Pith. sign in

REVIEW 4 major objections 5 minor 41 references

A 66-thousand-parameter gate at the output layer of a frozen sequential recommender recovers a meaningful share of the accuracy lost when users return after more than a year of inactivity.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 11:39 UTC pith:IUUZPKC2

load-bearing objection DeltaGate is a neat idea and the framing is useful, but the headline >365d gains are selected on the reported test buckets, so they're not credible until a clean validation split. the 4 major comments →

arxiv 2607.19802 v1 pith:IUUZPKC2 submitted 2026-07-22 cs.IR cs.LG

Zero-Observation User Reactivation with Gap-Driven Dimensional Gating

Jiandong Ding , Tianying Liu , Fuyuan Liu , Huijie Qin , Tiandeng Wu This is my paper
classification cs.IR cs.LG
keywords sequential recommendationuser reactivationzero-observation reactivationgap-driven routingdimensional gatingcold-starttemporal gapparameter-efficient adaptation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Sequential recommenders are usually evaluated as if the user's last interaction were recent, but many users reappear after months or years of silence. This paper tries to establish that those long-gap returns are a distinct failure mode and that the gap duration itself is a usable signal. Across three e-commerce categories and three backbone families, the top-10 hit rate (Hit@10) falls monotonically as the no-interaction gap grows and bottoms out after one year. The proposed DeltaGate, a roughly 66-thousand-parameter gate on top of a frozen backbone, blends each dimension of the stale user representation with a learned global prior according to gap length, lifting the >365-day Hit@10 from 0.031 to 0.047 on Video Games. The claim matters because it suggests dormant users can be served with a cheap plug-in rather than a full retraining.

Core claim

The paper's central claim is that after a long observation-free gap, a pre-gap user representation is only partially trustworthy and its trustworthiness varies by latent dimension. DeltaGate operationalizes this by computing, for each dimension, a retention weight g = σ(MLP([log(1+Δt) ∥ h_pers])) and forming h_final = g ⊙ h_pers + (1-g) ⊙ h_global, where h_global is a zero-initialized learned prior. The gate is trained while the backbone stays frozen. In the >365-day bucket, this recovers a substantial slice of the decay: DG-SASRec reaches 0.047 Hit@10 versus 0.031 for SASRec on Video Games, and DG-BERT4Rec reaches 0.046 versus 0.025 for BERT4Rec, with about 66,000 trainable parameters. Comp

What carries the argument

The central mechanism is DeltaGate's elementwise gate: g = σ(MLP([log(1+Δt) ∥ h_pers])), a sigmoid vector that decides per latent dimension how much of the frozen pre-gap representation to keep versus a shared learned prior, with final state h_final = g ⊙ h_pers + (1-g) ⊙ h_global. It is an output-layer plugin, so it costs O(d^2) time and about 66K parameters at d=128, independent of sequence length and backbone depth. The key work it performs is routing: it lets a stale dimension be replaced by the prior while a stable dimension is retained, conditioned jointly on gap length and the user's own representation.

Load-bearing premise

The load-bearing premise is that users in different gap buckets are otherwise comparable; if long-gap users differ systematically in pre-gap history length, item-era familiarity, or catalog drift, the paper's own Section 6 admits this protocol cannot remove that drift, and the measured decay-and-recovery pattern could be an artifact of cohort differences rather than the gap itself.

What would settle it

Build a matched-cohort test: for each >365-day user, find (or synthesize by truncating the same user's history at an earlier point) a counterpart with the same pre-gap sequence length and the same calendar window of item availability, so only the simulated gap differs; if Hit@10 no longer declines monotonically and DeltaGate's advantage disappears, the gap duration is not the operative cause.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Long inactivity gaps are common and costly: the >365-day cohort makes up 19.2% of Video Games test users, and every evaluated backbone loses accuracy monotonically as the gap widens.
  • A dimension-wise gap-conditioned gate can recover a portion of that loss without touching the backbone (e.g., Video Games >365d: 0.047 vs 0.031 Hit@10 for SASRec).
  • The gain comes from the learned per-dimension mixture, not from either endpoint: the shared-prior-only ablation lands at 0.005-0.008 Hit@10, far below the routed model.
  • End-to-end retraining with the gap feature reaches higher raw accuracy (0.080 vs 0.047 on Video Games >365d) but uses ~40x more trainable parameters and shifts item embeddings 139.7%, so the frozen plugin is the lower-risk deployment when embeddings are shared.
  • DeltaGate increases catalog coverage on all three datasets, indicating the prior route is not collapsing to a narrow popularity list.

