Simultaneous Relevance and Diversity: A New Recommendation Inference Approach

Qiyuan An; Trevor Hastie; Yang Yi; Yanzhi Wang; Yifang Liu; Zhentao Xu

REVIEW 4 major objections 4 minor 46 references

Simultaneous Relevance and Diversity: A New Recommendation Inference Approach

T0 review · 4 major / 4 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read A recommender that learns from both liked and skipped items can raise relevance and diversity in one model, dissolving the standard exploitation-exploration trade-off.

desk verdict An elegant negative-channel extension of CF with impressive but mechanism-unconfirmed gains; needs an ablation that uses impressions without the n2p structure. read the letter →

arxiv 2009.12969 v1 pith:47MYP7OZ submitted 2020-09-27 cs.LG cs.IRcs.SI

Yifang Liu , Zhentao Xu , Qiyuan An , Yang Yi , Yanzhi Wang , Trevor Hastie This is my paper · ORCID

classification cs.LGcs.IRcs.SI

keywords heterogeneousinferencenegative-to-positivedivergentrelevancecollaborativefilteringrecommendationdiversityfeedback-differentiatingencodingimplicitfeedbackcandidategeneration

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 sets out to show that recommendation relevance and diversity do not have to be competing objectives. It introduces heterogeneous inference (HI), an extension of collaborative filtering that adds a negative-to-positive inference channel: from items a user skipped or disliked, the model learns which other items the user is likely to like. The two channels share one model, so diversity emerges from the same relevance inference rather than from a separate re-ranking stage. On MovieLens and real-world production data, the authors report that HI beats conventional CF and diversity-aware baselines on relevance and diversity simultaneously. If correct, this removes a common trade-off in recommender design.

What carries the argument

The machinery is feedback-differentiating encoding (FEEDE), built on the observation matrix $O$ and the positive-feedback matrix $X$: the negative-feedback matrix $Y=O-X$ marks observed non-engagement as explicit negative feedback. HI then runs collaborative filtering on two similarity channels: p2p uses $X^T X$ (items co-liked by the same users), and n2p uses $Y^T X$ (items disliked versus items liked by the same users), each smoothed by rank-$k$ matrix factorization. The two channels can be mixed during candidate generation with a p2p:n2p ratio, embedded as separate sub-networks in a deep neural ranking model with an interaction matrix $W$, or concatenated as $H=[X^T X;\,Y^T X]$ for realtime recommendation. The mechanism carrying the argument is that n2p similarity does not concentrate around previously liked topics, because a negative leaves many possible positives open, so diversity is a by-product of the relevance inference itself.

What would settle it

Run a head-to-head test on a dataset with full impression logs: HI versus a conventional CF that also receives $O$ (by treating observed non-engagement as training negatives) plus an explicit diversity re-ranking stage. If the latter matches or beats HI on held-out AUC/mAP and list diversity, or if HI's edge disappears when impressions are only sampled randomly, the central claim that n2p makes diversity inherent rather than a separate objective is falsified.

Watch

Extended reading notes

Core claim

The central claim is that relevance and diversity can be two inherent outcomes of a single relevance inference process, which the paper calls divergent relevance (DR). Conventional CF is characterized as positive-to-positive (p2p) inference: item similarity is derived from users who liked both items, and this converges toward a narrowing topic. HI adds negative-to-positive (n2p) inference: the fact that a user disliked an item carries signal about which other items the same user may like. By encoding negative feedback as $Y=O-X$ (observed engagement minus positive engagement) and forming cross-similarity matrices such as $Y^T X$, the model spreads predicted relevance across items the user has not yet engaged with. The paper argues that n2p's many-possible-positives-per-negative structure is what makes diversity intrinsic, and reports experiments where HI outperforms p2p-only CF and re-ranking baselines on both relevance metrics and diversity.

Load-bearing premise

The method assumes the recommender knows which items each user actually saw (the observation matrix $O$), so that a skipped item can be tagged as a true negative; if impression logs are unavailable, the negative-to-positive channel cannot be formed.

