Pith. sign in

REVIEW 3 major objections 5 minor 63 references

MEGG claims a per-sample gradient score, GGscore, ranks historical user-item interactions by influence, and replaying only the most extreme-scoring samples keeps incrementally trained recommenders within about 0.1–0.2 percent RMSE of full r

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 ↗

A gradient-alignment influence score (GGscore) that selects the highest- and lowest-scoring old interactions for replay improves incremental neural recommendation slightly over random replay, mainly at large replay ratios.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A useful empirical replay-sampling paper for neural recommenders whose theoretical derivation in Eq. (9) does not hold as stated; it deserves a serious referee, but only a conditional accept after the derivation, the keep-both-ends rule, and the novelty claims are fixed. the 3 major comments →

arxiv 2509.07319 v1 pith:DLQIOPYE submitted 2025-09-09 cs.IR cs.AI

MEGG: Replay via Maximally Extreme GGscore in Incremental Learning for Neural Recommendation Models

Yunxiao Shi , Shuo Yang , Haimin Zhang , Li Wang , Yongze Wang , Qiang Wu , Min Xu This is my paper
classification cs.IR cs.AI
keywords incremental learningexperience replayrecommender systemscatastrophic forgettingsample influenceGGscoregradient alignmentdata-centric AI
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.

The reading

MEGG argues that in incremental learning for neural recommender models, the right question is not which past samples are prototypical or near decision boundaries, but which ones actually influence training. It defines GGscore, the dot product of a sample's gradient with a reference gradient, and shows that this quantity is proportional to the One Step Loss Change, an estimate of how much removing the sample would move the loss in one optimization step. The paper's proposal is to replay the samples with the most extreme GGscores—keeping both the lowest- and highest-scoring halves of the reservoir—and it reports that this rule consistently beats random replay, prototype replay, and boundary replay across three recommender architectures and four datasets. In most settings the replay set performs within a few tenths of a percent of retraining on all historical data, which is the practical target for an online recommender.

Core claim

The central claim is that the influence of a user-item interaction on a neural recommender's training can be measured by GGscore, GG(z,θ|V) = V · ∇θ L(z,θ), and that this score is proportional to the One Step Loss Change under mini-batch gradient descent. Because gradients need only be computed on the user embedding, item embedding, and final-layer parameters, the score is cheap even at large embedding sizes. Given that score, the paper claims the best replay subset is not the most prototypical or most forgotten samples but the maximally extreme ones, and it shows empirically that this keep-both-ends rule outperforms existing replay samplers and recent knowledge-distillation and meta-learnin

What carries the argument

The central object is the GGscore, a dot product between a sample's parameter gradient and a reference vector, which the paper derives as proportional to the One Step Loss Change in mini-batch gradient descent (Eq. 7–9). The reference vector is the gradient of the total loss at the converged parameters, while the sample gradient is evaluated at the penultimate-epoch parameters. The second load-bearing piece is the keep-both-ends rule (Eq. 10), which retains the lowest- and highest-scoring samples and discards the middle of the score distribution. The paper's justification for this two-sided rule is empirical rather than derived.

Load-bearing premise

The keep-both-ends rule assumes the least-aligned samples are as worth keeping as the most-aligned ones, yet the paper gives no ablation isolating the contribution of the low-score half.

What would settle it

Run the same three-model, four-dataset protocol with replay sets restricted to (a) only high-score samples, (b) only low-score samples, and (c) random samples of the same size; if either restricted set matches MEGG's margins over random replay, or if either half is inert, the two-sided rule is not the source of the gain.

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

If this is right

  • Plugging MEGG's reservoir into knowledge-distillation or meta-learning incremental frameworks gives larger gains than plugging in random replay, per the paper's combination experiments.
  • Because only user/item embeddings and final-layer parameters are needed for scoring, sampling cost rises only mildly as embedding size grows, unlike full-gradient samplers.
  • At replay ratios above 70 percent of the reservoir, MEGG dominates the other replay samplers; above 80 percent it approaches Full-Batch performance.
  • Since the method is model-agnostic, the same scoring and reservoir logic can be applied to any neural collaborative filtering architecture without changing its internals.

