GRAINRec: Graph and Attention Integrated Approach for Real-Time Session-Based Item Recommendations

Amit Pande; Bhavtosh Rath; David Relyea; Prathyusha Kanmanth Reddy; Pushkar Chennu

REVIEW 4 major objections 7 minor 29 references

GRAINRec: Graph and Attention Integrated Approach for Real-Time Session-Based Item Recommendations

T0 review · 4 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A graph-and-attention session recommender can be served in real time via a nearest-neighbor candidate matrix, beating its predecessor by 1.5% offline and lifting live click-through by 10%.

desk verdict An honest, production-oriented extension of LESSR with a useful real-time inference trick; the 1.5% offline gain over LESSR is believable but not fully pinned down because the candidate sets aren't shown to be matched. read the letter →

arxiv 2411.09152 v1 pith:HC2GORXR submitted 2024-11-14 cs.LG

Bhavtosh Rath , Pushkar Chennu , David Relyea , Prathyusha Kanmanth Reddy , Amit Pande This is my paper · ORCID

classification cs.LG

keywords session-basedrecommendationgraphneuralnetworkattentionmechanismreal-timeinferencenearestneighbormatrixe-commercerecommendationsA/BtestingLESSR

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

A session-based recommender predicts the next item a shopper will want from the ordered list of items already in the current basket, rather than from the single most recent item. The paper argues that GRAINRec, a graph-and-attention model built on the earlier LESSR architecture, can generate such dynamic recommendations in real time by restricting the inference candidate set with a precomputed nearest-neighbor matrix. If the claim holds, retailers do not need to precompute per-item recommendations: they can serve session-aware suggestions at production latency. The paper reports an average 1.5% improvement over LESSR across offline metrics, and a two-week A/B test showing a 10% increase in click-through rate and a 9% increase in attributable demand.

What carries the argument

The central mechanism is the nearest-neighbor matrix, a precomputed table of size $ ext{items} imes k$ built from item embeddings during training, which turns a softmax over a catalog of hundreds of thousands of items into a softmax over the union of $k$ nearest neighbors per session item. Around it, the architecture alternates GRU-based graph neural network layers and attention layers so that nearby items are propagated through the graph while distant items are linked by attention, and a readout layer concatenates local and global session embeddings. This combination carries the argument: the alternating layers improve representation quality, and the neighbor matrix makes those representations usable in real time.

What would settle it

Rerun the offline evaluation with GRAINRec and every baseline, especially LESSR, using the identical nearest-neighbor candidate matrix, the same category filter, and the same hyperparameter tuning budget; if GRAINRec's average 1.5% advantage disappears, the central claim is not supported. A second check is to redeploy the tuned model in the production A/B test and see whether the 10% click-through lift and 9% attributable-demand lift reproduce.

Watch

Extended reading notes

Core claim

On the paper's own terms, GRAINRec shows that the bottleneck to real-time session-based recommendation is not the model but the candidate search. The model treats a session as an ordered directed graph without self-loops, alternates GRU-based graph neural network layers with attention layers, and forms a session embedding by concatenating the last item's embedding with an attention-weighted global embedding. For inference it precomputes, during training, a matrix of the top-$k$ nearest item embeddings for every item; for a session of length $t$ the softmax is restricted to the union of those precomputed neighbors, around 300 candidates for a capped three-item session. This lets the model meet the platform's latency service-level agreement while scoring recommendations dynamically as the session evolves, and the paper credits the neighbor matrix, along with self-loop removal and cross-category filtering, for the offline improvement over LESSR.

Load-bearing premise

The central claim stands on the assumption that the baseline models in the offline comparison were evaluated under the same candidate-set restriction, category filter, and hyperparameter tuning as GRAINRec; if the baselines were not given those same conditions, the reported 1.5% improvement may come from the evaluation protocol rather than from the model itself.