Editorial extensions

If this is right

A single HI model can replace the common two-stage pipeline of relevance ranking plus diversity re-ranking, reducing recommendation latency while improving both metrics.
HI's benefit grows for new users: in the batch test, the AUC advantage over baselines is larger for day-1-to-7 users than for all users, suggesting negative feedback helps cold start.
Because n2p spreads relevance across less-similar items, recommendation lists cover more of the item space, showing up as higher recall at fixed precision and higher diversity without a relevance penalty.
HI can be deployed at different sophistication levels, from standalone matrix factorization for realtime recommendation to embedding modules inside a deep neural network for ranking, without changing the underlying inference logic.

Reading between the lines

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

If impression logs are not already recorded, a practical entry path is to approximate $O$ from UI signals such as feed position, scroll depth, or dwell time; the fidelity of that approximation is a testable limit of the method's reach.
The p2p:n2p mixture ratio (67:33 in the paper) is a knob the paper does not fully explore; tuning it per user or per item category could produce smooth control over the convergence-versus-spread behavior.
The divergent-relevance mechanism may also weaken the echo-chamber feedback loop over long horizons, because n2p keeps feeding varied items; the paper's experiments are short-term, so this remains an open extension.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

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

Short version: this is a simple and plausible idea—treat observed non-engagement as a separate negative channel and let it interact with positive feedback—and the reported gains are large enough to matter. But the experiments never isolate the new channel from the extra information, so the headline claim is not actually tested.

What's new: FEEDE, Y=O−X, is a clean way to encode the negative/unknown distinction, and the n2p similarity matrix Y^T X is a natural counterpart to p2p X^T X. The paper gives three instantiations (candidate generation, DNN ranking, realtime MF) and evaluates on timestamp-disjoint splits plus a production A/B test, where HI-NN shows a 32% lift on long-watch. It also states the key assumption—impressions must be observable—which is more honest than most industry papers. Related-work coverage is reasonable; previous negative-feedback methods are cited, and the novelty claim centers on diversity, which is fair.

The soft spots are real. The main one is experimental confound: every HI variant trains on X and Y=O−X, while every baseline trains on X alone. The paper's own Analysis 1 says FEEDE encodes more information, I(x~;O|X)>0. So the gains could come from having the impression matrix at all, not from the n2p factorization. A baseline that uses O in a conventional way—say, weighting observed negatives as zero feedback in the same MF—would settle this. Without it, the paper cannot support "divergent relevance" as the mechanism. Analysis 1 is also close to a tautology, and the divergence story is illustrated, not derived. The experiments report single numbers with no variance or significance, hyperparameters are incomplete, and no code or data is released. The production dataset is downsampled and described only vaguely. The diversity metric uses similarity from conventional CF, which is a reasonable external measure, but it still relies on p2p geometry.

I don't think the paper is broken; the empirical direction is plausible and the production A/B result is valuable even with an unclear mechanism. But the missing ablation and missing uncertainty quantification mean it should be read as a promising industry report, not a validated inference framework. A serious referee could fix it by demanding the O-using baseline.

Who this is for: recsys practitioners and researchers working on exposure bias or diversity; less for pure theory readers. I'd bring it to a reading group. I'd accept it for peer review—the empirical claim deserves referee time—but I'd flag the confound as a major revision condition.

Referee Report

4 major / 4 minor

Summary. The paper proposes Heterogeneous Inference (HI), an extension of collaborative filtering that combines positive-to-positive (p2p) inference with a new negative-to-positive (n2p) inference channel, enabled by a feedback-differentiating encoding (FEEDE) in which negative feedback is encoded as Y = O - X. The authors claim that n2p inference makes diversity an inherent outcome of relevance inference ('divergent relevance'), allowing relevance and diversity to improve simultaneously rather than trade off. The method is instantiated in three forms—realtime MF (HI-RT), neural ranking (HI-NN), and candidate generation—and evaluated on MovieLens and a downsampled production dataset against CF baselines with and without diversity re-ranking.