Where Pith is reading between the lines

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

  • Editorial inference: If representation dimensions really have different temporal half-lives, per-dimension gate values could be read directly as interest-decay rates and used to time re-engagement messages (e.g., approaching dormant users when a stable-interest dimension is still active).
  • Editorial inference: The single shared prior is a bottleneck; a prior conditioned on user type or refreshed periodically with the catalog would likely improve the fallback endpoint, since the paper itself notes a fixed prior cannot track fast-changing catalogs.
  • Editorial inference: The monotonic-decay result may be partly an artifact of cohort drift, because the protocol does not match pre-gap history length or control catalog evolution; a paired experiment that artificially truncates the same user's history at different gap lengths would isolate the causal contribution of Δt.
  • Editorial inference: The 19.2% macro-gap share and the gate's behavior are measured on product categories; on streaming or news platforms, where item relevance decays faster, the same gate might need a larger prior weight at shorter gaps, and the delta may be larger.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces the Zero-Observation Reactivation problem for sequential recommendation: a returning user with a pre-gap history but no behavioral signal during a long inactivity gap, Δt. The authors propose a Gap-Synthesize Protocol (user-level leave-one-out, natural-gap stratification into <30d, 30–90d, 90–180d, 180–365d, >365d) and document monotonic Hit@10 decay across gap buckets for GRU4Rec, SASRec, and BERT4Rec on three Amazon categories. They then propose DeltaGate, a frozen-backbone output-layer gate g = σ(MLP([log(1+Δt) ∥ h_pers])) that interpolates dimension-wise between the personalized representation and a learned zero-initialized global prior. The central empirical claim is that DeltaGate improves Hit@10 in the >365d bucket, e.g., DG-SASRec 0.047 vs SASRec 0.031 on Video Games, with 66K trainable parameters. The paper also reports routing-component ablations, an adaptation-strategy comparison, a controlled gate diagnostic varying Δt with fixed h_pers, and coverage/exclusive-hit analyses.

Significance. If the central claim survives scrutiny, this is a genuinely useful contribution: it identifies an under-studied but practically important evaluation slice, proposes a lightweight plugin with zero backbone drift, and provides a reproducible protocol (code is released). The controlled diagnostic in Figure 2 is a good idea for isolating the gate's sensitivity to Δt. The parameter efficiency (66K vs 2.6M) and the explicit comparison with end-to-end retraining are strong points. However, the current evidence base is undermined by checkpoint selection on the reported evaluation set and unmatched natural-gap cohorts, both of which the paper itself acknowledges. These issues are fixable and do not invalidate the core idea, but they must be addressed before the headline numbers can be treated as out-of-sample evidence.

