Pith. sign in

REVIEW 3 major objections 5 minor 33 references

Improving End-to-End Sequential Recommendations with Intent-aware Diversification

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read The paper claims that an end-to-end sequential recommender with implicit intent mining can raise recommendation diversity on MovieLens while keeping or improving accuracy, beating re-ranking baselines.

desk verdict New idea, shaky evaluation; the split description and the undefined training objective need fixing before the Table 2 claims carry weight. read the letter →

arxiv 1908.10171 v1 pith:3QDXZJBM submitted 2019-08-27 cs.IR

Wanyu Chen , Pengjie Ren , Fei Cai , Maarten de Rijke This is my paper · ORCID
classification cs.IR
keywords sequentialrecommendationdiversityintent-awarediversificationend-to-endlearningimplicitintentminingintra-listdistancerecurrentneuralnetworksMovieLens
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

This paper argues that sequential recommenders do not have to trade accuracy for diversity. It introduces IDSR, an end-to-end model that mines a user's latent intents from their behavior sequence and generates a recommendation list by selecting items that collectively cover those intents, supervised by a loss that rewards lists containing at least one relevant item per intent. On MovieLens 100K and 1M, the authors report that IDSR improves Intra-List Distance by roughly 20–58% over the best baseline while also improving Recall and MRR. If correct, this would mean diversity can be learned during training rather than bolted on by re-ranking.

What carries the argument

The load-bearing piece is the pairing of an Implicit Intent Mining (IIM) module with an Intent-aware Diversity Promoting (IDP) loss. IIM runs $M$ parallel attention heads over the GRU-encoded behavior sequence, each head producing an intent-specific user representation $F_u^i$, and estimates each intent's importance $P(a_i\mid u)$ by how well it matches the global preference. The IDP loss trains the decoder to place the ground-truth item near the top of the list while also maximizing $\sum_i P(a_i\mid u)\bigl(1-\prod_{v\in RL}(1-P(v\mid a_i))\bigr)$, the probability that every mined intent has at least one relevant item in the list. At decoding time, the score of each candidate item combines its relevance with an intent-satisfaction term $W(R_{t-1},a_i)$ that tracks which intents the list built so far has not yet covered, so diversity enters the model at training time and at generation time rather than in a separate re-ranking step.

What would settle it

Retrieve the timestamp-sorted interactions, split each user's sequence by time into 70% training, 10% validation, and 20% test, then retrain IDSR and the HRNN+MMR baseline; if IDSR's ILD and Recall gains over HRNN+MMR mostly disappear, the paper's central claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that an end-to-end sequential recommender can be explicitly trained to produce lists that are both accurate and diverse, and that doing so beats the two-stage alternative of training an accuracy-only model and then re-ranking with MMR. The paper's model, IDSR, assumes each behavior sequence reflects M latent intents; it uses a GRU encoder, a multi-head attention module called the Implicit Intent Mining (IIM) module to estimate intent importance, and a decoding step that greedily picks items using a score mixing relevance with an intent-satisfaction term that tracks which intents already-selected items cover. The Intent-aware Diversity Promoting (IDP) loss trains the whole system to place the ground-truth item high while maximizing the probability that every mined intent has at least one relevant item in the list. On ML100K and ML1M, the authors report higher Recall, MRR, and ILD than the baselines, including MMR-based post-processing.

Load-bearing premise

The reported gains rest on the assumption that the train/test split is leakage-free, but the paper describes the split as both random (70/10/20) and temporal (test contains the most recent behaviors), and those two descriptions cannot both be true.

Editorial extensions

If this is right

  • On both ML100K and ML1M, IDSR reports higher Recall and MRR than the strongest accuracy-only baseline (HRNN), so the diversity gains do not come from sacrificing accuracy.
  • The diversity gains are large: ILD improves by about 19.9–20.3% over HRNN on ML100K and 50.9–57.8% on ML1M, and by 7.4–11.8% over the best MMR re-ranking baseline.
  • Using a GRU encoder inside IDSR outperforms a Transformer encoder on these datasets, suggesting that the recurrent inductive bias is valuable when training data is limited.
  • A single trade-off parameter $\lambda$ controls the accuracy–diversity balance in both the decoder and the loss; best accuracy occurs around $\lambda=0.5$ on ML100K and $\lambda=0.8$ on ML1M, while best diversity occurs at lower $\lambda$.