Significance. If the central claim were established, the paper would make a valuable conceptual contribution: diversity emerging from the inference process itself, rather than from a separate re-ranking stage, is a practically appealing direction for recommender systems, and the reported production A/B lift (32.05% on long-watch) is noteworthy. The timestamp-disjoint train/test splits and the use of an external CF-based similarity matrix for the diversity metric are also strengths. However, the paper's current evidence does not isolate the proposed n2p mechanism from the mere availability of extra impression information O, and the theoretical analysis is largely definitional. The significance of the paper therefore rests on whether the authors can supply controlled experiments and stronger argumentation.

major comments (4)

[Experiments; Eq. (6)] The experimental comparison is confounded: every HI variant (HI-RT, HI-NN) is trained on both X and Y = O - X, while every baseline (CF-RT, CF-MF, CF-NN, CF-DA, CF-DM) is trained only on X. Analysis 1 explicitly establishes that FEEDE carries additional information, I(x~; O|X) > 0. Consequently, the observed gains in AUC, mAP, and diversity in Tables 1-3 could be caused simply by the extra impression matrix O rather than by the n2p inference structure. A controlled baseline that uses O in a conventional way (e.g., a CF model trained on X and Y as separate positive/negative label sets, or a model that receives O as auxiliary input) is needed to attribute the improvement to the n2p mechanism.
[Analysis 1; Proof 1] The claimed information-theoretic proof is definitional rather than substantive. The statement I(x~; O|X) > 0 when x~ and O are not independent given X is a restatement of the definition of conditional mutual information; it does not show that FEEDE yields better engagement prediction, nor does it explain why n2p inference produces 'divergent relevance.' The proof therefore does not support the central claim that diversity is an inherent outcome of the relevance inference process. The authors should either derive a testable property of n2p inference (e.g., a condition under which the recommended-item distribution is flatter) or rely on experiments specifically designed to test that property.
[Results and observations; Tables 1-3] The experimental results are reported as single point estimates with no variance, no number of repeated runs, and no significance tests. The text states that HI has a 'significant advantage' over baselines, but with one split for MovieLens and one downsampled production set, this claim is unsupported. The paper should provide confidence intervals or significance tests, and should describe how many random seeds or data samples were used. This is especially important because the differences in diversity metrics in Table 3 (e.g., 0.226 vs. 0.266) may be small relative to user-level variability.
[Feedback-Differentiating Encoding (FEEDE)] The method requires that the impression matrix O is observable, as the paper acknowledges: 'FEEDE relies on an assumption that the impression of an item on a user is observable, i.e., O is known.' Many implicit feedback datasets do not record impressions, so Y = O - X cannot be formed and the n2p channel is undefined. The paper should state this applicability limitation prominently and discuss which real-world systems actually log impressions, since the general framing of the introduction suggests broader applicability than the method can deliver.

minor comments (4)

[Abstract; Introduction] There are several typos and grammatical slips, including 'necessities' (should be 'necessitates') in the abstract and 'simultanuously' in the Results section; a careful proofreading pass is needed.
[Evaluation metrics; Eq. (5)] The diversity metric uses the CF similarity matrix from Eq. (5), which is a reasonable choice for comparability, but the paper should clarify that this is an external similarity measure and discuss whether the conclusions are sensitive to the choice of similarity matrix.
[Algorithms and tests] Several hyperparameters are stated without sensitivity analysis: the p2p:n2p ratio (67:33), the re-ranking weight phi in CF-DA, the loss weights alpha, gamma, lambda, and beta in Eq. (10). A brief sensitivity study or a statement that results are stable across reasonable choices would strengthen the empirical claims.
[Figure 3] The paper references Figure 3 for user-level diversity distributions, but the figure as provided appears to be a placeholder. The authors should ensure the final version includes the actual charts with labeled axes and legends.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's empirical claims are not forced by its definitions or by self-citation; the main weakness is an experimental confound, not a circular derivation.