Editorial extensions

If this is right

GRAINRec can be served on CPUs at under 50 ms 95th-percentile latency and about 60 requests per second per pod, so dynamic session-aware recommendations are practical in a production retailer setting.
The live metrics attributed to the model are large: a 10% increase in click-through rate and a 9% increase in attributable demand over the production link-prediction baseline.
Hyperparameter tuning accounts for a substantial part of the online gain, moving CTR lift from 6.4% to 10.1% and attributable-demand lift from 5.8% to 9.2%.
Increasing the nearest-neighbor count improves ndcg@10 with diminishing returns beyond roughly 100 neighbors, so the chosen matrix size is a deliberate accuracy-latency trade-off.
The model works better on frequent, less sparse item categories than on discretionary items, suggesting that data sparsity is the main remaining quality limiter.

Reading between the lines

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

Editorial extension: the nearest-neighbor candidate matrix is a general inference technique and could plausibly be bolted onto any session model, so part of GRAINRec's offline edge may come from the candidate set rather than the graph-attention architecture; the paper does not state whether baselines used the same restriction.
Editorial extension: because inference caps session length at 3 while training uses sequences up to 20, the deployed model deliberately discards older session context to meet latency; testing longer caps against the SLA would map the quality-versus-latency frontier directly.
Editorial extension: the same pattern of precomputing neighbor sets per item and then scoring dynamically against a union applies beyond this paper to other real-time personalization surfaces, such as search or homepage feeds, where a full-catalog softmax is infeasible.

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

Quick take: this is a straightforward engineering paper, not a new learning method, and it says so. The genuinely useful piece is the nearest-neighbor matrix heuristic for real-time inference, plus the deployment write-up (CPU serving, thread pinning, gRPC, 50ms p95). The model architecture is LESSR with three tweaks—no self-loops, same-category filtering, tuned hyperparameters—and the paper is upfront about that. The offline numbers are plausible but the evaluation has a real hole: Table III doesn't say whether the baselines were scored over the same ~300-item candidate set that GRAINRec uses, or the full catalog. Since the paper itself credits the nearest-neighbor matrix with part of the offline gain (Section VI.E), and Figure 4 shows ndcg@10 varies from 0.132 to 0.154 as the neighborhood size changes, the 1.5% margin over LESSR could partly be an artifact of a smaller candidate set. That's not an accusation—it's a missing description. The fix is easy: state that all baselines used the same candidate generation and category filter, or rerun with matched protocols.

Also worth noting: the online A/B is against an unspecified production link-prediction model, not against LESSR, so it says nothing about the offline claim. And there are no error bars or significance tests anywhere; with point estimates this close (0.258 vs 0.255 in hit@10 for frequency items), a few thousand test sequences could flip the ranking. That said, for a production report this level of rigor is about average, and the paper is honest about what it did.

What's actually new: the combination of a LESSR-style model with a precomputed nearest-neighbor lookup to meet a real-time SLA, validated on Target's data and traffic. That is a legitimate applied contribution, and the ablation studies give the reader a feel for the trade-offs. I don't see any sign of circular reasoning or invented results; the model is a standard softmax classifier over item embeddings, and the nearest-neighbor matrix is a heuristic to cut the search space.

Who's this for? Practitioners building real-time session recommenders in industry, and researchers who want a clear example of what a deployed system actually needs. It's not a methods paper. I'd send it to a venue with an applied track, or a workshop, where the evaluation protocol can be pushed on. If the authors clarify the candidate-set question and add significance testing, it becomes a useful reference.

Referee Report

4 major / 7 minor

Summary. GRAINRec is a session-based recommendation model that extends LESSR by alternating GNN and attention layers, removing self-loops, applying a same-category session filter, and introducing a nearest-neighbor matrix to restrict the candidate space for real-time inference. The paper reports offline evaluation on Target retail data showing an average improvement of about 1.5% over LESSR across hit@10, mrr@10, and ndcg@10, as well as an online A/B test reporting a 10% increase in click-through rate and a 9% increase in attributable demand against a production link-prediction model. The authors also describe their CPU-based inference deployment, which meets a 50 ms p95 latency SLA, and include ablation studies on layer ordering, nearest-neighbor size, and embedding dimension.

