Personalized Education with Ranking Alignment Recommendation

Haipeng Liu; Ting Long; Yuxuan Liu

REVIEW 4 major objections 6 minor 16 references

Personalized Education with Ranking Alignment Recommendation

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper shows that adding a ranking-alignment loss to any RL-based question recommender improves exploration efficiency and learning effects in simulated education environments.

desk verdict The loss in Eq. 14 is not ranking alignment—it is a one-sided diversity margin that never pulls similar students together—so the central attribution is unsupported, but the idea may still be salvageable as a diversity exploration trick. read the letter →

arxiv 2507.23664 v1 pith:XZSUNI3R submitted 2025-07-31 cs.AI cs.IR

Haipeng Liu , Yuxuan Liu , Ting Long This is my paper · ORCID

classification cs.AIcs.IR

keywords questionrecommendationreinforcementlearningexplorationrankingalignmentcollaborativeinformationpersonalizededucationknowledgetracingeffect

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Question recommenders that use reinforcement learning to lead students through exercises struggle to explore efficiently: within a limited number of training episodes they fail to identify the most useful questions for each student. This paper claims that the bottleneck is exploration, not policy expressiveness, and that it can be fixed by a collaborative ranking-alignment module. The module measures the difference between two students by the symmetric difference of their learning-target question sets, measures the difference between their recommended question sequences, and adds a loss that pushes the ranking of these two difference measures to agree. The authors report that the two resulting variants, RAR-S and RAR-A, achieve the best or second-best learning effects in all five simulated environments, and that plugging the module into two standard off-the-shelf RL algorithms improves them in most cases. If the claim holds, any RL-based question recommender can gain efficiency simply by adding this loss term.

What carries the argument

The ranking alignment module (the 'RAM') is the central mechanism: it takes any RL-based recommendation module's output probabilities and produces a differentiable regularizer. Its student-difference metric is the symmetric difference of learning-target sets; its recommendation-difference metric is the L2 distance between encoded recommendation sequences; and its rank loss $\text{Clip}(\psi d^t_{uv} - d^p_{uv}, 0, \omega)$ forces the two rankings to align. This loss converts collaborative information into an exploration signal, reducing exploration complexity for similar students while directing exploration toward differentiated questions for dissimilar ones.

What would settle it

A direct test would replace the simulated environments with a small deployment involving real students answering recommended questions, then compare the learning effects of RAR-S/RAR-A against the strongest baseline under identical conditions. If the ranking-alignment advantage does not appear or reverses with real responses, the central claim that RAR improves any RL-based recommender in practice would fail. More narrowly, one could vary the simulator's noise structure (e.g., adding guessing or slipping parameters) and check whether the benefit persists.

Watch

Extended reading notes

Core claim

The central claim is that efficient exploration for personalized question recommendation can be obtained by aligning the ordering of student differences with the ordering of recommendation differences. Student difference is defined as the cardinality of the symmetric difference of the two students' learning-target question sets, $d^t_{uv} = |T_u \cup T_v - T_u \cap T_v|$. Recommendation difference is the L2 distance between two representations of the recommended question sequence, either a sequential encoding (RAR-S) or the summed per-step recommendation probabilities (RAR-A), $d^p_{uv} = \|b_u - b_v\|_2$. The ranking alignment loss, $L_r = \sum_{u}\sum_{v} \text{Clip}(\psi d^t_{uv} - d^p_{uv}, 0, \omega)$, is added to the policy-gradient and knowledge-tracing losses, encouraging the recommender to differentiate recommendations for dissimilar students and to consolidate exploration for similar ones. The paper evaluates this in five simulated student environments built from a rule-based simulator and two deep knowledge-tracing models trained on two real datasets, and finds that RAR variants outperform all baselines in essentially all conditions and that the module transfers to DQN and SAC in most environments.