Reading between the lines

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

  • A natural extension the paper leaves implicit is applying IDSR to shared-account recommendation, where the behavior sequence mixes several users' intents; the model's latent intents could in principle separate them without any user identifier.
  • Because the diversity signal in the IDP loss is computed from item representations rather than hand-labeled categories, the same architecture could in principle diversify news, music, or e-commerce lists by redefining item distance, and could even target serendipity if the embedding space encodes novelty.
  • If the leakage-free split is verified, the comparison against MMR suggests a testable hypothesis: optimizing a coverage-style loss directly over mined intents generalizes better than a greedy pairwise-dissimilarity heuristic, because the loss matches the evaluation objective more closely.
  • The IDP loss is not architecture-specific; it could be dropped into non-sequential recommenders with candidate intent vectors, though the paper only demonstrates it in the sequential setting.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes IDSR, an end-to-end neural model for sequential recommendation that jointly optimizes recommendation accuracy and diversity. The model uses a GRU sequence encoder, an Implicit Intent Mining (IIM) module with multi-head attention over latent intents, and an intent-aware decoder that greedily builds a recommendation list by balancing a relevance score and a diversity score. Training is driven by an Intent-aware Diversity Promoting (IDP) loss that combines a relevance term and a diversity term. Experiments on MovieLens 100K and 1M compare IDSR against POP, FPMC, GRU4Rec, HRNN, and MMR-based post-processing variants, reporting Recall, MRR, and ILD. The authors report consistent gains in both accuracy and diversity, plus an encoder ablation and a sensitivity analysis for the trade-off parameter lambda.

Significance. If the results hold, the paper makes a useful contribution by incorporating diversity into the training objective of a sequential recommender rather than as a post-processing step, and by mining latent intents without explicit intent labels. The paper includes a reasonable set of baselines, an ablation over sequence encoders, and a parameter sensitivity study. The main claimed strengths are the end-to-end nature of the method and the consistent ILD improvements. However, the significance depends on resolving two load-bearing technical issues: the exact data-split protocol and the differentiability/training procedure of the discrete decoder. The paper is not internally inconsistent in its central derivation, but the missing details prevent the reader from verifying the claims.

major comments (3)
  1. [Section 4.1 (Datasets)] The split description is internally contradictory: the text states that the datasets are 'randomly divide[d]' into training (70%), validation (10%), and test (20%) sets, and then states that the test set contains the most recent behaviors, which happened later than those in the training and validation sets. A random split over samples cannot guarantee this temporal ordering, and a temporal split is not a random split. If a random split was applied, future interactions can appear in training while earlier interactions appear in test, which is leakage that sequential models are especially prone to exploit; if a temporal split was used, the word 'randomly' is wrong. Because all of the headline results in Table 2 depend on this split, please specify the exact split procedure (e.g., split by user, global timestamp cutoff, or per-user temporal split) and confirm that no test interaction has a timestamp earlier than any training interaction.
  2. [§3.4–3.5, Eqs. (2) and (11)] The training objective is not fully specified. In the IDP decoder, the next item is selected by the discrete operation v_t <- argmax_{v} S(v) (Eq. 2), but the relevance loss L_rel^RL in Eq. (11) depends on p*_t, 'the probability of the ground-truth item y* at t-th step,' which is never defined as a function of the model parameters. It is also not explained how gradients propagate through the argmax to the encoder, IIM, and decoder parameters. If p*_t is the softmax-normalized score over all items at step t, that distribution should be written out explicitly; if the model uses teacher forcing or a soft relaxation during training, that should be stated. As written, the claim that 'all parameters ... can be learned in an end-to-end back-propagation training paradigm' is not substantiated.
  3. [§3.4, Eq. (9)] The definition of W(R_{t-1}, a_i) has an index inconsistency: the numerator uses w_i^{t-1} while the denominator sums exp(w_j^t), and w_j^t is not defined anywhere. Since W appears in the item score S(v) in Eq. (3), this makes the method ambiguous. The denominator should presumably be Σ_j exp(w_j^{t-1}); please correct and confirm.
minor comments (5)
  1. [Section 4.1] The text says ML1M 'contains 4,607,047 ratings for movies,' but Table 1 reports 1,000,209 interactions for ML1M; the larger number appears to be an error and should be corrected.
  2. [Section 5.4] The sentence 'IDSR increases when λ changes from 0 to 0.1 and then decreases from 0.4 (0.2) to 1' is ambiguous; clarify which dataset corresponds to 0.4 and which to 0.2.
  3. [Figure 2] Figure 2 is difficult to read in the version provided; the labels inside the diagram (e.g., 'diversityrelevance' and 'softmax') are not legible, which hampers understanding of the architecture.
  4. [Section 4.4] The code link is given as a suppressed placeholder; an anonymous working link, or an explicit statement that code will be made available upon publication, is needed to support reproducibility.
  5. [Section 5.2] The paired t-test results are reported as significant, but the text does not state how many random seeds or evaluation runs were used; please add this information to support the statistical claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: IDSR's diversity gains are measured by a genre-based ILD metric that is not equivalent to the training loss, and all load-bearing components of the derivation are self-contained.

