pith. sign in

arxiv: 1907.01639 · v1 · pith:VSLTDTQVnew · submitted 2019-06-24 · 💻 cs.IR · cs.LG

Query-based Interactive Recommendation by Meta-Path and Adapted Attention-GRU

Yu Zhu , Yu Gong , Qingwen Liu , Yingcai Ma , Wenwu Ou , Junxiong Zhu , Beidou Wang , Ziyu Guan
show 1 more author
Deng Cai
This is my paper

Pith reviewed 2026-05-25 17:32 UTC · model grok-4.3

classification 💻 cs.IR cs.LG
keywords interactive recommendationquery recommendationmeta-pathattention-GRUpersonalized questionssingle-round interactione-commerce recommendation
0
0 comments X

The pith

Meta-path retrieval followed by adapted Attention-GRU ranking generates personalized questions from millions of candidates to drive single-round interactive recommendations.

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

The paper establishes a query-based interactive recommender that asks one personalized question and then produces recommendations tightly linked to the user's answer. It converts question generation into a retrieval-plus-ranking task: a Meta-Path model first pulls hundreds of candidate questions from a pool of millions of automatically constructed ones, after which an adapted Attention-GRU ranks the candidates for the current user. The resulting system is claimed to let users actively shape results while giving the platform more accurate user signals in a single exchange. Experiments on Taobao data and live traffic are presented as verification that the generated questions are both accurate and effective for refining recommendations.

Core claim

The paper claims that transforming single-round interaction into a query-recommendation task, solved by Meta-Path candidate retrieval from a large automatically-generated pool followed by adapted Attention-GRU ranking, produces personalized questions whose feedback yields recommendations closely related to user responses.

What carries the argument

Meta-Path model that retrieves hundreds of query candidates from millions of automatically constructed questions, followed by an adapted Attention-GRU that ranks the candidates for personalization.

If this is right

  • Users can actively intervene in recommendation results through one targeted question rather than passively receiving lists.
  • The platform obtains richer user signals from the feedback to the generated question, enabling more accurate follow-up recommendations.
  • The two-step pipeline (Meta-Path retrieval then Attention-GRU ranking) scales to production traffic on large e-commerce platforms.
  • The approach has been deployed in the Taobao App homepage since November 2018.

Where Pith is reading between the lines

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

  • The same candidate-generation-plus-ranking pattern could be tested on multi-round conversations if the single-round version succeeds.
  • Automatic construction of the initial question pool might be replaced by other generative methods without changing the downstream ranking stage.
  • The framework could transfer to non-e-commerce domains that maintain large sets of natural-language queries about items or content.

Load-bearing premise

The meta-path retrieval step surfaces a sufficiently complete and relevant set of query candidates without systematic omission of high-value questions for particular users.

What would settle it

An A/B test in which users who receive the system's generated questions show no improvement in recommendation metrics over a non-interactive baseline that skips the question step.

Figures

Figures reproduced from arXiv: 1907.01639 by Beidou Wang, Deng Cai, Junxiong Zhu, Qingwen Liu, Wenwu Ou, Yingcai Ma, Yu Gong, Yu Zhu, Ziyu Guan.

Figure 1
Figure 1. Figure 1: The Question&Recommendation framework. explore-exploit tradeoff in RS. They learn users’ preferences by recommending explored items and acquiring their feedback. How￾ever, personal assistants and reinforcement learning/multi-armed bandit/active learning methods usually need multi-round inter￾actions to well learn users, thus users who want quick&accurate recommendations would not be satisfied. The question… view at source ↗
Figure 2
Figure 2. Figure 2: The system generates 4 queries to consult the user [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Three types of meta paths to generate query candi [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Attention-GRU based model to rank query candidates. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Modified attention schema considering action [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: A/B test setting. (a) is the original setting. Our in [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
read the original abstract

