without overfitting to review text distributions.
minor comments (1)
  1. [Abstract] The sentence 'there has been few attempt' contains a grammatical error and should read 'there have been few attempts'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed comments. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that RNS 'obtains significant performance improvement' is unsupported by any quantitative results, ablation studies, statistical tests, dataset statistics, or error analysis, which is load-bearing for the central claim and prevents evaluation of whether post-hoc modeling choices drove the headline metric.

    Authors: We agree that the abstract would benefit from including concrete quantitative indicators to support the performance claim. The full manuscript reports relative improvements, ablation results, and significance tests across the three datasets in Section 5. We will revise the abstract to incorporate key performance gains (e.g., average improvements over the strongest baselines). revision: yes

  2. Referee: [Abstract] Abstract: no equations, fitting procedures, or architectural details are supplied for the aspect-aware representations or the hierarchical attention-over-attention mechanism, so it is impossible to assess whether these components isolate union-level and individual-level patterns without overfitting to review text distributions.

    Authors: Abstracts are intentionally concise and do not contain equations or implementation details; these are provided with full mathematical formulations in Sections 3.2 (aspect-aware encoding) and 3.3 (hierarchical attention-over-attention) of the manuscript. The abstract summarizes the modeling approach at a high level. revision: no

Circularity Check

0 steps flagged

No significant circularity

full rationale

The provided abstract and description contain no derivation chain, equations, or first-principles claims that reduce to inputs by construction. RNS is presented as an empirical neural architecture whose headline result is performance improvement on three datasets; this rests on experimental outcomes rather than any self-definitional mapping, fitted parameter renamed as prediction, or load-bearing self-citation. No steps meet the criteria for circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the unverified premise that review text yields stable aspect-aware vectors that capture intrinsic preference and that the attention mechanism isolates genuine sequential signals rather than dataset artifacts. No explicit free parameters, axioms, or invented entities are stated in the abstract.

pith-pipeline@v0.9.0 · 5689 in / 1145 out tokens · 26165 ms · 2026-05-25T11:49:14.182389+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 1 internal anchor

  1. [1]

    Mining sequential patterns

    [Agrawal and Srikant, 1995] Rakesh Agrawal and Ramakr- ishnan Srikant. Mining sequential patterns. In ICDE, pages 3–14,

    work page 1995

  2. [2]

    Top- icmf: Simultaneously exploiting ratings and reviews for recommendation

    [Bao et al., 2014] Yang Bao, Hui Fang, and Jie Zhang. Top- icmf: Simultaneously exploiting ratings and reviews for recommendation. In AAAI, pages 2–8,

    work page 2014

  3. [3]

    Blei, Andrew Y

    [Blei et al., 2003] David M. Blei, Andrew Y . Ng, and Michael I. Jordan. Latent dirichlet allocation. Journal of Machine Learning Research, 3:993–1022,

    work page 2003

  4. [4]

    [Catherine and Cohen, 2017] Rose Catherine and William W. Cohen. Transnets: Learning to trans- form for recommendation. In RecSys, pages 288–296,

    work page 2017

  5. [5]

    Sequential recommendation with user memory net- works

    [Chen et al., 2018] Xu Chen, Hongteng Xu, Yongfeng Zhang, Jiaxi Tang, Yixin Cao, Zheng Qin, and Hongyuan Zha. Sequential recommendation with user memory net- works. In WSDM, pages 108–116,

    work page 2018

  6. [6]

    Kankanhalli

    [Cheng et al., 2018] Zhiyong Cheng, Ying Ding, Lei Zhu, and Mohan S. Kankanhalli. Aspect-aware latent factor model: Rating prediction with ratings and reviews. In WWW, pages 639–648,

    work page 2018

  7. [7]

    Anr: Aspect-based neural recommender

    [Chin et al., 2018] Jin Yao Chin, Kaiqi Zhao, Shafiq Joty, and Gao Cong. Anr: Aspect-based neural recommender. In CIKM, pages 147–156,

    work page 2018

  8. [8]

    Fusing similarity models with markov chains for sparse sequential recommendation

    [He and McAuley, 2016] Ruining He and Julian McAuley. Fusing similarity models with markov chains for sparse sequential recommendation. In ICDM, pages 191–200,

    work page 2016

  9. [9]

    Trirank: Review-aware explainable recom- mendation by modeling aspects

    [He et al., 2015] Xiangnan He, Tao Chen, Min-Yen Kan, and Xiao Chen. Trirank: Review-aware explainable recom- mendation by modeling aspects. In CIKM, pages 1661– 1670,

    work page 2015

  10. [10]

    Session-based Recommendations with Recurrent Neural Networks

    [Hidasi et al., 2015] Bal´azs Hidasi, Alexandros Karat- zoglou, Linas Baltrunas, and Domonkos Tikk. Session- based recommendations with recurrent neural networks. CoRR, abs/1511.06939,

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  11. [11]

    Self-attentive sequential recommendation

    [Kang and McAuley, 2018] Wang-Cheng Kang and Julian McAuley. Self-attentive sequential recommendation. In ICDM, pages 197–206,

    work page 2018

  12. [12]