Where Pith is reading between the lines

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

  • If the dominant signal is the highest-scoring, most gradient-aligned samples, the lowest-scoring half may be functioning as a diverse representative buffer; a min-only versus max-only ablation would separate the two mechanisms, and the paper does not run one.
  • GGscore's proportionality to one-step loss change suggests a standalone data-pruning or denoising tool for static training sets, a use the paper mentions as motivation but does not develop.
  • The same keep-both-ends rule might transfer to class-incremental or NLP continual learning, where negative-influence samples are rarely replayed; this is an untested extension.
  • Because the efficiency argument rests on embedding locality, the score could be computed at user or item level instead of per interaction, which the paper does not explore.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 5 minor

Summary. The paper proposes MEGG, an experience-replay method for incremental learning in neural recommender systems. The core idea is a new per-sample influence score, GGscore (Eq. 8), defined as the dot product of a sample gradient with a reference gradient vector, which is claimed to be proportional to the One Step Loss Change (Eq. 9). MEGG retains the samples with the most extreme (lowest and highest) GGscores from a reservoir and replays them together with each incoming data block. Experiments cover three backbone models (WDL, DCN, NFM) and four datasets, with rating prediction (RMSE) and classification (AUC) metrics. The paper reports that MEGG outperforms replay baselines (iCaRL, MIR, GDumb) and incremental methods (IncCTR, SML) in most settings, approaches Full-Batch performance, and also improves IncCTR/SML when used as a replay component. Additional analyses address hyperparameters, sampling efficiency, and replay ratio.

Significance. If the theoretical connection were established, MEGG would be a simple, model-agnostic, and efficient replay sampler for a relatively underexplored problem: incremental learning for neural recommenders. The empirical study is a genuine strength: Table 1 reports means and standard deviations across 24 model-dataset cells, including several baselines, and multiple follow-up analyses (integration with IncCTR/SML, efficiency, hyperparameters, replay ratio) are provided. The paper also openly discloses the settings where GDumb beats MEGG and where replay ratios below 70% are not advantageous. However, the paper's central theoretical claim—that GGscore is proportional to One Step Loss Change—is not justified by the derivation as written, and the both-ends selection rule is not supported by any ablation. The contribution is therefore best viewed currently as an empirical heuristic whose theoretical interpretation and key design choice need substantial revision.