Load-bearing premise

All reported gains are measured against simulated students, so the claim that RAR improves any RL-based recommender depends on these simulators faithfully reproducing real students' learning dynamics.

Editorial extensions

If this is right

Any RL-based question recommender can potentially improve by adding the rank-loss term without changing its policy network or training loop.
The collaborative principle is applied to exploration itself: similar learners share exploration, dissimilar learners receive differentiated exploration, which reduces the number of episodes needed to find good questions.
Both representation choices for recommendation differences work, so a practitioner can use the cheap additive encoding (RAR-A) when computational cost matters.
The module transfers to off-the-shelf RL algorithms such as DQN and SAC, indicating it is a general plug-in rather than a component tied to one architecture.
Gains appear at both short interaction horizons (10 steps) and long horizons (200 steps), suggesting the exploration efficiency benefit is not limited to a transient early-training phase.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper implicitly treats the symmetric difference of learning targets as the whole of student similarity; an extension left open is to incorporate the student's current mastery state or knowledge-tracing embeddings into the student-difference metric, which could yield finer alignment and larger gains.
The same ranking-alignment principle could transfer to other recommendation settings where user similarity is available as set differences and item recommendations as probability distributions, for example content or course recommendation, though the paper only evaluates question recommendation.
A stress test that the paper does not run is to vary the simulator's response noise (e.g., guessing and slip probabilities) to see whether the ranking-alignment benefit degrades gracefully; this would indicate whether the mechanism is robust to less predictable learners.
The DQN counterexample reported in the compatibility study suggests the module helps most when the base algorithm can actually learn from the exploration data; if a base policy is too weak to exploit the signal, rank alignment alone may not rescue it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The new thing here is a plug-in exploration loss for RL-based question recommenders, plus evidence that it helps in simulated environments and is compatible with DQN and SAC. The paper is clearly written, the code is released, and Table 1 shows consistent gains for RAR-S and RAR-A across five simulators. That part is real and worth taking seriously.

The problem is that the loss does not do what the paper says it does. The stated principle in Section 1 is that similar students should receive similar recommendations and different students should get different ones. But Eq. 14 is L_r = Σ Clip(ψ d^t_uv − d^p_uv, 0, ω). For pairs with identical learning targets, d^t_uv = 0, so the term is always zero regardless of how different the recommendations are. There is no gradient that ever encourages similar students to receive similar questions. For pairs with different targets, the loss is satisfied once d^p_uv ≥ ψ d^t_uv; it never compares one pair against another, so it enforces at most a diversity lower bound, not a ranking of differences. The stress-test note is right: this is a one-sided hinge, not a ranking alignment. The central claim that the gains come from ranking alignment is unsupported by the loss as written.

The rest of the evaluation has standard soft spots: no error bars or significance tests, hyperparameters chosen from small lists without a clear held-out protocol, and the claim of “any RL-based recommender” rests on only two compatibility experiments. Eq. 15 also looks like a typo—the definition of \hat r is self-referential without a time index, and the policy gradient sum should run over a trajectory with proper discounting. The simulator-fidelity concern is real but not fatal; it is normal for this subfield to evaluate in simulation, and the paper is honest about it.

If the loss were reframed as “margin-based diversity exploration” and the explanation revised accordingly, the empirical results would still be interesting. As written, however, the paper’s main conceptual contribution is misdescribed, and the missing ablation isolating the rank loss makes it hard to tell how much of the gain is actually due to this term.

This deserves a serious referee—the idea is usable and the empirical work is not sloppy—but the authors should be asked to fix the loss description, add an ablation with the rank loss disabled, report variance, and correct the return definition. I would not cite it in its current form, but I would send it to review with a request for major revision.