major comments (4)
  1. [§5.1, Table 3] Checkpoint selection on the reported evaluation set invalidates the headline >365d Hit@10 as an unbiased estimate. The paper states: 'we select checkpoints by mean Hit@10 over the four non-active buckets of the held-out leave-one-out set. Because that set is also reported, these results are descriptive under the historical protocol rather than unbiased final-test estimates.' The four non-active buckets are exactly the rows reported in Table 3. Every number in that table is thus the maximum over checkpoints on the same data used to support the central claim. This inflates expected performance, especially for models with additional stochastic capacity (the gate MLP and h_global). A clean validation split that is never used for checkpoint selection or any model choice must be introduced, with test results reported separately. Without this, the quantitative gains are not out-of-sample estima
  2. [§3.3, §6] Natural-gap cohorts are not matched for pre-gap history properties or catalog drift. Users with >365d gaps may systematically differ from active users in history length, item popularity at the time of the target, catalog composition, and user maturity. The paper acknowledges this: 'Natural-gap stratification also cannot remove catalog drift over multi-year intervals.' This confounds the monotonic decay pattern and DeltaGate's advantage. The authors should either match cohorts on observable covariates (e.g., history length, item frequency, time of last interaction) or perform a within-user gap synthesis that uses active users and masks a long gap, or at minimum show that the reported pattern holds after such adjustment.
  3. [Table 3, §5.2] Table 3 omits the standard deviations that the text says are computed ('Results are means ± standard deviations over five random seeds'). The table shows only point estimates. This is not merely presentational: significance claims such as p<0.01 for DG-BERT4Rec cells cannot be assessed without variance information or confidence intervals. The authors should report full distributions, standard deviations, or confidence intervals in the main table, or at least in an appendix with explicit references from the main text.
  4. [§4.2, §5.3, RQ4] The 'gap-driven' claim lacks a no-Δt control. Since the gate input is [log(1+Δt) ∥ h_pers], the gate could in principle learn to rely exclusively on h_pers; the controlled diagnostic varies Δt with fixed h_pers, but it does not show that the Δt pathway contributes to prediction accuracy. The routing-component controls (Table 4) compare the learned route against the two endpoints (g=0 and g=1), but not against a gate that receives only h_pers. Adding an ablation where Δt is removed from the gate input (i.e., gate conditioned only on h_pers) would isolate the contribution of the gap signal and directly test the paper's central design hypothesis.
minor comments (5)
  1. [Figure 2] The text says Dimension 115 increases from g=0.00 to g=0.97, and Dimension 37 decreases from g=1.00 to g=0.49. Since σ has range (0,1), exact values 0.00 and 1.00 are impossible; this should be stated as rounding, or the endpoint values recomputed with more precision.
  2. [§5.5] The text says 'Among dimensions monotone over all seven supplied gaps' but Figure 2 shows only three gaps (0d, 180d, 730d). The seven supplied gap values should be listed explicitly.
  3. [Table 6] The columns ΔHit and ΔNDCG are described as relative changes, but the name Δ suggests an absolute difference. Clarify in the caption or header (e.g., '%ΔHit').
  4. [§3.1] The Time-Decay Popularity baseline is described only in one sentence. Provide the decay formula or a reference so the reader can reproduce this non-personalized reference.
  5. [Table 5] The Gain/1M values appear to be based on unrounded differences; with the rounded values shown, DeltaGate Plugin's gain/1M would be 0.242, not 0.245. State explicitly that the computation uses unrounded values (as the text does for Table 6, but restate here for clarity).

Circularity Check

0 steps flagged

No significant circularity: DeltaGate's gains are empirical and held-out; self-citations are contextual only.

full rationale

I walked the derivation chain. The core claims are empirical: DeltaGate's gate g=σ(MLP([log(1+Δt)∥h_pers])) and global prior h_global are trained with full-catalog cross-entropy (Eq. 7) on training users and evaluated on held-out leave-one-out targets. Neither Eq. 5 nor Eq. 6 is defined in terms of the reported Hit@10 values, so the central 'prediction' is not forced by construction. The monotonic decay across gap buckets and the Figure 2 gate diagnostic are descriptive post-hoc measurements, not fitted parameters relabeled as predictions. The only self-citations are Refs. [4] and [5], both co-authored by present authors; they appear solely in related-work context (hierarchical/lifelong recommenders and RPE4Rec) and are not load-bearing for DeltaGate's design, no 'uniqueness theorem' is imported, and no ansatz is smuggled in via citation. The paper explicitly admits its own evaluation limitations: Sec. 5.1 says checkpoint selection on the four non-active buckets of the held-out leave-one-out set makes Table 3 'descriptive under the historical protocol rather than unbiased final-test estimates,' and Sec. 6 admits natural-gap stratification 'cannot remove catalog drift over multi-year intervals.' These are validity/fairness concerns about test-set selection and confounding, not circularity: the evaluated numbers do not reduce to the model's inputs by definition. The derivation chain is self-contained, and no circular step is present.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 1 invented entities

DeltaGate depends on a pre-trained frozen backbone, a log-transformed gap input, a zero-initialized shared prior, and an evaluation assumption that natural-gap cohorts are comparable. The two learned components (66K gate weights and h_global) are model parameters fitted to data.