major comments (3)
  1. [Section 4.2, Eqs. (7) and (9)] The claimed proportionality ΔL_i^k ∝ GG(z_k, θ_i | ∇θ_{i+1} L(D_train, θ_{i+1})) is not derived from Definition 2. Starting from Eq. (7), the exact first-order estimate is ΔL_i^k ≈ α/(B−1)[V·g(z_k, θ_i) − V·g(Z_i, θ_i)], where V = g(D_train, θ_{i+1}). The 'interaction term' V·g(Z_i, θ_i) is sample-dependent because Z_i contains z_k, so discarding it changes the ranking over k, not just the scale. The proposed fix—replacing B−1 with B in Eq. (4)—does cancel that term, but it changes the parameter update and therefore the quantity being computed is no longer the One Step Loss Change defined in Definition 2. The paper calls this a 'minor departure,' but it is the only route to Eq. (9). As written, the claim that GGscore measures sample influence is an unsupported construction, not a consequence of the paper's own definitions. The authors should either provide conditions under which the batc
  2. [Section 4.4, Eq. (10)] The 'keep-both-ends' selection rule is not justified. If GGscore measures influence, retaining the highest-score samples is plausible, but retaining the lowest-score samples—those least aligned with the reference gradient—is not. No ablation is provided that isolates the contribution of the low-score half (e.g., min-only, max-only, middle, or random within-score-quantile variants). The only related experiment, Section 5.7, shows MEGG's advantage over GDumb appears only when the replay ratio exceeds 70%; below that, no consistent benefit is visible. Since the default configuration uses K = M − M′ = 90% of the reservoir, the reported Table 1 results do not establish that the both-ends criterion, rather than the high replay volume, is responsible for the gains. This is a load-bearing issue for the proposed algorithm's distinctiveness.
  3. [Section 5.7, Fig. 5] The replay-ratio analysis directly limits the practical scope of the claims. Figure 5 shows MEGG outperforms iCaRL, MIR, and GDumb only when the replay ratio exceeds 70%, and approaches Full-Batch only above 80%. In many incremental-learning deployments, a 70–90% replay ratio of a reservoir that already holds two-thirds of all data is a large memory footprint. The abstract and introduction claim 'superior performance' without this qualification; the conclusion should be revised to state that the advantage is demonstrated at high replay ratios, and the method's behavior in lower-capacity regimes remains unexplained. This also connects to the lack of an ablation for the selection rule: the high-ratio regime is exactly where random selection also becomes competitive.
minor comments (5)
  1. [Table 1] The dataset column header 'Taobao204' appears to be a typo for 'Taobao2014'.
  2. [Algorithm 1, line 4] The reference 'with (??)' is an unresolved placeholder; it should cite Eq. (10).
  3. [Section 5.2] The statement that 'an AUC increase at the 0.001 level is considered statistically significant' is presented without a citation or test; as written it conflates practical significance with statistical significance. Please clarify or remove.
  4. [Section 4.1 / Fig. 2] The MNIST validation of One Step Loss Change is described only briefly. It is unclear how the SGD-Influence Estimator is used to compute 10 One Step Loss Change values per sample and how the cumulative Loss Change is defined. Adding details would strengthen the empirical motivation.
  5. [Section 4.3] The parameter selection rule (embedding parameters plus final fully connected layer) is justified only by prior work. Since the paper emphasizes efficiency, a brief statement of how many parameters are retained relative to the full model for the three backbones would be useful.

Circularity Check

1 steps flagged

Eq. (9)'s proportionality is achieved by redefining the update and dropping the sample-dependent batch term, so the GGscore–influence link is constructed rather than derived.

specific steps
  1. self definitional [Section 4.2, Equation (9) and following paragraph]
    "The assertion that the One-Step Loss Change is proportional to the GGScore can be derived with relative simplicity. This conclusion is reached by omitting the constant term α/(B−1) in Equation (7) and disregarding the interaction term g(D_train, θ_{i+1})·g(Z_i, θ_i). ... To mitigate this, we propose a modification to Equation (4), replacing B−1 with B. This adjustment effectively eliminates the term g(Z_i, θ_i) from ∆L_i^k in Equation (7)."

    Under the actual removal update, Eq. (4), the one-step parameter difference is (α/(B−1))[g(z_k,θ_i) − g(Z_i,θ_i)] (Eq. 5), and because z_k ∈ Z_i, the term g(Z_i,θ_i) depends on the very sample being scored. Dropping it changes the ranking, not merely the scale. The paper then changes the update rule from B−1 to B, which makes the Taylor-expanded loss difference exactly proportional to V·g(z_k,θ_i), i.e., to GGscore by definition. Thus Eq. (9) is not a consequence of the originally defined One Step Loss Change; it is enforced by an update-rule modification introduced specifically to make the interaction term vanish. The central claim that GGscore quantifies sample influence therefore reduces, by construction, to the definition of GGscore.

full rationale

The paper's empirical evaluation is largely self-contained: MEGG is compared against GDumb, iCaRL, MIR, IncCTR, and SML on four public benchmarks, and the MNIST correlation study in Section 3.2 independently supports One Step Loss Change as a surrogate for the original Loss Change. Those parts are not circular. However, the load-bearing theoretical step is Eq. (9), which claims One Step Loss Change is proportional to GGscore. That step is constructed rather than derived: Eq. (7) contains a batch-interaction term V·g(Z_i) that depends on the candidate sample z_k; the paper 'disregards' it and then replaces B−1 with B in the update rule, which changes the quantity being computed. The resulting proportionality is therefore true by definition of the modified update, not by the paper's own Definition 2. This is a partial circularity in the central influence justification. There are no load-bearing self-citations: references [31], [32], and [49] are unrelated prior works by overlapping authors and are not used to justify the core derivation. No uniqueness theorem or ansatz is imported via self-citation. The score of 6 reflects that one key 'prediction' (GGscore as influence) reduces by construction, while the empirical replay comparisons remain independent evidence.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 2 invented entities