    Keiningham, Bruce Cooil, Lerzan Aksoy, Tor W

    [Keiningham et al., 2007] Timothy L. Keiningham, Bruce Cooil, Lerzan Aksoy, Tor W. Andreassen, and Jay Weiner. The value of different customer satisfaction and loyalty metrics in predicting customer retention, recommendation, and share-of-wallet. Managing Service Quality: An Inter- national Journal, 17(4):361–384,

    work page 2007

  13. [13]

    Convolutional neural networks for sentence classification

    [Kim, 2014] Yoon Kim. Convolutional neural networks for sentence classification. In EMNLP, pages 1746–1751,

    work page 2014

  14. [14]

    Bell, and Chris V olinsky

    [Koren et al., 2009] Yehuda Koren, Robert M. Bell, and Chris V olinsky. Matrix factorization techniques for rec- ommender systems. IEEE Computer, 42(8):30–37,

    work page 2009

  15. [15]

    A capsule network for recommendation and explaining what you like and dislike

    [Li et al., 2019] Chenliang Li, Cong Quan, Peng Li, Yunwei Qi, Yuming Deng, and Libing Wu. A capsule network for recommendation and explaining what you like and dislike. In SIGIR,

    work page 2019

  16. [16]

    A hybrid of sequential rules and collabo- rative filtering for product recommendation

    [Liu et al., 2009] Duen-Ren Liu, Chin-Hui Lai, and Wang- Jung Lee. A hybrid of sequential rules and collabo- rative filtering for product recommendation. Inf. Sci. , 179(20):3505–3519,

    work page 2009

  17. [17]

    Re- current neural network based language model

    [Mikolov et al., 2010] Tomas Mikolov, Martin Karafi ´at, Luk´as Burget, Jan Cernock´y, and Sanjeev Khudanpur. Re- current neural network based language model. In Proc. of INTERSPEECH, pages 1045–1048,

    work page 2010

  18. [18]

    Factorizing personal- ized markov chains for next-basket recommendation

    [Rendle et al., 2010] Steffen Rendle, Christoph Freuden- thaler, and Lars Schmidt-Thieme. Factorizing personal- ized markov chains for next-basket recommendation. In WWW, pages 811–820,

    work page 2010

  19. [19]

    Probabilistic matrix factorization

    [Salakhutdinov and Mnih, 2007] Ruslan Salakhutdinov and Andriy Mnih. Probabilistic matrix factorization. In NIPS, pages 1257–1264,

    work page 2007

  20. [20]

    Person- alized top-n sequential recommendation via convolutional sequence embedding

    [Tang and Wang, 2018] Jiaxi Tang and Ke Wang. Person- alized top-n sequential recommendation via convolutional sequence embedding. In WSDM, pages 565–573,

    work page 2018

  21. [21]

    Multi-pointer co-attention networks for recommen- dation

    [Tay et al., 2018] Yi Tay, Anh Tuan Luu, and Siu Cheung Hui. Multi-pointer co-attention networks for recommen- dation. In KDD, pages 2309–2318,

    work page 2018

  22. [22]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    [Vaswani et al., 2017] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 6000–6010,

    work page 2017

  23. [23]

    [Wang and Blei, 2011] Chong Wang and David M. Blei. Collaborative topic modeling for recommending scientific articles. In KDD, pages 448–456,

    work page 2011

  24. [24]

    Learn- ing hierarchical representation model for nextbasket rec- ommendation

    [Wang et al., 2015] Pengfei Wang, Jiafeng Guo, Yanyan Lan, Jun Xu, Shengxian Wan, and Xueqi Cheng. Learn- ing hierarchical representation model for nextbasket rec- ommendation. In SIGIR, pages 403–412,

    work page 2015

  25. [25]

    PARL: let strangers speak out what you like

    [Wu et al., 2018] Libing Wu, Cong Quan, Chenliang Li, and Donghong Ji. PARL: let strangers speak out what you like. In CIKM, pages 677–686,

    work page 2018

  26. [26]

    A context-aware user-item representation learning for item recommendation

    [Wu et al., 2019] Libing Wu, Cong Quan, Chenliang Li, Qian Wang, Bolong Zheng, and Xiangyang Luo. A context-aware user-item representation learning for item recommendation. ACM Trans. Inf. Syst. , 37(2):22:1– 22:29,

    work page 2019

  27. [27]

    A dynamic recurrent model for next basket recommendation

    [Yu et al., 2016] Feng Yu, Qiang Liu, Shu Wu, Liang Wang, and Tieniu Tan. A dynamic recurrent model for next basket recommendation. In SIGIR, pages 729–732,

    work page 2016

  28. [28]

    Explicit factor models for explainable recommendation based on phrase-level sentiment analysis

    [Zhang et al., 2014] Yongfeng Zhang, Guokun Lai, Min Zhang, Yi Zhang, Yiqun Liu, and Shaoping Ma. Explicit factor models for explainable recommendation based on phrase-level sentiment analysis. In SIGIR, pages 83–92,

    work page 2014

  29. [29]

    [Zheng et al., 2017] Lei Zheng, Vahid Noroozi, and Philip S. Yu. Joint deep modeling of users and items using reviews for recommendation. In WSDM, pages 425–434, 2017

    work page 2017