free parameters (2)
  • Gate MLP weights and h_global (66,304 parameters) = 66,304
    Learned by full-catalog cross-entropy on training data; these weights determine the gate outputs and the global prior, so improvements depend entirely on their fitted values.
  • Gate hidden dimension (2d=256) = 256
    Chosen by hand as twice the embedding dimension d=128; no ablation.
axioms (3)
  • domain assumption Pre-trained backbones (SASRec, GRU4Rec, BERT4Rec) yield informative h_pers when frozen, so that a frozen-backbone plugin can improve scoring.
    The method updates only the gate; if the frozen h_pers is poor for dormant users, the gate has no useful source signal.
  • domain assumption Natural-gap cohorts are comparable across buckets absent explicit control for catalog drift and history-length differences.
    The protocol stratifies by naturally observed terminal gap without matching; the paper states this limitation in Section 6.
  • ad hoc to paper log(1+Δt) is a sufficient encoding of gap duration for a 2-layer MLP gate.
    The paper picks the log transform to compress range but provides no principled derivation.
invented entities (1)
  • h_global (learned global prior) no independent evidence
    purpose: Shared fallback vector that the gate blends with h_pers when gate values are low, providing a non-personalized alternative for stale dimensions.
    A d-dimensional vector initialized to zero and trained with the gate. It has no meaning outside the model and no falsifiable prediction; it is a model artifact, not an independent entity.

pith-pipeline@v1.3.0-alltime-deepseek · 11895 in / 11948 out tokens · 121180 ms · 2026-08-01T11:39:38.227007+00:00 · methodology

0 comments
read the original abstract

Sequential recommendation (SR) models capture continuously observed behavior, but a returning user may have no interactions for months or years. We define this setting as Zero-Observation Reactivation: the user has a pre-gap history, while the platform observes no behavioral signals during a macro-gap Delta t. Under a chronologically aligned Gap-Synthesize Protocol on three Amazon datasets (Video Games, CDs & Vinyl, and Movies & TV), Hit@10 decreases monotonically across the evaluated gap buckets and reaches its lowest level beyond one year. The pattern appears across recurrent, unidirectional, and bidirectional SR backbones. We propose DeltaGate, a lightweight output-layer plugin that keeps the backbone frozen and routes each representation dimension between the personalized history and a learned, zero-initialized global prior. The gate is conditioned jointly on Delta t and the personalized representation. In a controlled diagnostic, we hold the personalized representation fixed and vary Delta t to isolate the trained gate's response to the gap input. In the >365d Video Games bucket, DG-SASRec reaches 0.047 Hit@10 versus 0.031 for SASRec, while DG-BERT4Rec reaches 0.046 versus 0.025 for BERT4Rec, with 66K trainable parameters (2--4% overhead). End-to-end retraining attains higher absolute accuracy but changes the backbone embeddings; the frozen plugin preserves zero backbone drift, uses about 40x fewer trainable parameters, and retains observable dimension-wise routing. The source code is available at https://github.com/jdding/DeltaGate.

Figures

Figures reproduced from arXiv: 2607.19802 by Fuyuan Liu, Huijie Qin, Jiandong Ding, Tiandeng Wu, Tianying Liu.

Figure 1
Figure 1. Figure 1: Hit@10 across gap buckets on Movies & TV. All [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Controlled gate diagnostic with a fixed user-history [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: DeltaGate architecture. One selected frozen backbone produces [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

41 extracted references · 1 canonical work pages

  1. [1]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. InProceedings of the 17th ACM Conference on Recommender Systems. 1007–1014

    2023

  2. [2]

    Léa Briand, Guillaume Salha-Galvan, Walid Bendada, Mathieu Morlon, and Viet- Anh Tran. 2021. A Semi-Personalized System for User Cold Start Recommendation on Music Streaming Apps. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 2601–2609

    2021

  3. [3]

    Jianxin Chang, Chenbin Zhang, Zhiyi Fu, Xiaoxue Zang, Lin Guan, Jing Lu, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, and Kun Gai. 2023. TWIN: TWo-stage Interest Network for Lifelong User Behavior Modeling in CTR Prediction at Kuaishou. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3815–3825

    2023

  4. [4]