Significance. If the reported improvement over LESSR and the real-time inference results are reliable, GRAINRec constitutes a useful industry case study in making graph-and-attention session-based recommenders deployable at scale. The paper provides a detailed system description, including deployment infrastructure, latency measurements, and A/B test results, which is valuable for practitioners. The offline evaluation, however, is the only quantitative comparison against LESSR, and its fairness is not established. The paper does not release code or data, and the online test baseline is not an academic baseline, so the scientific contribution hinges on the validity of the offline protocol. On the positive side, the paper clearly identifies which components contribute to performance (self-loop removal, category filtering, nearest-neighbor matrix) via ablation studies, and it reports a real production SLA achievement.

major comments (4)

[Section VI.E, Table III] The claimed average 1.5% improvement over LESSR is based on point estimates without confidence intervals or significance tests. For example, on frequency items the mrr@10 difference is 0.122 vs. 0.124 (1.64%), which is likely within run-to-run variance for neural recommenders. Since this margin is the central quantitative claim of the paper, the authors should provide error bars over multiple runs or a significance test (e.g., paired bootstrap) for each metric and category.
[Section IV.C and Section VI.E] The offline evaluation protocol is not shown to be apples-to-apples. GRAINRec's inference restricts the candidate space to the union of the top-100 nearest neighbors of each session item (about 300 items) and applies a same-category filter and a session-length cap of 3. The text states that 'what contributes to the performance improvement is removal of self-loops, cross category dependency and nearest neighbor matrix,' yet it never states whether LESSR and the other baselines in Table III were evaluated with the same candidate restriction and session filters. If the baselines scored over the full catalog while GRAINRec scored over ~300 candidates, the reported improvements could be an artifact of the evaluation protocol. The authors must clarify this and, if the protocols differ, re-run the baselines under the same candidate set and filtering rules.
[Section VI.D] The online A/B test compares GRAINRec to an unspecified production 'link prediction' model, not to LESSR or any offline baseline. Therefore the reported 10% CTR and 9% AD improvements do not support the claim of superiority over LESSR. In addition, the table reports no confidence intervals, no statistical test, and no description of how the before/after hyperparameter-tuning A/B tests were controlled (e.g., same time period, same traffic split). This is load-bearing because the offline comparison is the only quantitative evidence for the advantage over LESSR.
[Section VI.B] The inferencing setup applies a session-length cap of 3 and a same-category filter 'for both training and inference,' but it is unclear whether these constraints were applied identically to all baseline models in the offline evaluation. If the baselines were trained or evaluated on unfiltered, full-length sessions while GRAINRec used the capped and filtered versions, the comparison is not fair. The authors should specify exactly which preprocessing and candidate-generation steps were shared across all models in Table III.

minor comments (7)

[Abstract] The phrase 'To addressing this challenge' should be 'To address this challenge'.
[Section VIII] The sentence 'we saw a improvement across all metrics' contains a grammatical error; it should be 'we saw an improvement.'
[Figure 5 caption] The caption reads 'ndcg@10 for different neighborhood sizes in the nearest neighbor matrix,' but the experiment varies the embedding dimension, not the neighborhood size. Please correct the caption.
[Section VIII] The phrase 'pertained embeddings' should be 'pre-trained embeddings.'
[Table III] The 'Improvement' row should explicitly define the percentage as relative improvement over LESSR and should state whether the average is taken over all nine metric-category combinations or computed differently. The abstract's 'average improvement of 1.5%' does not exactly match the row values, so the computation should be clarified.
[Section IV.B, Eqs. (14)-(15)] The notation h_n = s^T V_n is ambiguous because V_n was introduced as the value matrix in Eq. (6). It should be clarified that h_n is a scalar score for item n and that the softmax in Eq. (15) is over item scores for all candidate items.
[References] Reference [24] (Pei et al., Interacting Attention-Gated Recurrent Networks) does not appear to be cited anywhere in the text. Please either cite it or remove it.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the model's improvements are empirical comparisons against external baselines; the nearest-neighbor candidate heuristic is an ablated inference component, not a prediction that reduces to its own input.