Recently, interactive recommender systems are becoming increasingly popular. The insight is that, with the interaction between users and the system, (1) users can actively intervene the recommendation results rather than passively receive them, and (2) the system learns more about users so as to provide better recommendation. We focus on the single-round interaction, i.e. the system asks the user a question (Step 1), and exploits his feedback to generate better recommendation (Step 2). A novel query-based interactive recommender system is proposed in this paper, where \textbf{personalized questions are accurately generated from millions of automatically constructed questions} in Step 1, and \textbf{the recommendation is ensured to be closely-related to users' feedback} in Step 2. We achieve this by transforming Step 1 into a query recommendation task and Step 2 into a retrieval task. The former task is our key challenge. We firstly propose a model based on Meta-Path to efficiently retrieve hundreds of query candidates from the large query pool. Then an adapted Attention-GRU model is developed to effectively rank these candidates for recommendation. Offline and online experiments on Taobao, a large-scale e-commerce platform in China, verify the effectiveness of our interactive system. The system has already gone into production in the homepage of Taobao App since Nov. 11, 2018 (see https://v.qq.com/x/page/s0833tkp1uo.html on how it works online). Our code and dataset are public in https://github.com/zyody/QueryQR.

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 / 2 minor

Summary. The paper proposes a single-round query-based interactive recommender system. Step 1 transforms question generation into a query recommendation task: a meta-path model retrieves hundreds of candidate questions from a pool of millions of automatically constructed queries, after which an adapted Attention-GRU ranks them for personalization. Step 2 treats recommendation as a retrieval task that incorporates user feedback. Offline and online experiments on Taobao data plus production deployment since November 2018 are reported, with code and dataset released.

Significance. If the central claims hold, the work demonstrates a deployable interactive system on a large e-commerce platform that combines graph-based candidate retrieval with neural ranking, supported by real user-interaction data rather than synthetic or parameter-fitted proxies. Public release of code and dataset strengthens reproducibility.

major comments (2)
  1. [Abstract, §3] Abstract and §3 (meta-path retrieval description): the claim that 'personalized questions are accurately generated from millions of automatically constructed questions' rests on the meta-path step retrieving a sufficiently complete and non-systematically biased set of candidates; no coverage metric, recall@K against the full pool, or ablation comparing meta-path retrieval to random or exhaustive baselines is provided to rule out user-specific omissions.
  2. [§5] §5 (experiments): offline and online results are reported without error bars, statistical significance tests, or component ablations isolating the contribution of the meta-path retrieval versus the Attention-GRU ranker; this weakens the ability to confirm that the retrieval step does not create blind spots that the subsequent ranking cannot recover.
minor comments (2)
  1. [§3, §4] Notation for meta-path definitions and the exact form of the adapted Attention-GRU (e.g., how the attention is modified from standard GRU) could be stated more formally with equations.
  2. [§5] Figure captions and table headers should explicitly state whether reported metrics are averages over multiple runs or single-run values.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and commit to revisions that strengthen the evaluation of the meta-path component and experimental reporting.

read point-by-point responses

  1. Referee: [Abstract, §3] Abstract and §3 (meta-path retrieval description): the claim that 'personalized questions are accurately generated from millions of automatically constructed questions' rests on the meta-path step retrieving a sufficiently complete and non-systematically biased set of candidates; no coverage metric, recall@K against the full pool, or ablation comparing meta-path retrieval to random or exhaustive baselines is provided to rule out user-specific omissions.

    Authors: We agree that the manuscript would be strengthened by quantitative evidence on coverage and potential bias. The meta-path construction is intended to surface only queries reachable via user-item paths in the heterogeneous graph, which inherently limits the candidate set to relevant ones; exhaustive search over millions of queries is computationally infeasible at serving time. In the revision we will add (i) a coverage analysis reporting the fraction of the full query pool reachable via meta-paths for held-out users, (ii) recall@K of the meta-path step versus random sampling of the same cardinality, and (iii) an ablation replacing meta-path retrieval with random or popularity-based candidates before the Attention-GRU ranker. These results will appear in an expanded Section 3 and in the experimental section. revision: yes

  2. Referee: [§5] §5 (experiments): offline and online results are reported without error bars, statistical significance tests, or component ablations isolating the contribution of the meta-path retrieval versus the Attention-GRU ranker; this weakens the ability to confirm that the retrieval step does not create blind spots that the subsequent ranking cannot recover.

