pith. sign in

arxiv: 1906.10910 · v2 · pith:LGAVLYK5new · submitted 2019-06-26 · 💻 cs.LG · cs.CL· stat.ML

Creating A Neural Pedagogical Agent by Jointly Learning to Review and Assess

Youngnam Lee , Youngduck Choi , Junghyun Cho , Alexander R. Fabbri , HyunBin Loh , Chanyou Hwang , Yongku Lee , Sang-Wook Kim
show 1 more author
Dragomir Radev
This is my paper

Pith reviewed 2026-05-25 15:33 UTC · model grok-4.3

classification 💻 cs.LG cs.CLstat.ML
keywords neural pedagogical agentuser modelingresponse correctness predictionrecurrent neural networksattention mechanismintelligent tutoring systemsmobile educationreal-time modeling
0
0 comments X

The pith

A bidirectional RNN with attention updates user knowledge in real time to predict response correctness more accurately than prior methods, especially for new users.

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

The paper introduces a neural pedagogical agent that maintains and refreshes a representation of each user's knowledge as soon as new question responses arrive, without requiring full model retraining. It processes sequences of embedded question-response pairs with bidirectional recurrent layers and an attention mechanism drawn from sequence modeling techniques. Experiments on a dataset of 559k users and 66M responses from a mobile TOEIC app show improved accuracy in forecasting whether a user will answer correctly, with the largest gains for users who have short interaction histories. The same attention weights, paired with expert topic tags on the questions, are used to surface explanations and to drive a smart review system that selects material for revisiting.

Core claim

Our model updates user features in real-time via bidirectional recurrent neural networks with an attention mechanism over embedded question-response pairs. Our model outperforms existing approaches over several metrics in predicting user response correctness, notably out-performing other methods on new users without large question-response histories. Additionally, our attention mechanism and annotated tag set allow us to create an interpretable education platform, with a smart review system that addresses the issue of varied user attention and problem exhaustion.

What carries the argument

bidirectional recurrent neural networks with an attention mechanism over embedded question-response pairs

If this is right

  • Real-time user feature updates become possible on mobile platforms where users arrive and change frequently.
  • Prediction accuracy improves for users who have accumulated only short response histories.
  • An interpretable review system can be built that uses attention weights and topic tags to select material for revisiting.

Where Pith is reading between the lines

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

  • The architecture could be extended to joint prediction of both correctness and next-problem recommendations within the same model.
  • Similar sequence-modeling approaches may transfer to other domains that require real-time updating of user state from sparse interaction logs.
  • The interpretability claim would be strengthened by measuring whether attention-highlighted topics actually change user behavior when presented in the review system.
  • The performance edge on new users suggests the model could reduce the cold-start problem in other sequential recommendation settings.

Load-bearing premise

The attention mechanism over embedded question-response pairs combined with expert topic tags produces reliable interpretability for the smart review system without separate validation of the explanations.

What would settle it

A direct comparison of the model's attention weights against independent expert ratings or user self-reports on which past questions most influence current correctness predictions would test whether the interpretability holds.

Figures

Figures reproduced from arXiv: 1906.10910 by Alexander R. Fabbri, Chanyou Hwang, Dragomir Radev, HyunBin Loh, Junghyun Cho, Sang-Wook Kim, Yongku Lee, Youngduck Choi, Youngnam Lee.