full rationale

Walking the derivation chain, GRAINRec's architecture is a standard GNN-plus-attention session encoder with a readout layer, followed by a softmax classification loss; none of the equations define a target quantity in terms of the quantity it is supposed to predict. The claimed 1.5% offline improvement over LESSR is an empirical result from Table III, computed with standard ranking metrics, and the online A/B test compares against a separate production link-prediction model. The nearest-neighbor matrix is generated from the model's learned item embeddings and restricts the inference candidate space, but the paper presents this as a heuristic latency-accuracy trade-off and ablates its size in Figure 4; it is not a parameter fitted so that a stated prediction becomes true by construction. There is no load-bearing self-citation, no imported uniqueness theorem, and no renaming of a known result. The main defensible concern is whether Table III is apples-to-apples with respect to candidate restriction and session filtering for the baselines; that is an evaluation-protocol and correctness risk, not a circularity of the kind this pass scores. Therefore no specific circular step can be quoted, and the appropriate finding is no significant circularity.

Assumptions & free parameters 10 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a set of empirical preprocessing and inference choices whose efficacy is asserted from the authors' own experiments, plus a proprietary dataset. There are no physical or mathematical postulates beyond standard deep learning machinery, but several domain assumptions (self-loop removal, category filtering, candidate-set restriction) are load-bearing for the reported gains.

free parameters (10)

embedding_dimension = 256
Tuned with scikit-optimize; ablation in Fig. 5 explores 32-1024 and shows diminishing returns.
dropout_rate = 0.146
Hyperparameter chosen by tuning, not from first principles.
learning_rate = 0.00045
Tuned hyperparameter.
batch_size = 1024
Tuned hyperparameter affecting training dynamics.
decay_rate = 0.0001
Weight decay chosen by tuning.
layer_configuration = 2-layer GNN-Attention
Selected to balance accuracy and latency; ablation in Fig. 3 and Fig. 5.
n_neighbors = 100
Chosen from ablation in Fig. 4 as a tradeoff between ndcg@10 and SLA latency.
inference_session_cap = 3
Upper limit on session length during inference; directly affects the candidate set and the comparison.
min_item_frequency = 10
Items appearing fewer than 10 times in the sequence data were excluded to mitigate sparsity.
sequence_length_cap = 20
Training sequences are capped at 20 items.

assumptions (5)

domain assumption Removing self-loops (consecutive duplicate items) improves recommendation performance.
Section V states this was found empirically; no theoretical justification is given.
domain assumption Excluding cross-category items produces more coherent session-based recommendations.
Section V and Section VI.B apply this filter during both training and inference.
domain assumption The union of top-100 nearest neighbors of session items is a sufficient candidate set for relevant recommendations.
Section IV.C and Fig. 4 motivate this choice; it directly affects inference and possibly the offline comparison.
domain assumption Session context is summarized by the last item (local embedding) plus an attention-weighted aggregate (global embedding).
Readout layer in Section IV.B.3; this is a standard design assumption, not derived.
standard math Cross-entropy loss and softmax over the item catalog are appropriate training objectives for next-item prediction.
Standard deep learning machinery assumed reliable for this task.

how reviews work

0 comments

Cite this review

Pith. "Pith review of GRAINRec: Graph and Attention Integrated Approach for Real-Time Session-Based Item Recommendations." pith.science (2026). https://pith.science/paper/HC2GORXR