full rationale

The claimed derivation is an empirical training-and-evaluation pipeline, not a derivation by construction. The IDP loss (Eq. 10) combines a next-item relevance loss L_rel (Eq. 11) with a coverage-based diversity loss L_div, while the reported diversity metric ILD (Eq. 12) is the average pairwise genre-vector distance of the generated list; ILD is not a term in the loss and is computed from genre metadata that the model is not directly trained to optimize. Neither the relevance nor the diversity improvement is therefore forced by the objective function. The latent intents mined by IIM are supervised only by the IDP loss, which is a self-referential training signal for a latent representation, but the paper's claims are about recommendation accuracy and diversity on held-out data, not about the ontological reality of the intents, so this is not circularity. The only self-citations (Ren et al. 2019, Li et al. 2017) appear in related-work context and are not load-bearing for the model's design or evaluation. The contradictory dataset-split description in Section 4.1 and the suppressed code URL are correctness and reproducibility risks, not circular reductions; they do not raise the circularity score.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central empirical claim rests on the latent-intent decomposition, a greedy generation approximation, and the choice of genre-based ILD as the diversity ground truth. The model also depends on unreported hyperparameters (lambda, M) and a contradictory data-split description, so the ledger is not clean.

free parameters (4)
  • lambda (relevance-diversity trade-off) = not stated for final results; Fig. 3 scans 0 to 1
    Used in Eq. (3) and Eq. (10) to balance accuracy and diversity; the value used to produce Table 2 is not reported, and optimal lambda differs by metric.
  • M (number of latent intents / attention heads) = not stated
    Defines the number of intents in Eq. (1) and attention heads in IIM (Eq. 5); never specified in Section 4.4, and it controls model capacity and diversity behavior.
  • theta for MMR baselines = 0.5
    Set by hand in Section 4.2 to balance relevance and dissimilarity; no tuning or sensitivity analysis, so MMR baselines may be understated.
  • input sequence length (window size) = 9
    Each sample uses the former 9 behaviors as input (Section 4.1); no ablation or justification for this length.
assumptions (5)
  • domain assumption There are exactly M latent intents behind each user behavior sequence (Section 3.1).
    The model assumes a fixed, discrete set of intents with which every behavior can be associated; the text gives no external evidence for this structure.
  • domain assumption User satisfaction factorizes as a mixture over intents: P(RL|u,S_u)=sum_i P(a_i|u)P(RL|a_i,u,S_u) (Eq. 1).
    This decomposition assumes the intents are exhaustive and conditionally independent given the user; no empirical check is provided.
  • domain assumption Greedy item-by-item selection (Eq. 2) is a sufficient approximation to optimal list generation.
    The paper optimizes greedily without bounding the gap to the optimal diversified list; standard practice but an unstated approximation.
  • domain assumption The IDP diversity loss L_div is a valid training proxy for the ILD metric (Section 3.5 vs Section 4.3).
    L_div optimizes expected intent coverage while ILD measures pairwise genre distance; the alignment between them is not demonstrated.
  • domain assumption Item genre vectors are a meaningful dissimilarity measure for diversity (Section 4.3).
    ILD is computed on Euclidean distance between genre vectors; this assumes genre categories capture the diversity that matters to users.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving End-to-End Sequential Recommendations with Intent-aware Diversification." pith.science (2026). https://pith.science/paper/3QDXZJBM