full rationale

The paper's derivation chain is self-contained rather than circular. HI is defined as an extension of CF with an additional n2p channel built on Y = O - X, but that is a modeling choice, not a restatement of the relevance/diversity outcomes. Diversity is not optimized by the HI objective (Eq. 10); it is measured externally using the conventional CF similarity matrix *X defined in Eq. 5, so the reported diversity gains are not equal by construction to the model's own scoring rule. Both Test-RT and Test-BT use timestamp-disjoint training/test splits, so the AUC, precision/recall, and production long-watch results are genuine predictions on future or held-out engagement rather than refits. The only formal result, Analysis 1, shows that FEEDE carries more conditioning information than CONFE via I(x~; O|X) > 0; this is a definitional consequence of Y = O - X and is not used as the proof of the observed empirical gains. A real weakness is experimental: no baseline consumes O without the n2p structure, so the contribution of the n2p mechanism versus the extra impression information O is not isolated. That is a confound/attribution problem, not a circular derivation. The sole overlapping-author citation (Hastie et al. 2015, on ALS) is a standard related-work reference and is not load-bearing.

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

The method rests on observable impressions and a domain assumption about negative feedback being predictive. Several tuning parameters (alpha, gamma, lambda, beta, ratio, threshold) are unreported, so the reader cannot determine how much of the claimed gain depends on free parameters chosen by the authors.

free parameters (5)

rank k = 10
Target rank of all matrix factorization models; set to 10 uniformly across algorithms.
loss weights alpha, gamma, lambda = not reported
Weights in Equation 10 for the n2p loss, interaction loss, and regularization; values are not given.
p2p:n2p mixing ratio = 67:33
Preset ratio for HI candidate generation; the paper states this ratio but does not justify it.
beta = not reported
Weight on Y in Equation 17 for the realtime HI variant; value is not reported.
feedback threshold = not reported
Threshold converting ratings or long-watch behavior into binary positive or negative feedback; no threshold is specified.

assumptions (4)

domain assumption O is observable, meaning impressions are known.
FEEDE defines Y = O - X; without O, negative feedback cannot be distinguished from unknown. The paper states this in Our Approach.
domain assumption Negative feedback implies positive correlation with other items.
n2p inference assumes users not interested in one item tend to be interested in others; stated in the Introduction.
domain assumption I(x~; O|X) is greater than 0.
Analysis 1 assumes conditioning on O reduces uncertainty about engagement, claimed to be usually true in feedback-based problems.
standard math Low-rank matrix factorization approximates item-item similarity matrices.
Standard CF practice used in Steps 1 and 2 of candidate generation; assumed throughout the MF-based algorithms.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Simultaneous Relevance and Diversity: A New Recommendation Inference Approach." pith.science (2026). https://pith.science/paper/47MYP7OZ

@misc{pith2026200912969,
  author       = {Pith},
  title        = {Pith review of: Simultaneous Relevance and Diversity: A New Recommendation Inference Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/47MYP7OZ}},
  note         = {Machine review of arXiv:2009.12969}
}

read the original abstract

Relevance and diversity are both important to the success of recommender systems, as they help users to discover from a large pool of items a compact set of candidates that are not only interesting but exploratory as well. The challenge is that relevance and diversity usually act as two competing objectives in conventional recommender systems, which necessities the classic trade-off between exploitation and exploration. Traditionally, higher diversity often means sacrifice on relevance and vice versa. We propose a new approach, heterogeneous inference, which extends the general collaborative filtering (CF) by introducing a new way of CF inference, negative-to-positive. Heterogeneous inference achieves divergent relevance, where relevance and diversity support each other as two collaborating objectives in one recommendation model, and where recommendation diversity is an inherent outcome of the relevance inference process. Benefiting from its succinctness and flexibility, our approach is applicable to a wide range of recommendation scenarios/use-cases at various sophistication levels. Our analysis and experiments on public datasets and real-world production data show that our approach outperforms existing methods on relevance and diversity simultaneously.