Referee Report

4 major / 6 minor

Summary. The paper proposes Ranking Alignment Recommendation (RAR), a framework for RL-based personalized question recommendation. The recommendation module can be any RL-based recommender; the ranking alignment module computes pairwise differences between students' learning targets (symmetric difference of target question sets) and between recommendation sequences (L2 distance over sequential or summed probability representations), and adds a rank loss L_r (Eq. 14) to the policy-gradient and knowledge-tracing losses. Experiments on five simulated environments (KSS and DKT/IEKT simulators trained on ASSIST09 and Junyi) report that RAR-S and RAR-A achieve the best or second-best learning effects in Table 1, and that adding the module to DQN and SAC improves performance in most cases in Table 2. The paper argues that the module can be plugged into any RL-based question recommender.

Significance. If the empirical claims held, the paper would offer a simple, portable exploration mechanism for RL-based recommenders, and the released code would be a useful resource for the community. However, the central mechanism as written in Eq. (14) does not implement the stated ranking alignment or the collaborative principle from Section 1, and the empirical evaluation lacks error bars, significance tests, and a clear model-selection protocol. The main contribution is therefore not yet established; the paper requires careful correction of the loss definition and a stronger empirical analysis before its claims can be accepted.

major comments (4)

[Section 3.2, Eq. (14)] Equation (14) does not implement the ranking alignment described in the paper. The loss is L_r = sum Clip(psi d_t_uv - d_p_uv, 0, omega), which is a one-sided lower-bound hinge. For similar students, d_t_uv = 0, so the term becomes Clip(-d_p_uv, 0, omega) = 0 for every recommendation difference, meaning the loss provides no gradient to make similar students' recommendations more similar. For dissimilar students, the loss is zero as soon as d_p_uv >= psi d_t_uv, and it never compares one pair against another, so it does not align the ranking of target differences with the ranking of recommendation differences. At most it enforces a diversity lower bound proportional to the target distance. Since the paper's central claim is that this ranking-alignment loss improves exploration, the attribution of the reported gains to ranking alignment is unsupported as written. Please replace Eq. (14) with a loss that actually compares pairs (for example, a margin or listwise loss over both similarity and dissimilarity constraints) or re-frame the contribution as a diversity regularization mechanism and revise the motivation accordingly.
[Section 3.3, Eq. (15)] The definition of the return in Eq. (15) is self-referential and does not define a cumulative reward: it states that \hat r(s_t, q_t) = r(s_t, q_t) + \gamma \hat r(s_t, q_t), which implies \hat r = r / (1 - \gamma) and has no dependence on future rewards. This is presumably a typographical error, but as written the policy-gradient objective is not well-defined. Please correct the recursion (e.g., \hat r_t = r_t + \gamma \hat r_{t+1}) and specify the bootstrapping or return estimator actually used in the experiments.
[Section 4.3 and 4.4] Tables 1 and 2 report no error bars, confidence intervals, or significance tests, and hyperparameters (alpha, beta, learning rate) are selected from small lists without a described held-out protocol. Given the stochasticity of RL training and the small differences in some entries (e.g., RAR-S 0.616 vs RAR-A 0.614 at t=10 in KSS), the claim that RAR-S and RAR-A achieve the best or second-best performance in all cases is not statistically supported. Please report multiple independent runs with means and standard deviations, and clearly separate model selection from final evaluation.
[Section 4.1 and Conclusion] The evaluation is conducted entirely with simulated student environments (KSS and deep knowledge tracing simulators), and the authors explicitly state that they avoid real student interaction for ethical and practical reasons. However, the abstract and conclusion present the performance gains as general, and the claim that the framework can be applied to any RL-based question recommender is broader than what is tested (only DQN and SAC are considered in Table 2). Please temper the generalization claims, discuss simulator-fidelity threats, and report compatibility results on a wider range of RL recommenders or clearly scope the contribution to simulation settings.

minor comments (6)