@misc{pith2026190810171,
  author       = {Pith},
  title        = {Pith review of: Improving End-to-End Sequential Recommendations with Intent-aware Diversification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3QDXZJBM}},
  note         = {Machine review of arXiv:1908.10171}
}
read the original abstract

Sequential Recommendation (SRs) that capture users' dynamic intents by modeling user sequential behaviors can recommend closely accurate products to users. Previous work on SRs is mostly focused on optimizing the recommendation accuracy, often ignoring the recommendation diversity, even though it is an important criterion for evaluating the recommendation performance. Most existing methods for improving the diversity of recommendations are not ideally applicable for SRs because they assume that user intents are static and rely on post-processing the list of recommendations to promote diversity. We consider both recommendation accuracy and diversity for SRs by proposing an end-to-end neural model, called Intent-aware Diversified Sequential Recommendation (IDSR). Specifically, we introduce an Implicit Intent Mining module (IIM) into SRs to capture different user intents reflected in user behavior sequences. Then, we design an Intent-aware Diversity Promoting (IDP) loss to supervise the learning of the IIM module and force the model to take recommendation diversity into consideration during training. Extensive experiments on two benchmark datasets show that IDSR significantly outperforms state-of-the-art methods in terms of recommendation diversity while yielding comparable or superior recommendation accuracy.

Figures

Figures reproduced from arXiv: 1908.10171 by the authors.

Figure 1
Figure 1. A example showing sequential recommenda [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of IDSR. Blue, purple and green colors denote different user intents. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Performance of IDSR on two datasets with the parameter [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: An example of recommendation results gener [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 28 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

  2. [2]

    S.; Farooq, U.; Khan, S

    Abid, A.; Hussain, N.; Abid, K.; Ahmad, F.; Farooq, M. S.; Farooq, U.; Khan, S. A.; Khan, Y. D.; Naeem, M. A.; and Sabir, N. 2016. A survey on search results diversification techniques. Neural Computing and Applications 27(5):1207--1229

    work page 2016

  3. [3]

    Adomavicius, G., and Tuzhilin, A. 2005. Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE Transactions on Knowledge and Data Engineering. 17(6):734--749

    work page 2005

  4. [4]

    Agrawal, R.; Gollapudi, S.; Halverson, A.; and Ieong, S. 2009. Diversifying search results. In WSDM '09 , 5--14. New York, NY, USA: ACM

    work page 2009

  5. [5]

    Ashkan, A.; Kveton, B.; Berkovsky, S.; and Wen, Z. 2015. Optimal greedy diversity for recommendation. In IJCAI'15 , 1742--1748. AAAI Press

    work page 2015

  6. [6]

    Carbonell, J., and Goldstein, J. 1998. The use of mmr, diversity-based reranking for reordering documents and producing summaries. In SIGIR '98 , 335--336. New York, NY, USA: ACM

    work page 1998

  7. [7]

    Chen, L.; Zhang, G.; and Zhou, H. 2018. Fast greedy map inference for determinantal point process to improve recommendation diversity. In NIPS'18 , 5627--5638. USA: Curran Associates Inc

    work page 2018

  8. [8]

    Cheng, P.; Wang, S.; Ma, J.; Sun, J.; and Xiong, H. 2017. Learning to recommend accurate and diverse items. In WWW '17 , 183--192

    work page 2017

Show all 33 references
  1. [9]

    Glorot, X., and Bengio, Y. 2010. Understanding the difficulty of training deep feedforward neural networks. In AI&Statistics '10 , 249--256. Chia Laguna Resort, Sardinia, Italy: PMLR

    2010

  2. [10]

    He, R., and McAuley, J. 2016. Fusing similarity models with markov chains for sparse sequential recommendation. In International Conference on Data Mining , 191--200. IEEE

    2016

  3. [11]

    He, X.; He, Z.; Song, J.; Liu, Z.; Jiang, Y.; and Chua, T. 2018. Nais: Neural attentive item similarity model for recommendation. IEEE Transactions on Knowledge and Data Engineering 30(12):2354--2366

    2018

  4. [12]

    Hidasi, B.; Karatzoglou, A.; Baltrunas, L.; and Tikk, D. 2016a. Session-based recommendations with recurrent neural networks. In ICLR '16

  5. [13]

    Hidasi, B.; Karatzoglou, A.; Baltrunas, L.; and Tikk, D. 2016b. Session-based recommendations with recurrent neural networks. In ICLR'16 , 1--10