Figure 1
Figure 1. Figure 1: Changes in the number of users and questions. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: User interface of SantaTOEIC. embedding representations aim to capture co-occurrence relation￾ships in the embedding space. Additionally, sequence modeling, in particular tasks such as machine translation, have shown great im￾provements through the adoption of neural sequence-to-sequence models [5] and subsequent introduction of aŠention-based models [11]. AŠention mechanisms helped address the problem of … view at source ↗
Figure 4
Figure 4. Figure 4: Example of review. We call the sequence of question-response pairs the exhausted sequence. Œe user vectors ui are mapped using the exhausted ques￾tions qk and the responses rk ∈ {0, 1} to them. Œe details of the method will be introduced in Section 4.3. To capture the temporal properties of exhausted sequences, we add time indices to the po￾tential questions and the exhausted sequence. Œus Eq.(1) becomes: … view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the architecture. Once our model is trained, the model can adaptively map user vectors in the latent space without extra training. Œis enables our model to map new users without updating the model. 4.2 Data Representation Typically, words (or other units such as characters, sentences) are embedded in vector spaces of appropriate dimensions, and these vec￾tors are €Šed to capture linguistic prop… view at source ↗
Figure 6
Figure 6. Figure 6: Attention and tag matching [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Bidirectional LSTM output heatmap. 5 Experiments Here we introduce our dataset, training seŠings as well as experi￾mental results and analysis. 5.1 Dataset We use the SantaTOEIC dataset, which is a set of user responses of multiple-choice questions collected over the last four years in the SantaTOEIC service (Android and iOS). Œe main features of the user-question response data are the following four colum… view at source ↗
Figure 8
Figure 8. Figure 8: ‡e F1-score by timestep [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: ‡e attention trajectory. Modelling. arXiv preprint arXiv:1807.00154 (2018). [8] Christian Desrosiers and George Karypis. 2011. A Comprehensive Survey of Neighborhood-Based Recommendation Methods. In Recommender systems hand￾book. Springer, 107–144. [9] Benedict du Boulay. 2016. ARTIFICIAL INTELLIGENCE AS AN EFFECTIVE CLASSROOM ASSISTANT. IEEE Intelligent Systems 31, 6 (2016), 76–81. [10] Arthur K Ellis, Da… view at source ↗
read the original abstract

Machine learning plays an increasing role in intelligent tutoring systems as both the amount of data available and specialization among students grow. Nowadays, these systems are frequently deployed on mobile applications. Users on such mobile education platforms are dynamic, frequently being added, accessing the application with varying levels of focus, and changing while using the service. The education material itself, on the other hand, is often static and is an exhaustible resource whose use in tasks such as problem recommendation must be optimized. The ability to update user models with respect to educational material in real-time is thus essential; however, existing approaches require time-consuming re-training of user features whenever new data is added. In this paper, we introduce a neural pedagogical agent for real-time user modeling in the task of predicting user response correctness, a central task for mobile education applications. Our model, inspired by work in natural language processing on sequence modeling and machine translation, updates user features in real-time via bidirectional recurrent neural networks with an attention mechanism over embedded question-response pairs. We experiment on the mobile education application SantaTOEIC, which has 559k users, 66M response data points as well as a set of 10k study problems each expert-annotated with topic tags and gathered since 2016. Our model outperforms existing approaches over several metrics in predicting user response correctness, notably out-performing other methods on new users without large question-response histories. Additionally, our attention mechanism and annotated tag set allow us to create an interpretable education platform, with a smart review system that addresses the aforementioned issue of varied user attention and problem exhaustion.

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 introduces a neural pedagogical agent for real-time user modeling in mobile education apps. It uses bidirectional RNNs with an attention mechanism over embedded question-response pairs to predict user response correctness, updating features without retraining. Experiments on the SantaTOEIC dataset (559k users, 66M responses, 10k expert-tagged problems) claim outperformance over existing methods on several metrics, especially for new users lacking large histories, while the attention and tags enable an interpretable smart review system.

Significance. If the new-user generalization holds without data leakage, the real-time update capability without retraining would be a meaningful advance for dynamic user modeling in intelligent tutoring systems on mobile platforms. The combination of sequence modeling and expert annotations for interpretability is a constructive direction, though the absence of separate validation for the generated explanations reduces the immediate practical significance.

major comments (2)
  1. [Abstract] Abstract: The central empirical claim is outperformance on new users without large question-response histories. However, the manuscript provides no description of the train/test split procedure (user-level vs. response-level). A response-level split would allow shared RNN parameters and embeddings to encode user-specific patterns from a user's training responses, rendering the reported gains on 'new' users non-generalizable and undermining the headline result.
  2. [Experiments] Experiments section: The abstract asserts outperformance 'over several metrics' and 'notably' on new users, yet reports no numerical values, baseline implementations, statistical significance tests, or error analysis. This absence makes it impossible to assess whether the gains are substantive or whether they survive proper user-stratified evaluation.
minor comments (1)
  1. The abstract states that the model 'jointly learns to review and assess' in the title but the described architecture focuses on correctness prediction with attention used post-hoc for review; a clearer statement of the joint objective would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive feedback. The two major comments highlight important gaps in the description of the evaluation protocol and the reporting of results. We address each below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central empirical claim is outperformance on new users without large question-response histories. However, the manuscript provides no description of the train/test split procedure (user-level vs. response-level). A response-level split would allow shared RNN parameters and embeddings to encode user-specific patterns from a user's training responses, rendering the reported gains on 'new' users non-generalizable and undermining the headline result.