[Section 2.1, Eq. (1)] The learning-effect definition in Eq. (1) is undefined when m_sup = m_b because the denominator is zero; please add a convention for this degenerate case.
[Section 3.2, Eq. (13)] The distance function dist(.) is said to be L2 distance, but it is unclear whether the recommendation representations are normalized before computing this distance; please specify this detail.
[Section 3.2] The text does not specify how the set U_u of m peer students is sampled (e.g., uniformly from the batch, or by some similarity criterion), nor whether the symmetric difference in Eq. (10) is over question sets or concept sets; please clarify.
[Section 4.5, Table 2] The compatibility study shows a substantial performance drop for DQN+RAM on IEKTA09 (0.6445 to 0.3736) and on DKTJu (0.0070 to 0.0057), and the text attributes this to DQN's inability to learn suitable questions; this explanation is speculative without additional diagnostics or variance estimates.
[Section 1 and Conclusion] The phrase 'any RL-based question recommender' is an overstatement because only DQN and SAC are tested; please qualify the claim to the tested algorithms.
[Throughout] There are several notation ambiguities, such as f_t used both as a learning target encoder (Eq. 6) and as a subscript for time step, and r_t used both as a reward and as a representation; please standardize the notation for readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ranking-alignment loss is an auxiliary training objective, not a re-statement of the evaluation metric, and the empirical comparisons use independently defined simulators.

full rationale

The paper's derivation chain is not circular. The learning effect in Eq. (1) is defined from mastery before and after recommendation, independently of the proposed rank loss. The ranking-alignment loss in Eq. (14) is an auxiliary training signal added to the policy-gradient and KT losses in Eq. (17); it is never used as the evaluation metric, and no fitted parameter is relabeled as a prediction. The central empirical claim (RAR-S and RAR-A best or second-best in Table 1; compatibility gains in Table 2) is evaluated against externally defined simulators (KSS, DKT, IEKT trained on ASSIST09 and Junyi, Section 4.1). The only author-overlapping citation is the IEKT simulator reference [11], which is used as an external benchmark environment rather than as a load-bearing uniqueness theorem or ansatz; under the hard rules, such independently trainable benchmarks do not raise the circularity score. The critique that Eq. 14 is a one-sided hinge that may not literally implement pairwise ranking alignment is a soundness and effectiveness concern about whether the loss matches its name, not evidence that the result is equivalent to its inputs by construction.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The main unknowns are seven hyperparameters and two domain assumptions (simulator fidelity and collaborative alignment). The method introduces no new physical or mathematical entities beyond the ranking alignment objective itself.

free parameters (7)

alpha (weight on KT loss) = selected from {0, 0.1, 0.5, 1}
Eq. (17); no per-environment value or validation split reported.
beta (weight on rank loss) = selected from {0, 0.1, 0.5, 1}
Eq. (17); this weight directly controls the proposed ranking alignment contribution, yet the exact value for each reported result is not given.
psi (scale of target difference in rank loss)
Eq. (14); no value reported anywhere in the paper.
omega (clip cap in rank loss)
Eq. (14); no value reported.
m (number of peer students sampled per student)
Section 3.2; size of the student set used to build the rank loss is not specified.
learning rate = selected from {1e-3, 5e-4, 1e-4}
Section 4.3; optimizer setting chosen by hand with no reported per-environment values.
gamma (discount factor)
Appears in the return definition around Eq. (15); value is never reported.

assumptions (5)

domain assumption Simulated students faithfully represent real students' learning effects
Section 4.1 evaluates only in KSS and DKT/IEKT-based simulators; the conclusion about real large-scale platforms depends on this fidelity.
domain assumption Students with more similar learning targets should receive more similar recommendation sequences
Core premise of the ranking alignment loss in Eq. (14); no empirical or theoretical support is given for it.
domain assumption Symmetric difference cardinality of learning target sets is a sufficient student difference metric
Eq. (10) reduces student differences to set size differences, ignoring question difficulty, prior mastery, and prerequisite structure.
domain assumption Probability-sequence or sum representations with L2 distance capture recommendation differences
Eqs. (11)-(13); no validation that these embeddings behave like real recommendation differences.
standard math Policy gradient (REINFORCE) with an auxiliary knowledge-tracing loss is a valid optimization framework
Eqs. (15)-(17) rely on standard RL gradients; the return in Eq. (15) is written recursively with no terminal case, indicating a typo.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Personalized Education with Ranking Alignment Recommendation." pith.science (2026). https://pith.science/paper/XZSUNI3R