@misc{pith2026241109152,
  author       = {Pith},
  title        = {Pith review of: GRAINRec: Graph and Attention Integrated Approach for Real-Time Session-Based Item Recommendations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HC2GORXR}},
  note         = {Machine review of arXiv:2411.09152}
}

read the original abstract

Recent advancements in session-based recommendation models using deep learning techniques have demonstrated significant performance improvements. While they can enhance model sophistication and improve the relevance of recommendations, they also make it challenging to implement a scalable real-time solution. To addressing this challenge, we propose GRAINRec: a Graph and Attention Integrated session-based recommendation model that generates recommendations in real-time. Our scope of work is item recommendations in online retail where a session is defined as an ordered sequence of digital guest actions, such as page views or adds to cart. The proposed model generates recommendations by considering the importance of all items in the session together, letting us predict relevant recommendations dynamically as the session evolves. We also propose a heuristic approach to implement real-time inferencing that meets Target platform's service level agreement (SLA). The proposed architecture lets us predict relevant recommendations dynamically as the session evolves, rather than relying on pre-computed recommendations for each item. Evaluation results of the proposed model show an average improvement of 1.5% across all offline evaluation metrics. A/B tests done over a 2 week duration showed an increase of 10% in click through rate and 9% increase in attributable demand. Extensive ablation studies are also done to understand our model performance for different parameters.

Figures

Figures reproduced from arXiv: 2411.09152 by the authors.

**Figure 1.** Architecture of GRAINRec. Data processing generates ordered sequences of items interacted within a session, which [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Architecture of model inferencing V. COMPARISON OF GRAINREC AND LESSR The proposed GRAINRec architecture is an enhanced version of the LESSR model [23], incorporating several key modifications to improve performance (Table III). Unlike the original LESSR model, which included self-loops allowing consecutive item repetitions, GRAINRec was trained without self-loops, which we found improved performance. Crosscategor… view at source ↗

**Figure 4.** ndcg@10 for different neighborhood sizes in the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: ndcg@10 for different neighborhood sizes in the nearest neighbor matrix [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 26 canonical work pages

[1]

& Lian, D

Wang, S., Cao, L., Wang, Y ., Sheng, Q., Orgun, M. & Lian, D. A survey on session-based recommender systems. ACM Computing Surveys (CSUR). 54, 1-38 (2021)

work page 2021
[2]

Song, Y ., Elkahky, A. & He, X. Multi-rate deep learning for temporal recommendation. Proceedings Of The 39th International ACM SIGIR Conference On Research And Development In Information Retrieval . pp. 909-912 (2016)

work page 2016
[3]

& Dragut, E

Shen, C., Han, C., He, L., Mukherjee, A., Obradovic, Z. & Dragut, E. Session-based News Recommendation from Temporal User Comment- ing Dynamics. 2022 IEEE/ACM International Conference On Advances In Social Networks Analysis And Mining (ASONAM). pp. 163-170 (2022) Fig. 5: ndcg@10 for different neighborhood sizes in the nearest neighbor matrix

work page 2022
[4]

Wang, D., Deng, S., Zhang, X. & Xu, G. Learning music embedding with metadata for context aware recommendation. Proceedings Of The 2016 ACM On International Conference On Multimedia Retrieval . pp. 249-253 (2016)

work page 2016
[5]

& Mokbel, M

Bao, J., Zheng, Y . & Mokbel, M. Location-based and preference-aware recommendation using sparse geo-social networking data. Proceedings Of The 20th International Conference On Advances In Geographic Information Systems. pp. 199-208 (2012)

work page 2012
[6]

Wang, X., Liu, X., Li, L., Chen, X., Liu, J. & Wu, H. Time-aware user modeling with check-in time prediction for next POI recommendation. 2021 IEEE International Conference On Web Services (ICWS) . pp. 125- 134 (2021)

work page 2021
[7]

& Chi, E

Beutel, A., Covington, P., Jain, S., Xu, C., Li, J., Gatto, V . & Chi, E. Latent cross: Making use of context in recurrent recommender systems. Proceedings Of The Eleventh ACM International Conference On Web Search And Data Mining . pp. 46-54 (2018)

work page 2018
[8]

Yap, G., Li, X. & Yu, P. Effective next-items recommendation via personalized sequential pattern mining. International Conference On Database Systems For Advanced Applications . pp. 48-64 (2012)

work page 2012

Show all 29 references

[9]

& Zhang, Z

Hu, H., He, X., Gao, J. & Zhang, Z. Modeling personalized item frequency information for next-basket recommendation. Proceedings Of The 43rd International ACM SIGIR Conference On Research And Development In Information Retrieval . pp. 1071-1080 (2020)