    Authors: We agree that the absence of an explicit description of the train/test split is a serious omission that prevents proper assessment of the new-user results. The current manuscript does not state whether the split was performed at the user or response level. In the revision we will add a clear statement that all experiments use a user-stratified split (entire user histories are assigned wholly to train or test), which is the only procedure that supports the claimed generalization to new users. We will also move this description from the supplementary material into the main Experiments section. revision: yes

  2. Referee: [Experiments] Experiments section: The abstract asserts outperformance 'over several metrics' and 'notably' on new users, yet reports no numerical values, baseline implementations, statistical significance tests, or error analysis. This absence makes it impossible to assess whether the gains are substantive or whether they survive proper user-stratified evaluation.

    Authors: We acknowledge that the Experiments section (and abstract) currently lacks the concrete numbers, baseline implementation details, significance tests, and error analysis needed to evaluate the claims. In the revision we will insert a table of exact metric values (AUC, accuracy, etc.) for all methods, describe the baseline implementations, report paired statistical tests, and add a short error analysis focused on the new-user regime. These additions will be placed in the main text rather than only in supplementary material. revision: yes

Circularity Check

0 steps flagged

No significant circularity; model derivation is self-contained

full rationale

The paper describes a standard neural sequence model (bidirectional RNN + attention over embedded question-response pairs) trained on observed response data from SantaTOEIC to predict held-out response correctness. No equations, self-citations, or ansatzes are presented that reduce the claimed predictions or interpretability features to the inputs by construction; the central empirical claims rest on external data splits and evaluation metrics rather than definitional equivalence or load-bearing self-references. The derivation chain from embeddings through the RNN/attention layers to correctness predictions is independent and falsifiable on held-out data.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review based on abstract only; no explicit free parameters, axioms, or invented entities stated. Relies on standard supervised learning assumptions for sequence models.

axioms (1)
  • standard math Standard neural network training assumptions hold, including that gradient-based optimization finds useful representations from the given data.
    Implicit in any RNN training description.

pith-pipeline@v0.9.0 · 5855 in / 1150 out tokens · 39074 ms · 2026-05-25T15:33:58.263732+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

48 extracted references · 48 canonical work pages · 8 internal anchors

  1. [1]

    Jun Araki, Dheeraj Rajagopal, Sreecharan Sankaranarayanan, Susan Holm, Yukari Yamakawa, and Teruko Mitamura. 2016. Generating /Q_uestions and Multiple- Choice Answers using Semantic Analysis of Texts. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers. 1125–1136

    work page 2016

  2. [2]

    Frank B. Baker. 1985. /T_he Basics of Item Response /T_heory. ERIC Clearinghouse on Assessment and Evaluation

    work page 1985

  3. [3]

    Robert M Bell and Yehuda Koren. 2007. Improved Neighborhood-Based Collabo- rative Filtering. In KDD cup and workshop at the 13th ACM SIGKDD international conference on knowledge discovery and data mining . Citeseer, 7–14

    work page 2007

  4. [4]

    JESUS Bobadilla, Francisco Serradilla, Antonio Hernando, et al. 2009. Collabora- tive Filtering Adapted to Recommender Systems of E-Learning.Knowledge-Based Systems 22, 4 (2009), 261–265

    work page 2009

  5. [5]

    Kyunghyun Cho, Bart Van Merri¨enboer, Dzmitry Bahdanau, and Yoshua Ben- gio. 2014. On the Properties of Neural Machine Translation: Encoder-Decoder Approaches. arXiv preprint arXiv:1409.1259 (2014)

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  6. [6]

    Edward Choi, Mohammad Taha Bahadori, Jimeng Sun, Joshua Kulas, Andy Schuetz, and Walter Stewart. 2016. RETAIN: An Interpretable Predictive Model for Healthcare using Reverse Time A/t_tention Mechanism. InAdvances in Neural Information Processing Systems. 3504–3512

    work page 2016

  7. [7]

    Cristina Conati, Kaska Porayska-Pomsta, and Manolis Mavrikis. 2018. AI in Education needs interpretable machine learning: Lessons from Open Learner Creating A Neural Pedagogical Agent by Jointly Learning to Review and Assess Conference’17, July 2017, Washington, DC, USA Figure 9: /T_he attention trajectory. Modelling. arXiv preprint arXiv:1807.00154 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  8. [8]