    Authors: We accept that the current experimental section lacks error bars, significance testing, and component ablations. Because the code and dataset are already public, we can recompute all offline metrics with standard deviations and paired t-tests (or Wilcoxon tests) across multiple random seeds. We will also insert an ablation table that reports end-to-end performance when the meta-path retriever is replaced by random or exhaustive top-K baselines before ranking. These additions will directly address whether the retrieval stage introduces irrecoverable omissions. The revised Section 5 will contain the new tables and statistical results. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation and evaluation are self-contained against external benchmarks.

full rationale

The paper proposes a two-step interactive recommender: meta-path retrieval of query candidates from a constructed pool, followed by adapted Attention-GRU ranking. These steps are presented as a new model without reducing to fitted parameters or self-citations that bear the central load. Validation relies on offline experiments and online production deployment on the external Taobao platform, plus public code/dataset release, rather than quantities defined internally by the paper's own equations. No self-definitional, fitted-input-as-prediction, or uniqueness-imported patterns appear in the described chain.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The approach rests on standard assumptions of graph-based retrieval and recurrent neural ranking; no new entities are postulated. Free parameters are typical neural network hyperparameters and the choice of hundreds of candidates.

free parameters (2)
  • number of query candidates retrieved by meta-path
    Chosen to balance coverage and efficiency; value stated as hundreds in the abstract.
  • Attention-GRU model hyperparameters
    Standard neural network parameters fitted during training on Taobao data.
axioms (2)
  • domain assumption Meta-paths in the user-item-query graph capture relevant query candidates for a given user
    Invoked when transforming Step 1 into a query recommendation task.
  • domain assumption User feedback on the selected query is sufficient to retrieve closely related recommendations
    Central to the retrieval task in Step 2.

pith-pipeline@v0.9.0 · 5835 in / 1265 out tokens · 39629 ms · 2026-05-25T17:32:01.278693+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

34 extracted references · 34 canonical work pages · 1 internal anchor

  1. [1]

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv:1409.0473 (2014)

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  2. [2]

    Shi-Yong Chen, Yang Yu, Qing Da, Jun Tan, Hai-Kuan Huang, and Hai-Hong Tang. 2018. Stabilizing reinforcement learning in dynamic environment with application to online recommendation. In KDD. ACM, 1187–1196

    work page 2018

  3. [3]

    Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, et al

    work page

  4. [4]

    In Proceedings of the 1st workshop on deep learning for recommender systems

    Wide & deep learning for recommender systems. In Proceedings of the 1st workshop on deep learning for recommender systems . ACM, 7–10

    work page

  5. [5]

    Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv (2014)

    work page 2014

  6. [6]

    Jan K Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio. 2015. Attention-based models for speech recognition. In NIPS. 577–585

    work page 2015

  7. [7]

    Konstantina Christakopoulou, Alex Beutel, Rui Li, Sagar Jain, and Ed H Chi. 2018. Q&r: A two-stage approach toward interactive recommendation. In KDD. ACM, 139–148

    work page 2018

  8. [8]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for youtube recommendations. In Recsys. ACM, 191–198

    work page 2016

  9. [9]

    Mehdi Elahi, Matthias Braunhofer, Francesco Ricci, and Marko Tkalcic. 2013. Personality-based active learning for collaborative filtering recommender sys- tems. In Congress of the Italian Association for Artificial Intelligence . Springer, 360–371

    work page 2013

  10. [10]

    Chen He, Denis Parra, and Katrien Verbert. 2016. Interactive recommender systems: A survey of the state of the art and future research challenges and opportunities. Expert Systems with Applications 56 (2016), 9–27

    work page 2016

  11. [11]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

    work page

  12. [12]

    Session-based recommendations with recurrent neural networks. In ICLR

    work page

  13. [13]

    Balázs Hidasi, Massimo Quadrana, Alexandros Karatzoglou, and Domonkos Tikk

    work page

  14. [14]

    In RecSys

    Parallel recurrent neural network architectures for feature-rich session- based recommendations. In RecSys. ACM, 241–248

    work page

  15. [15]

    Geoffrey Hinton, Li Deng, Dong Yu, George Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Brian Kings- bury, et al. 2012. Deep neural networks for acoustic modeling in speech recogni- tion. IEEE Signal processing magazine 29 (2012)

    work page 2012

  16. [16]

    Zhipeng Huang, Bogdan Cautis, Reynold Cheng, and Yudian Zheng. 2016. Kb- enabled query recommendation for long-tail queries. In IJCAI. ACM, 2107–2112

    work page 2016

  17. [17]

    Michael Jugovac and Dietmar Jannach. 2017. Interacting with recommenders: overview and research directions. TiiS 7, 3 (2017), 10

    work page 2017

  18. [18]