  6. [14]

    Kang, W., and McAuley, J. J. 2018. Self-attentive sequential recommendation. In ICDM '18 , 197--206

    2018

  7. [15]

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

    2014 arXiv

  8. [16]

    Kulesza, A., and Taskar, B. 2012. Determinantal Point Processes for Machine Learning . Hanover, MA, USA: Now Publishers Inc

    2012

  9. [17]

    Kunaver, M., and Porl, T. 2017. Diversity in recommender systems a survey. Know.-Based Syst. 123(C):154--162

    2017

  10. [18]

    Li, J.; Ren, P.; Chen, Z.; Ren, Z.; Lian, T.; and Ma, J. 2017. Neural attentive session-based recommendation. In CIKM '17 , 1419--1428. New York, NY, USA: ACM

    2017

  11. [19]

    Liu, Q.; Zeng, Y.; Mokhosi, R.; and Zhang, H. 2018. Stamp: Short-term attention/memory priority model for session-based recommendation. In KDD '18 , 1831--1839. New York, NY, USA: ACM

    2018

  12. [20]

    Quadrana, M.; Karatzoglou, A.; Hidasi, B.; and Cremonesi, P. 2017. Personalizing session-based recommendations with hierarchical recurrent neural networks. In RecSys '17 , 130--137. New York, NY, USA: ACM

    2017

  13. [21]

    Quadrana, M.; Cremonesi, P.; and Jannach, D. 2018. Sequence-aware recommender systems. ACM Computing Surveys 51(4):66:1--66:36

    2018

  14. [22]

    Rakkappan, L., and Rajan, V. 2019. Context-aware sequential recommendations with stacked recurrent neural networks. In The Web Conference , 3172--3178

    2019

  15. [23]

    Ren, P.; Chen, Z.; Li, J.; Ren, Z.; Ma, J.; and de Rijke, M. 2019. Repeatnet: A repeat aware neural recommendation machine for session-based recommendation. In AAAI '19 . AAAI

    2019

  16. [24]

    Rendle, S.; Freudenthaler, C.; and Schmidt-Thieme, L. 2010. Factorizing personalized markov chains for next-basket recommendation. In WWW '10 , 811--820. New York, NY, USA: ACM

    2010

  17. [25]

    Sha, C.; Wu, X.; and Niu, J. 2016. A framework for recommending relevant and diverse items. In IJCAI'16 , 3868--3874. AAAI Press

    2016

  18. [26]

    Sun, F.; Liu, J.; Wu, J.; Pei, C.; Lin, X.; Ou, W.; and Jiang, P. 2019. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. CoRR abs/1904.06690

    2019 arXiv

  19. [27]

    N.; Kaiser, L.; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017a. Attention is all you need. CoRR abs/1706.03762

    arXiv

  20. [28]

    u.; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, Aidan Nand Kaiser, L. u.; and Polosukhin, I. 2017b. Attention is all you need. In Advances in Neural Information Processing Systems 30 , 5998--6008. Curran Associates, Inc

  21. [29]

    Wu, Q.; Liu, Y.; Miao, C.; Zhao, Y.; Guan, L.; and Tang, H. 2019. Recent advances in diversified recommendation. CoRR abs/1905.06589

    2019 arXiv

  22. [30]

    S.; Cui, Z.; Zhou, X.; and Xiong, H

    Xu, C.; Zhao, P.; Liu, Y.; Xu, J.; S.Sheng, V. S.; Cui, Z.; Zhou, X.; and Xiong, H. 2019. Recurrent convolutional neural network for sequential recommendation. In WWW '19 , 3398--3404. New York, NY, USA: ACM

    2019

  23. [31]

    Zhang, M., and Hurley, N. 2008. Avoiding monotony: Improving the diversity of recommendation lists. In RecSys '08 , 123--130. New York, NY, USA: ACM

    2008

  24. [32]

    Zhang, S.; Yao, L.; and Sun, A. 2017. Deep learning based recommender system: A survey and new perspectives. arXiv preprint arXiv:1707.07435

    2017 arXiv

  25. [33]

    M.; and Meek, C

    Zimdars, A.; Chickering, D. M.; and Meek, C. 2001. Using temporal data for making recommendations. In UAI '01 , 580--588. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc

    2001

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.