@misc{pith2026250723664,
  author       = {Pith},
  title        = {Pith review of: Personalized Education with Ranking Alignment Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XZSUNI3R}},
  note         = {Machine review of arXiv:2507.23664}
}

read the original abstract

Personalized question recommendation aims to guide individual students through questions to enhance their mastery of learning targets. Most previous methods model this task as a Markov Decision Process and use reinforcement learning to solve, but they struggle with efficient exploration, failing to identify the best questions for each student during training. To address this, we propose Ranking Alignment Recommendation (RAR), which incorporates collaborative ideas into the exploration mechanism, enabling more efficient exploration within limited training episodes. Experiments show that RAR effectively improves recommendation performance, and our framework can be applied to any RL-based question recommender. Our code is available in https://github.com/wuming29/RAR.git.

Figures

Figures reproduced from arXiv: 2507.23664 by the authors.

**Figure 2.** The pipeline of our method. This problem is typically formalized as a Markov Decision Process (MDP) [10,9], where the state st is generated by the student’s historical records Hu t and learning target Tu. The action at is the question recommended by the recommender based on st. The reward rt represents the reward for taking action at in state st, which is typically determined based on the learning effect ∆u. 3 Meth… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 11 canonical work pages

[1]

Advances in neural information processing systems33, 1877–1901 (2020)

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems33, 1877–1901 (2020)

2020
[2]

Chang, H.S., Hsu, H.J., Chen, K.T.: Modeling exercise relationships in e-learning: A unified approach. In: EDM. pp. 532–535 (2015)

work page 2015
[3]

Set-to-Sequence Ranking-based Concept-aware Learning Path Recommendation

Chen, X., Shen, J., Xia, W., Jin, J., Song, Y., Zhang, W., Liu, W., Zhu, M., Tang, R., Dong, K., et al.: Set-to-sequence ranking-based concept-aware learning path recommendation. arXiv preprint arXiv:2306.04234 (2023)

work page Pith review arXiv 2023
[4]

User modeling and user-adapted inter- action19, 243–266 (2009)

Feng, M., Heffernan, N., Koedinger, K.: Addressing the assessment challenge with an online system that tutors as it assesses. User modeling and user-adapted inter- action19, 243–266 (2009)

work page 2009
[5]

In: International conference on machine learning

Haarnoja, T., Zhou, A., Abbeel, P., Levine, S.: Soft actor-critic: Off-policy maxi- mum entropy deep reinforcement learning with a stochastic actor. In: International conference on machine learning. pp. 1861–1870. PMLR (2018)

work page 2018
[6]

arXiv preprint arXiv:1412.6980 (2014)

Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

arXiv 2014
[7]

arXiv preprint arXiv:2108.00268 (2021)

Kubotani, Y., Fukuhara, Y., Morishima, S.: Rltutor: Reinforcement learning based adaptive tutoring system by modeling virtual student with fewer interactions. arXiv preprint arXiv:2108.00268 (2021)

arXiv 2021
[8]

In: Interna- tional conference on machine learning

Lee, J., Lee, Y., Kim, J., Kosiorek, A., Choi, S., Teh, Y.W.: Set transformer: A framework for attention-based permutation-invariant neural networks. In: Interna- tional conference on machine learning. pp. 3744–3753. PMLR (2019)

work page 2019

Show all 16 references

[9]

In: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management

Li, Q., Xia, W., Yin, L., Shen, J., Rui, R., Zhang, W., Chen, X., Tang, R., Yu, Y.: Graph enhanced hierarchical reinforcement learning for goal-oriented learning path recommendation. In: Proceedings of the 32nd ACM International Conference on Information and Knowledge Manageme...