Figures

Figures reproduced from arXiv: 2009.12969 by the authors.

**Figure 1.** Distribution of relevance over items. (a) p2p convergent relevance: precise and narrow relevance prediction. (b) n2p [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Model architecture. Input vectors pass through embedding modules ( [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** User-level diversity distribution over all users. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 44 canonical work pages

[1]

, " * write output.state after.block = add.period write newline

ENTRY address author booktitle chapter doi edition editor eid howpublished institution isbn issn journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Adomavicius, G.; and Kwon, Y. 2011. Maximizing Aggregate Recommendation Diversity: A Graph-Theoretic Approach. In DiveRS

work page 2011
[4]

Adomavicius, G.; and Kwon, Y. 2012. Improving Aggregate Recommendation Diversity Using Ranking-Based Techniques. In IEEE Transactions on Knowledge and Data Engineering, 896--911

work page 2012
[5]

Antikacioglu, A.; and Ravi, R. 2017. Post processing recommender systems for diversity. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 707--716

work page 2017
[6]

Bradley, K.; and Smyth, B. 2001. Improving recommendation diversity. In Proceedings of the Twelfth Irish Conference on Artificial Intelligence and Cognitive Science, Maynooth, Ireland, 85--94. Citeseer

work page 2001
[7]

Castagnos, S.; Brun, A.; and Boyer, A. 2013. When Diversity Is Needed... But Not Expected!

work page 2013
[8]

Cheng, H.-T.; Koc, L.; Harmsen, J.; Shaked, T.; Chandra, T.; Aradhye, H.; Anderson, G.; Corrado, G.; Chai, W.; Ispir, M.; et al. 2016. Wide & deep learning for recommender systems. In ACM Recsys, 7--10

work page 2016

Show all 46 references

[9]

Covington, P.; Adams, J.; and Sargin, E. 2016. Deep Neural Networks for YouTube Recommendations. In ACM conference on recommender systems, 191--198

2016
[10]

V.; Gargi, U.; Gupta, S.; He, Y.; Lambert, M.; Livingston, B.; and Sampath, D

Davidson, J.; Liebald, B.; Liu, J.; Nandy, P.; Vleet, T. V.; Gargi, U.; Gupta, S.; He, Y.; Lambert, M.; Livingston, B.; and Sampath, D. 2010. The YouTube Video Recommendation System. In ACM Recsys

2010
[11]

V.; Dieleman, S.; and Schrauwen, B

den Oord, A. V.; Dieleman, S.; and Schrauwen, B. 2013. Deep content-based music recommendation. In NIPS, 2643--2651

2013
[12]

Gauci, J.; Conti, E.; Liang, Y.; Virochsiri, K.; He, Y.; Kaden, Z.; Narayanan, V.; Ye, X.; Chen, Z.; and Fujimoto, S. 2018. Horizon: Facebook's open source applied reinforcement learning platform. arXiv preprint arXiv:1811.00260

2018 arXiv
[13]

Ge, Y.; Zhao, S.; Zhou, H.; Pei, C.; Sun, F.; Ou, W.; and Zhang, Y. 2020. Understanding Echo Chambers in E-commerce Recommender Systems. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, 2261--2270

2020
[14]

A.; and Hunt, N

Gomez-Uribe, C. A.; and Hunt, N. 2016. The netflix recommender system: Algorithms, business value, and innovation. In TMIS

2016
[15]

Hastie, T.; Mazumder, R.; Lee, J.; and Zadeh, R. 2015. Matrix Completion and Low-Rank SVD via Fast Alternating Least Squares. In JLMR

2015
[16]

He, X.; Pan, J.; Jin, O.; Xu, T.; Liu, B.; Xu, T.; Shi, Y.; Atallah, A.; Herbrich, R.; Bowers, S.; and et al. 2014. Practical lessons from predicting clicks on ads at facebook. In International Workshop on Data Mining for Online Advertising, 1--9