2020
[10]

& Lauw, H

Le, D., Fang, Y . & Lauw, H. Modeling sequential preferences with dynamic user and context factors. Joint European Conference On Machine Learning And Knowledge Discovery In Databases . pp. 145- 161 (2016)

2016
[11]

& Tikk, D

Hidasi, B., Karatzoglou, A., Baltrunas, L. & Tikk, D. Session- based recommendations with recurrent neural networks. ArXiv Preprint arXiv:1511.06939. (2015)

2015 arXiv
[12]

Li, J., Ren, P., Chen, Z., Ren, Z., Lian, T. & Ma, J. Neural attentive session-based recommendation. Proceedings Of The 2017 ACM On Conference On Information And Knowledge Management . pp. 1419- 1428 (2017)

2017
[13]

& McAuley, J

Kang, W. & McAuley, J. Self-attentive sequential recommendation. 2018 IEEE International Conference On Data Mining (ICDM) . pp. 197-206 (2018)

2018
[14]

& Caverlee, J

Wang, J., Ding, K., Hong, L., Liu, H. & Caverlee, J. Next-item recommendation with sequential hypergraphs. Proceedings Of The 43rd International ACM SIGIR Conference On Research And Development In Information Retrieval . pp. 1101-1110 (2020)

2020
[15]

& Wang, K

Tang, J. & Wang, K. Personalized top-n sequential recommendation via convolutional sequence embedding. Proceedings Of The Eleventh ACM International Conference On Web Search And Data Mining. pp. 565-573 (2018)

2018
[16]

& Philip, S

Shi, C., Hu, B., Zhao, W. & Philip, S. Heterogeneous information network embedding for recommendation. IEEE Transactions On Knowl- edge And Data Engineering . 31, 357-370 (2018)

2018
[17]

& Zhang, H

Liu, Q., Zeng, Y ., Mokhosi, R. & Zhang, H. STAMP: short-term attention/memory priority model for session-based recommendation. Proceedings Of The 24th ACM SIGKDD International Conference On Knowledge Discovery & Data Mining . pp. 1831-1839 (2018)

2018
[18]

& Others A survey of graph neural networks for recommender systems: Challenges, methods, and directions

Gao, C., Zheng, Y ., Li, N., Li, Y ., Qin, Y ., Piao, J., Quan, Y ., Chang, J., Jin, D., He, X. & Others A survey of graph neural networks for recommender systems: Challenges, methods, and directions. ACM Transactions On Recommender Systems . 1, 1-51 (2023)

2023
[19]

& Tan, T

Wu, S., Tang, Y ., Zhu, Y ., Wang, L., Xie, X. & Tan, T. Session-based recommendation with graph neural networks. Proceedings Of The AAAI Conference On Artificial Intelligence . 33, 346-353 (2019)

2019
[20]

& De Rijke, M

Pan, Z., Cai, F., Chen, W., Chen, H. & De Rijke, M. Star graph neural networks for session-based recommendation. Proceedings Of The 29th ACM International Conference On Information & Knowledge Management. pp. 1195-1204 (2020)

2020
[21]

& Wang, Z

Wang, H., Zhang, F., Zhang, M., Leskovec, J., Zhao, M., Li, W. & Wang, Z. Knowledge-aware graph neural networks with label smoothness regularization for recommender systems. Proceedings Of The 25th ACM SIGKDD International Conference On Knowledge Discovery & Data Mining. pp. 9...