2023
[10]

In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining

Liu, Q., Tong, S., Liu, C., Zhao, H., Chen, E., Ma, H., Wang, S.: Exploiting cog- nitive structure for adaptive learning. In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. pp. 627–635 (2019)

2019
[11]

In: Proceedings of the 44th Inter- national ACM SIGIR Conference on Research and Development in Information Retrieval

Long, T., Liu, Y., Shen, J., Zhang, W., Yu, Y.: Tracing knowledge state with individual cognition and acquisition estimation. In: Proceedings of the 44th Inter- national ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 173–182 (2021)

2021
[12]

arXiv preprint arXiv:1312.5602 (2013)

Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., Riedmiller, M.: Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 (2013)

2013 arXiv
[13]

Advances in neural information processing systems28(2015)

Piech, C., Bassen, J., Huang, J., Ganguli, S., Sahami, M., Guibas, L.J., Sohl- Dickstein, J.: Deep knowledge tracing. Advances in neural information processing systems28(2015)

2015
[14]

In: International conference on machine learning

Silver, D., Lever, G., Heess, N., Degris, T., Wierstra, D., Riedmiller, M.: Determin- istic policy gradient algorithms. In: International conference on machine learning. pp. 387–395. Pmlr (2014)

2014
[15]

Advances in neural information pro- cessing systems30(2017)

Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017)

2017
[16]

In: Proceedings of the ACM web conference 2022

Zhou, K., Yu, H., Zhao, W.X., Wen, J.R.: Filter-enhanced mlp is all you need for sequential recommendation. In: Proceedings of the ACM web conference 2022. pp. 2388–2399 (2022)

2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Advances in neural information processing systems33, 1877–1901 (2020)

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems33, 1877–1901 (2020)

2020

[2] [2]

Chang, H.S., Hsu, H.J., Chen, K.T.: Modeling exercise relationships in e-learning: A unified approach. In: EDM. pp. 532–535 (2015)

work page 2015

[3] [3]

Set-to-Sequence Ranking-based Concept-aware Learning Path Recommendation

Chen, X., Shen, J., Xia, W., Jin, J., Song, Y., Zhang, W., Liu, W., Zhu, M., Tang, R., Dong, K., et al.: Set-to-sequence ranking-based concept-aware learning path recommendation. arXiv preprint arXiv:2306.04234 (2023)

work page Pith review arXiv 2023

[4] [4]

User modeling and user-adapted inter- action19, 243–266 (2009)

Feng, M., Heffernan, N., Koedinger, K.: Addressing the assessment challenge with an online system that tutors as it assesses. User modeling and user-adapted inter- action19, 243–266 (2009)

work page 2009

[5] [5]

In: International conference on machine learning

Haarnoja, T., Zhou, A., Abbeel, P., Levine, S.: Soft actor-critic: Off-policy maxi- mum entropy deep reinforcement learning with a stochastic actor. In: International conference on machine learning. pp. 1861–1870. PMLR (2018)

work page 2018

[6] [6]

arXiv preprint arXiv:1412.6980 (2014)

Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

arXiv 2014

[7] [7]

arXiv preprint arXiv:2108.00268 (2021)

Kubotani, Y., Fukuhara, Y., Morishima, S.: Rltutor: Reinforcement learning based adaptive tutoring system by modeling virtual student with fewer interactions. arXiv preprint arXiv:2108.00268 (2021)

arXiv 2021

[8] [8]

In: Interna- tional conference on machine learning

Lee, J., Lee, Y., Kim, J., Kosiorek, A., Choi, S., Teh, Y.W.: Set transformer: A framework for attention-based permutation-invariant neural networks. In: Interna- tional conference on machine learning. pp. 3744–3753. PMLR (2019)

work page 2019