    Zerui Chen, Heng Chang, Tianying Liu, Chuantian Zhou, Yi Cao, Jiandong Ding, Ming Liu, and Bing Qin. 2026. Beyond the Flat Sequence: Hierarchical and Preference-Aware Generative Recommendations. InProceedings of the ACM Web Conference 2026. Association for Computing Machinery, 7999–8007. doi:10.1145/ 3774904.3792790

    arXiv 2026

  5. [5]

    Ke Cheng, Heng Chang, Pengyang Wang, Liang Gu, Jiandong Ding, Yi Cao, Junchen Ye, and Bowen Du. 2026. RPE4Rec: Enhancing Dynamic Node Re- trieval with Efficient Relative Position Encoding for Recommendation Systems. InProceedings of the Nineteenth ACM International Conference on Web Search and Data Mining. Association for Computing Machinery, 89–98. doi:1...

    work page doi:10.1145/3773966 2026

  6. [6]

    Sung Min Cho, Eunhyeok Park, and Seungjoon Yoo. 2020. MEANTIME: Mixture of Attention Mechanisms with Multi-temporal Embeddings for Sequential Recom- mendation. InProceedings of the 14th ACM Conference on Recommender Systems (Virtual Event, Brazil)(RecSys ’20). Association for Computing Machinery, New York, NY, USA, 68–77. doi:10.1145/3383313.3412216

    arXiv 2020

  7. [7]

    Ziwei Fan, Zhiwei Liu, Jiawei Zhang, Yun Xiong, Lei Zheng, and Philip S. Yu

  8. [8]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  9. [9]

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-Efficient Transfer Learning for NLP. InInternational Conference on Machine Learning. PMLR, 2790–2799

    2019

  10. [10]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InInternational Conference on Learning Representations

    2022

  11. [11]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Rec- ommendation. InProceedings of the 2018 IEEE International Conference on Data Mining (ICDM). 197–206

    2018

  12. [12]

    Walid Krichene and Steffen Rendle. 2020. On sampled metrics for item recom- mendation. InProceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. 1748–1757

    2020

  13. [13]

    Srijan Kumar, Xikun Zhang, and Jure Leskovec. 2019. Predicting dynamic em- bedding trajectory in temporal interaction networks. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 1269–1278

    2019

  14. [14]

    Jiacheng Li, Yujie Wang, and Julian McAuley. 2020. Time Interval Aware Self- Attention for Sequential Recommendation. InProceedings of the 13th ACM Inter- national Conference on Web Search and Data Mining (WSDM). 322–330

    2020

  15. [15]

    Chengkai Liu, Jianghao Lin, Jianling Wang, Hanzhou Liu, and James Caverlee

  16. [16]

    Jie Luo, Wenyu Zhang, Xinming Zhang, and Yuan Fang. 2026. Time-Aware Adaptive Side Information Fusion for Sequential Recommendation. InProceedings of the 19th ACM International Conference on Web Search and Data Mining (WSDM)

    2026

  17. [17]

    Jiaqi Ma, Zhe Zhao, Yi Xinyang, Jilin Chen, Lichan Hong, and Ed H Chi. 2018. Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture- of-Experts. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 1930–1939

    2018

  18. [18]

    Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying Recommendations using Distantly-Labeled Reviews and Fine-Grained Aspects. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP). 188–197

    2019

  19. [19]

    2009.Causality: Models, Reasoning, and Inference(2nd ed.)

    Judea Pearl. 2009.Causality: Models, Reasoning, and Inference(2nd ed.). Cambridge University Press

    2009

  20. [20]

    Qi Pi, Xiaoqiang Zhu, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, and Kun Gai. 2020. Search-based User Interest Modeling with Lifelong Sequential Behavior Data for Click-Through Rate Prediction. InProceedings of the 29th ACM International Conference on Information & Knowledge Management. 2685–2692

    2020

  21. [21]

    Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2010. Factor- izing personalized markov chains for next-basket recommendation. InProceedings of the 19th international conference on World wide web. 811–820

    2010

  22. [22]

    Sculley, Gary Holt, Daniel Golovin, Eugene Davydov, Todd Phillips, Dietmar Ebner, Vinay Chaudhary, Michael Young, Jean-François Crespo, and Dan Denni- son