2019
[22]

& Zhang, J

Li, X., Wang, X., Zhang, H. & Zhang, J. Session-based Recommendation with Temporal Graph Neural Network and Contrastive Learning. 2023 3rd International Conference On Neural Networks, Information And Communication Engineering (NNICE) . pp. 10-14 (2023)

2023
[23]

& Wong, R

Chen, T. & Wong, R. Handling information loss of graph neural networks for session-based recommendation. Proceedings Of The 26th ACM SIGKDD International Conference On Knowledge Discovery & Data Mining. pp. 1172-1180 (2020)

2020
[24]

& Tax, D

Pei, W., Yang, J., Sun, Z., Zhang, J., Bozzon, A. & Tax, D. Interacting attention-gated recurrent networks for recommendation. Proceedings Of The 2017 ACM On Conference On Information And Knowledge Management. pp. 1459-1468 (2017)

2017
[25]

& Others The YouTube video recommendation system

Davidson, J., Liebald, B., Liu, J., Nandy, P., Van Vleet, T., Gargi, U., Gupta, S., He, Y ., Lambert, M., Livingston, B. & Others The YouTube video recommendation system. Proceedings Of The Fourth ACM Conference On Recommender Systems . pp. 293-296 (2010)

2010
[26]

& Schmidt-Thieme, L

Rendle, S., Freudenthaler, C. & Schmidt-Thieme, L. Factorizing person- alized markov chains for next-basket recommendation. Proceedings Of The 19th International Conference On World Wide Web . pp. 811-820 (2010)

2010
[27]

& Yahav, E

Alon, U. & Yahav, E. On the bottleneck of graph neural networks and its practical implications. ArXiv Preprint arXiv:2006.05205 . (2020)

2020 arXiv
[28]

& Pashami, S

Del Moral, P., Nowaczyk, S. & Pashami, S. Why is multiclass classifi- cation hard?. IEEE Access. 10 pp. 80448-80462 (2022)

2022
[29]

& Bengio, Y

Veli ˇckovi´c, P., Cucurull, G., Casanova, A., Romero, A., Lio, P. & Bengio, Y . Graph attention networks.ArXiv Preprint arXiv:1710.10903. (2017)

2017 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

& Lian, D

Wang, S., Cao, L., Wang, Y ., Sheng, Q., Orgun, M. & Lian, D. A survey on session-based recommender systems. ACM Computing Surveys (CSUR). 54, 1-38 (2021)

work page 2021

[2] [2]

Song, Y ., Elkahky, A. & He, X. Multi-rate deep learning for temporal recommendation. Proceedings Of The 39th International ACM SIGIR Conference On Research And Development In Information Retrieval . pp. 909-912 (2016)

work page 2016

[3] [3]

& Dragut, E

Shen, C., Han, C., He, L., Mukherjee, A., Obradovic, Z. & Dragut, E. Session-based News Recommendation from Temporal User Comment- ing Dynamics. 2022 IEEE/ACM International Conference On Advances In Social Networks Analysis And Mining (ASONAM). pp. 163-170 (2022) Fig. 5: ndcg@10 for different neighborhood sizes in the nearest neighbor matrix

work page 2022

[4] [4]

Wang, D., Deng, S., Zhang, X. & Xu, G. Learning music embedding with metadata for context aware recommendation. Proceedings Of The 2016 ACM On International Conference On Multimedia Retrieval . pp. 249-253 (2016)

work page 2016

[5] [5]

& Mokbel, M

Bao, J., Zheng, Y . & Mokbel, M. Location-based and preference-aware recommendation using sparse geo-social networking data. Proceedings Of The 20th International Conference On Advances In Geographic Information Systems. pp. 199-208 (2012)

work page 2012

[6] [6]

Wang, X., Liu, X., Li, L., Chen, X., Liu, J. & Wu, H. Time-aware user modeling with check-in time prediction for next POI recommendation. 2021 IEEE International Conference On Web Services (ICWS) . pp. 125- 134 (2021)

work page 2021

[7] [7]

& Chi, E

Beutel, A., Covington, P., Jain, S., Xu, C., Li, J., Gatto, V . & Chi, E. Latent cross: Making use of context in recurrent recommender systems. Proceedings Of The Eleventh ACM International Conference On Web Search And Data Mining . pp. 46-54 (2018)

work page 2018

[8] [8]

Yap, G., Li, X. & Yu, P. Effective next-items recommendation via personalized sequential pattern mining. International Conference On Database Systems For Advanced Applications . pp. 48-64 (2012)

work page 2012