    Myunghwan Kim and Jure Leskovec. 2013. Nonparametric multi-group member- ship model for dynamic networks. In NIPS. 1385–1393

    work page 2013

  19. [19]

    Lihong Li, Wei Chu, John Langford, and Robert E Schapire. 2010. A contextual- bandit approach to personalized news article recommendation. In WWW. ACM, 661–670

    work page 2010

  20. [20]

    H Brendan McMahan, Gary Holt, David Sculley, Michael Young, Dietmar Ebner, Julian Grady, Lan Nie, Todd Phillips, Eugene Davydov, Daniel Golovin, et al. 2013. Ad click prediction: a view from the trenches. In KDD. ACM, 1222–1230

    work page 2013

  21. [21]

    Volodymyr Mnih, Nicolas Heess, Alex Graves, et al. 2014. Recurrent models of visual attention. In NIPS. 2204–2212

    work page 2014

  22. [22]

    Michael J Pazzani and Daniel Billsus. 2007. Content-based recommendation systems. In The adaptive web. Springer, 325–341

    work page 2007

  23. [23]

    Alexander Rakhlin, Ohad Shamir, and Karthik Sridharan. 2012. Making Gradient Descent Optimal for Strongly Convex Stochastic Optimization. InICML. 449–456

    work page 2012

  24. [24]

    Neil Rubens, Mehdi Elahi, Masashi Sugiyama, and Dain Kaplan. 2015. Active learning in recommender systems. In Recommender systems handbook . Springer, 809–846

    work page 2015

  25. [25]

    Badrul Munir Sarwar, George Karypis, Joseph A Konstan, John Riedl, et al. 2001. Item-based collaborative filtering recommendation algorithms. Www 1 (2001), 285–295

    work page 2001

  26. [26]

    Yizhou Sun, Jiawei Han, Xifeng Yan, Philip S Yu, and Tianyi Wu. 2011. Pathsim: Meta path-based top-k similarity search in heterogeneous information networks. VLDB 4, 11 (2011), 992–1003

    work page 2011

  27. [27]

    Jiliang Tang, Xia Hu, and Huan Liu. 2013. Social recommendation: a review. Social Network Analysis and Mining 3, 4 (2013), 1113–1133

    work page 2013

  28. [28]

    Xiao Yu, Xiang Ren, Quanquan Gu, Yizhou Sun, and Jiawei Han. 2013. Collabora- tive filtering with entity similarity regularization in heterogeneous information networks. IJCAI HINA 27 (2013). Query-based Interactive Recommendation by Meta-Path and Adapted Attention-GRU CIKM ’19, November 03–07, 2019, Beijing, China

    work page 2013

  29. [29]

    Xiao Yu, Xiang Ren, Yizhou Sun, Quanquan Gu, Bradley Sturt, Urvashi Khandel- wal, Brandon Norick, and Jiawei Han. 2014. Personalized entity recommendation: A heterogeneous information network approach. In WSDM. ACM, 283–292

    work page 2014

  30. [30]

    Huan Zhao, Quanming Yao, Jianda Li, Yangqiu Song, and Dik Lun Lee. 2017. Meta- graph based recommendation fusion over heterogeneous information networks. In KDD. ACM, 635–644

    work page 2017

  31. [31]

    Zhou Zhao, Ruihua Song, Xing Xie, Xiaofei He, and Yueting Zhuang. 2015. Mobile query recommendation via tensor function learning. In IJCAI

    work page 2015

  32. [32]

    Zhi-Dan Zhao and Ming-Sheng Shang. 2010. User-based collaborative-filtering recommendation algorithms on hadoop. In 2010 Third International Conference on Knowledge Discovery and Data Mining . IEEE, 478–481

    work page 2010

  33. [33]

    Yu Zhu, Hao Li, Yikang Liao, Beidou Wang, Ziyu Guan, Haifeng Liu, and Deng Cai. 2017. What to do next: Modeling user behaviors by time-lstm. In IJCAI. 3602–3608

    work page 2017

  34. [34]

    Yu Zhu, Junxiong Zhu, Jie Hou, Yongliang Li, Beidou Wang, Ziyu Guan, and Deng Cai. 2018. A Brand-level Ranking System with the Customized Attention-GRU Model. In IJCAI. ACM, 3947–3953

    work page 2018