    D. Sculley, Gary Holt, Daniel Golovin, Eugene Davydov, Todd Phillips, Dietmar Ebner, Vinay Chaudhary, Michael Young, Jean-François Crespo, and Dan Denni- son. 2015. Hidden Technical Debt in Machine Learning Systems. InAdvances in Neural Information Processing Systems, Vol. 28. 2503–2511

    2015

  23. [23]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang

  24. [24]

    Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. 2022. LST: Ladder Side-Tuning for Parameter and Memory Efficient Transfer Learning. InAdvances in Neural Information Processing Systems, Vol. 35. 12991–13005

    2022

  25. [25]

    Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommenda- tion via convolutional sequence embedding. InProceedings of the eleventh ACM international conference on web search and data mining. 565–573

    2018

  26. [26]

    Sheng, and Mehmet Orgun

    Shoujin Wang, Liang Hu, Yan Wang, Longbing Cao, Quan Z. Sheng, and Mehmet Orgun. 2021. A Survey on Session-based Recommender Systems.Comput. Surveys 54, 7 (2021), 1–38

    2021

  27. [27]

    Tianxin Wei, Fuli Feng, Jiawei Chen, Ziwei Wu, Jinfeng Yi, and Xiangnan He

  28. [28]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, Hui Xiong, and Enhong Chen. 2024. A Survey on Large Language Models for Recommendation.World Wide Web27, 5 (2024), 60

    2024

  29. [29]

    Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, and Tieniu Tan. 2019. Session-based recommendation with graph neural networks. InProceedings of the AAAI conference on artificial intelligence, Vol. 33. 346–353

    2019

  30. [30]

    Xiangyu Zhang, Zongqiang Kuang, Zehao Zhang, Fan Huang, and Xianfeng Tan

  31. [31]

    Yang Zhang, Fuli Feng, Xiangnan He, Tianxin Wei, Chonggang Song, Guohui Ling, and Yongdong Zhang. 2021. Causal intervention for leveraging popularity bias in recommendation. InProceedings of the 44th International ACM SIGIR Conference. 11–20

    2021

  32. [32]

    InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining

    Model-agnostic counterfactual reasoning for eliminating popularity bias in recommender system. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 1791–1800

  33. [33]

    Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. 2020. S 3-Rec: Self-Supervised Learning for Sequential Recommendation with Mutual Information Maximization. InPro- ceedings of the 29th ACM International Conference on Information and Knowledge Management. 1893–1902

    2020

  34. [34]

    Yu Zhu, Hao Li, Yikang Liao, Beidou Wang, Ziyu Guan, Haifeng Liu, and Deng Cai

  35. [38]

    Yu Zheng, Chen Gao, Xiang Li, Xiangnan He, Yong Li, and Depeng Jin. 2021. Disentangling user interest and conformity for recommendation with causal embedding. InProceedings of the Web Conference 2021. 2980–2991

    2021

  36. [2016]

    InInter- national Conference on Learning Representations (ICLR)

    Session-based recommendations with recurrent neural networks. InInter- national Conference on Learning Representations (ICLR)

  37. [2017]

    InProceedings of the 26th International Joint Conference on Artificial Intelligence

    What to do next: Modeling user behaviors by time-LSTM. InProceedings of the 26th International Joint Conference on Artificial Intelligence. 3602–3608

  38. [2019]

    InProceedings of the 28th ACM International Conference on Information and Knowledge Management (CIKM)

    BERT4Rec: Sequential Recommendation with Bidirectional Encoder Rep- resentations from Transformer. InProceedings of the 28th ACM International Conference on Information and Knowledge Management (CIKM). 1441–1450

  39. [2021]

    InProceedings of the 30th ACM International Conference on Information and Knowledge Management

    Continuous-Time Sequential Recommendation with Temporal Graph Col- laborative Transformer. InProceedings of the 30th ACM International Conference on Information and Knowledge Management. 433–442

  40. [2023]

    InDatabase Systems for Advanced Applications: 28th International Conference, DASFAA 2023

    Cold & Warm Net: Addressing Cold-Start Users in Recommender Systems. InDatabase Systems for Advanced Applications: 28th International Conference, DASFAA 2023. 532–543

    2023

  41. [2024]

    Mamba4Rec: Towards Efficient Sequential Recommendation with Selective State Space Models.arXiv preprint arXiv:2403.03900(2024)

    Pith/arXiv arXiv 2024