    Christian Desrosiers and George Karypis. 2011. A Comprehensive Survey of Neighborhood-Based Recommendation Methods. In Recommender systems hand- book. Springer, 107–144

    work page 2011

  9. [9]

    Benedict du Boulay. 2016. ARTIFICIAL INTELLIGENCE AS AN EFFECTIVE CLASSROOM ASSISTANT. IEEE Intelligent Systems 31, 6 (2016), 76–81

    work page 2016

  10. [10]

    Arthur K Ellis, David W Denton, and John B Bond. 2014. An Analysis of Research on Metacognitive Teaching Strategies. InProcedia - Social and Behavioral Sciences, Volume 116. 4015–4024

    work page 2014

  11. [11]

    Bahdanau et al. 2014. NEURAL MACHINE TRANSLATION BY JOINTLY LEARN- ING TO ALIGN AND TRANSLATE. arXiv preprint arXiv:1409.0473 (2014)

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  12. [12]

    Mikolov et al. 2013. Efficient Estimation of Word Representations in Vector Space. arXiv preprint arXiv:1301.3781 (2013)

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  13. [13]

    Leilani H Gilpin, David Bau, Ben Z Yuan, Ayesha Bajwa, Michael Specter, and Lalana Kagal. 2018. Explaining Explanations: An Overview of Interpretability of Machine Learning. In 2018 IEEE 5th International Conference on Data Science and Advanced Analytics (DSAA). IEEE, 80–89

    work page 2018

  14. [14]

    Xavier Glorot and Yoshua Bengio. 2010. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on arti/f_icial intelligence and statistics. 249–256

    work page 2010

  15. [15]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural Collaborative Filtering. InProceedings of the 26th International Conference on World Wide Web . International World Wide Web Conferences Steering Commi/t_tee, 173–182

    work page 2017

  16. [16]

    Sepp Hochreiter and J ¨urgen Schmidhuber. 1997. Long Short-Term Memory. Neural computation 9, 8 (1997), 1735–1780

    work page 1997

  17. [17]

    Aarij Mahmood Hussaan and Karim Sehaba. 2014. Learn and Evolve the Domain Model in Intelligent Tutoring Systems. In CSEDU 2014 - Proceedings of the 6th International Conference on Computer Supported Education . 197–204

    work page 2014

  18. [18]

    TabMCQ: A Dataset of General Knowledge Tables and Multiple-choice Questions

    Sujay Kumar Jauhar, Peter D. Turney, and Eduard H. Hovy. 2016. TabMCQ: A Dataset of General Knowledge Tables and Multiple-choice /Q_uestions. CoRR abs/1602.03960 (2016)

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  19. [19]

    Diederik P Kingma and Jimmy Ba. 2014. ADAM: A METHOD FOR STOCHASTIC OPTIMIZATION. arXiv preprint arXiv:1412.6980 (2014)

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  20. [20]

    Andr´e Klaßen, Marcus Eibrink-Lunzenauer, and Till Gl¨oggler. 2013. Requirements for mobile learning applications in higher education. In 2013 IEEE International Symposium on Multimedia. IEEE, 492–497

    work page 2013

  21. [21]

    Kopp, Amy M

    Kristopher J. Kopp, Amy M. Johnson, Sco/t_t A. Crossley, and Danielle S. McNamara

    work page

  22. [22]

    InArti/f_icial Intelligence in Education - 18th International Conference, AIED 2017, Wuhan, China, June 28 - July 1, 2017, Proceedings

    Assessing /Q_uestion /Q_uality using NLP. InArti/f_icial Intelligence in Education - 18th International Conference, AIED 2017, Wuhan, China, June 28 - July 1, 2017, Proceedings. 523–527

    work page 2017

  23. [23]

    Yehuda Koren. 2008. Factorization Meets the Neighborhood: A Multifaceted Col- laborative Filtering Model. In Proceedings ACM SIGKDD International Conference on Knowledge Discovery and Data mining . ACM, 426–434

    work page 2008

  24. [24]

    Vassilis Kostakos and Mirco Musolesi. 2017. Avoiding Pitfalls When Using Machine Learning in HCI Studies. interactions 24, 4 (2017), 34–37

    work page 2017

  25. [25]

    Oleksii Kuchaiev and Boris Ginsburg. 2017. Training Deep AutoEncoders for Collaborative Filtering. arXiv preprint arXiv:1708.01715 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  26. [26]

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. 2015. Deep Learning. nature 521, 7553 (2015), 436

    work page 2015

  27. [27]