2014
[17]

Hurley, N.; and Zhang, M. 2011. Novelty and Diversity in Top-N Recommendation -- Analysis and Evaluation. In ACM Transactions on Internet Technology (TOIT), 14

2011
[18]

Jiang, R.; Chiappa, S.; Lattimore, T.; Gy \"o rgy, A.; and Kohli, P. 2019. Degenerate feedback loops in recommender systems. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, 383--390

2019
[19]

P.; Sivakumar, S.; and Wilkinson, D

Knijnenburg, B. P.; Sivakumar, S.; and Wilkinson, D. 2016. Recommender systems for self-actualization. In Proceedings of the 10th ACM Conference on Recommender Systems, 11--14

2016
[20]

Koren, Y. 2008. Factorization meets the neighborhood: a multifaceted collaborative filtering model. In ACM SIGKDD international conference on knowledge discovery and data mining, 426--434

2008
[21]

Koren, Y.; Bell, R.; and Volinsky, C. 2009. Matrix Factorization Techniques for Recommender Systems. In Computer, 30--37

2009
[22]

Krichene, W.; Mayoraz, N.; Rendle, S.; Zhang, L.; Yi, X.; Hong, L.; Chi, E.; and Anderson, J. 2018. Effcient training on very large corpora via gramian estimation. In arXiv

2018
[23]

D.; and Seung, H

Lee, D. D.; and Seung, H. S. 1999. Learning the parts of objects by non-negative matrix factorization. In Nature

1999
[24]

L'Huillier, A.; Castagnos, S.; and Boyer, A. 2014. Understanding usages by modeling diversity over time

2014
[25]

Mnih, A.; and Salakhutdinov, R. R. 2008. Probabilistic matrix factorization. In NIPS, 1257--1264

2008
[26]

T.; Hui, P.-M.; Harper, F

Nguyen, T. T.; Hui, P.-M.; Harper, F. M.; Terveen, L.; and Konstan, J. A. 2014. Exploring the filter bubble: the effect of using recommender systems on content diversity. In Proceedings of the 23rd international conference on World wide web, 677--686

2014
[27]

D.; Rosati, J.; Tomeo, P.; and Sciascio, E

Noia, T. D.; Rosati, J.; Tomeo, P.; and Sciascio, E. D. 2017. Adaptive multi-attribute diversity for recommender systems. volume 382-383, 234 -- 253

2017
[28]

Okura, S.; Tagami, Y.; Ono, S.; and Tajima, A. 2017. Embedding-based News Recommendation for Millions of Users. In SIGKDD

2017
[29]

Paudel, B.; F.Christoffel; and C.Newell, A. 2016. Updatable, Accurate, Diverse, and Scalable Recommendations for Interactive Applications. In ACM Transactions on Interactive Intelligent Systems

2016
[30]

Paudel, B.; Haas, T.; and Bernstein, A. 2017. Fewer Flops at the Top: Accuracy, Diversity, and Regularization in Two-Class Collaborative Filtering. In RecSys, 215--223

2017
[31]

Rendle, S. 2010. Factorization Machines. In ICDM: IEEE International Conference on Data Mining, 995--1000

2010
[32]

Rendle, S.; Freudenthaler, C.; and Schmidt-Thieme, L. 2010. Factorizing personalized markov chains for next-basket recommendation. In WWW: the 19th international conference on world wide web, 811--820

2010
[33]

Rendle, S.; and Schmidt-Thieme, L. 2010. Pairwise interaction tensor factorization for personalized tag recommendation. In WSDM: the third ACM international conference on web search and data mining, 81--90

2010
[34]

Salakhutdinov, R.; and Mnih, A. 2008. Bayesian probabilistic matrix factorization using Markov chain Monte Carlo. In ACM ICML, 880--887

2008
[35]