    Kangwook Lee, Jichan Chung, Yeongmin Cha, and Changho Suh. 2016. Ma- chine Learning Approaches for Learning Analytics: Collaborative Filtering Or Regression With Experts?. In NIPS Workshop. NIPS

    work page 2016

  28. [28]

    Minh-/T_hang Luong, Hieu Pham, and Christopher D Manning. 2015. Effective Approaches to A/t_tention-based Neural Machine Translation. arXiv preprint arXiv:1508.04025 (2015)

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  29. [29]

    Peter Afflerbach Marcel V. J. Veenman, Bernade/t_te H. A. M. Van Hout-Wolters

    work page

  30. [30]

    Kluwer Academic Publishers

    Metacognition and learning. Kluwer Academic Publishers

    work page

  31. [31]

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed Representations of Words and Phrases and their Compositionality. In Advances in Neural Information Processing Systems . 3111–3119

    work page 2013

  32. [32]

    Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig. 2013. Linguistic Regularities in Continuous Space Word Representations. InProceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 746–751

    work page 2013

  33. [33]

    Andriy Mnih and Ruslan R Salakhutdinov. 2008. Probabilistic Matrix Factoriza- tion. In Advances in Neural Information Processing Systems . 1257–1264

    work page 2008

  34. [34]

    Fumiya Okubo, Takayoshi Yamashita, Atsushi Shimada, and Hiroaki Ogata. 2017. A Neural Network Approach for Students’ Performance Prediction. InProceedings of the Seventh International Learning Analytics & Knowledge Conference . ACM, 598–599

    work page 2017

  35. [35]

    Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global Vectors for Word Representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) . 1532–1543

    work page 2014

  36. [36]

    Turney, and Oren Etzioni

    Carissa Schoenick, Peter Clark, Oyvind Ta/f_jord, Peter D. Turney, and Oren Etzioni

    work page

  37. [37]

    Moving Beyond the Turing Test with the Allen AI Science Challenge. Commun. ACM 60, 9 (2017), 60–64

    work page 2017

  38. [38]

    Mike Schuster and Kuldip K Paliwal. 1997. Bidirectional Recurrent Neural Networks. IEEE Transactions on Signal Processing 45, 11 (1997), 2673–2681

    work page 1997

  39. [39]

    Suvash Sedhain, Aditya Krishna Menon, Sco/t_t Sanner, and Lexing Xie. 2015. AutoRec: Autoencoders Meet Collaborative Filtering. In Proceedings of the 24th International Conference on World Wide Web. ACM, 111–112

    work page 2015

  40. [40]

    Nguyen /T_hai-Nghe, Lucas Drumond, Artus Krohn-Grimberghe, and Lars Schmidt- /T_hieme. 2010. Recommender System for Predicting Student Performance.Proce- dia Computer Science 1, 2 (2010), 2811–2819

    work page 2010

  41. [41]

    Andreas Toscher and Michael Jahrer. 2010. Collaborative Filtering Applied to Educational Data Mining. KDD cup (2010)

    work page 2010

  42. [42]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. A/t_tention Is All You Need. In Advances in neural information processing systems . 5998–6008

    work page 2017

  43. [43]

    Adrian Weller. 2017. Challenges for Transparency.arXiv preprint arXiv:1708.01870 (2017)

    work page arXiv 2017

  44. [44]

    Darrell M West. 2015. Connected learning: How mobile technology can improve education. Center for Technology Innovation at Brookings. Retrieved March 25 (2015), 2016

    work page 2015

  45. [45]

    Chao-Yuan Wu, Amr Ahmed, Alex Beutel, Alexander J Smola, and How Jing. 2017. Recurrent Recommender Networks. In Proceedings of the 10th ACM International Conference on Web Search and Data Mining . ACM, 495–503

    work page 2017

  46. [46]

    Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. 2015. Show, A/t_tend and Tell: Neu- ral Image Caption Generation with Visual A/t_tention. InInternational conference on machine learning. 2048–2057

    work page 2015

  47. [47]

    Vincent Aleven Yanjin Long. 2017. Enhancing learning outcomes through self- regulated learning support with an Open Learner Model. In User Modeling and User-Adapted Interaction. 55–88

    work page 2017

  48. [48]

    Shuai Zhang, Lina Yao, Aixin Sun, and Yi Tay. 2019. Deep Learning based Recommender System: A Survey and New Perspectives.ACM Computing Surveys (CSUR) 52, 1 (2019), 5

    work page 2019