Salakhutdinov, R.; Mnih, A.; and Hinton, G. 2007. Restricted Boltzmann machines for collaborative filtering. In ICML, 791--798

2007
[36]

Srebro, N.; Rennie, J.; and Jaakkola, T. S. 2012. Maximum-margin matrix factorization. In NIPS, 1329--1336

2012
[37]

Steck, H. 2018. Calibrated recommendations. In Proceedings of the 12th ACM conference on recommender systems, 154--162

2018
[38]

Vargas, S.; Baltrunas, L.; Karatzoglou, A.; and Castells, P. 2014. Coverage, redundancy and size-awareness in genre diversity for recommender systems. In Proceedings of the 8th ACM Conference on Recommender systems, 209--216

2014
[39]

Wang, H.; Wang, N.; and Yeung, D.-Y. 2015. Collaborative deep learning for recommender systems. In SIGKDD, 1235--1244

2015
[40]

Xie, R.; Ling, C.; Wang, Y.; Wang, R.; Xia, F.; and Lin, L. 2020. Deep Feedback Network for Recommendation. In Proceedings of IJCAI-PRICAI

2020
[41]

L.; and Darrell, T

Zhai, A.; Kislyuk, D.; Jing, Y.; Feng, M.; Tzeng, E.; Donahue, J.; Du, Y. L.; and Darrell, T. 2017. Visual discovery at pinterest. In International Conference on World Wide Web Companion, 515--524

2017
[42]

Zhang, S.; Yao, L.; Sun, A.; and Tay, Y. 2019. Deep learning based recommender system: A survey and new perspectives. ACM Computing Surveys (CSUR) 52(1): 1--38

2019
[43]

Zhang, Z.; Zheng, X.; and Zeng, D. 2016. A framework for diversifying recommendation lists by user interest expansion. In Knowledge-Based Systems

2016
[44]

Zhao, X.; Zhang, L.; Ding, Z.; Xia, L.; Tang, J.; and Yin, D. 2018. Recommendations with Negative Feedback via Pairwise Deep Reinforcement Learning. In ACM SIGKDD, 1040–--1048

2018
[45]

Zhao, Z.; Cheng, Z.; Hong, L.; and Chi, E. H. 2015. Improving user topic interest profiles by behavior factorization. In International Conference on World Wide Web, 1406--1416

2015
[46]

Zou, L.; Xia, L.; Ding, Z.; Song, J.; Liu, W.; and Yin, D. 2019. Reinforcement Learning to Optimize Long-term User Engagement in Recommender Systems. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2810--2818

2019

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address author booktitle chapter doi edition editor eid howpublished institution isbn issn journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Adomavicius, G.; and Kwon, Y. 2011. Maximizing Aggregate Recommendation Diversity: A Graph-Theoretic Approach. In DiveRS

work page 2011

[4] [4]

Adomavicius, G.; and Kwon, Y. 2012. Improving Aggregate Recommendation Diversity Using Ranking-Based Techniques. In IEEE Transactions on Knowledge and Data Engineering, 896--911

work page 2012

[5] [5]

Antikacioglu, A.; and Ravi, R. 2017. Post processing recommender systems for diversity. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 707--716

work page 2017

[6] [6]

Bradley, K.; and Smyth, B. 2001. Improving recommendation diversity. In Proceedings of the Twelfth Irish Conference on Artificial Intelligence and Cognitive Science, Maynooth, Ireland, 85--94. Citeseer

work page 2001

[7] [7]

Castagnos, S.; Brun, A.; and Boyer, A. 2013. When Diversity Is Needed... But Not Expected!

work page 2013

[8] [8]

Cheng, H.-T.; Koc, L.; Harmsen, J.; Shaked, T.; Chandra, T.; Aradhye, H.; Anderson, G.; Corrado, G.; Chai, W.; Ispir, M.; et al. 2016. Wide & deep learning for recommender systems. In ACM Recsys, 7--10

work page 2016