The method's contribution rests on the GGscore and the extremes rule. GGscore is a gradient-alignment score whose proportionality to loss change is established by dropping terms and altering the update rule, and the both-ends rule is unjustified and unablated. The evaluation then shows small margins over random replay at a high memory budget, so the ledger contains a moderate number of hand-chosen parameters and assumptions relative to the size of the claimed advance.

free parameters (5)
  • Reservoir size fraction M = 10/15 of the dataset = 10/15 (about two thirds of all data)
    Hand-chosen in Section 5.2; the buffer holds 66.7 percent of all data, so the replay selection is not memory-constrained in the usual continual-learning sense.
  • Replay subset size K = M - M' = 9/15 of the total data (90 percent of the reservoir)
    Default replay ratio; Section 5.7 shows MEGG beats random sampling only when this ratio exceeds 70 percent, so the reported advantage lives in the high-retention regime.
  • Hyperparameter scheme A: penultimate epoch theta' = epoch 4, final epoch theta_hat = epoch 5 = epochs (4, 5) of a 5-epoch run
    Chosen in Section 4.4; Section 5.6 reports that schemes A-D 'exert a substantial influence' on performance, and Table 1 uses only scheme A without stating it was fixed a priori.
  • Extreme-split sizes k_l = floor(K/2), k_r = ceil(K/2) = half and half of the retained budget
    Ad hoc design in Eq. (10); no derivation or ablation supports keeping the low-GGscore half.
  • Training schedule: 5 epochs per stage, batch 1024, learning rate 0.001, embedding dim 64 = 5 / 1024 / 0.001 / 64
    Hand-chosen standard settings in Section 5.2; the GGscore uses epoch-4 and epoch-5 checkpoints, so the schedule is load-bearing for the metric.
axioms (6)
  • standard math First-order Taylor expansion of L(Dtrain, theta_i+1_k) around theta_i+1 with negligible remainder o(||theta_i+1_k - theta_i+1||)
    Invoked in Eqs. (6)-(7) in Section 4.1; requires the parameter change from removing one sample to be small, which is not guaranteed under the original MBGD update.
  • domain assumption A single gradient step approximates the whole training process (one-step SGD-influence assumption)
    Borrowed from [58] in Section 3.2; the paper explicitly replaces full retraining analysis with a single-step analysis.
  • ad hoc to paper Replacing B-1 by B in Eq. (4) yields a valid estimate of the loss change (the modified update cancels the batch-gradient term g(Zi, theta_i))
    Section 4.2: the proportionality in Eq. (9) rests on this substitution; the modified update is no longer the actual MBGD rule, so the derivation is adjusted to obtain the desired dot-product form.
  • domain assumption Parameter-affected locality: only the interacting user and item embeddings plus the final fully-connected layer matter for influence
    Section 4.3, citing [57,60]; gradients on all other parameters are set to zero, and this locality is imported rather than re-validated on the four benchmark datasets.
  • domain assumption One Step Loss Change correlates with true Loss Change near convergence, transferred from a toy MNIST experiment to sparse bipartite recommendation data
    Fig. 2 in Section 3.2 validates on two MNIST classes (200 instances, two-layer net); the transfer to user-item interaction data is asserted, not tested.
  • ad hoc to paper The most extreme GGscore samples, both lowest and highest, are the most valuable replay samples
    Eq. (10) keeps floor(K/2) lowest and ceil(K/2) highest scores; no derivation or ablation supports retaining the low-score half, and Section 5.7 shows the rule only beats random above a 70 percent replay ratio.
invented entities (2)
  • GGscore (Eq. 8: V dot grad_theta L(z, theta)) no independent evidence
    purpose: Quantify per-sample influence for replay selection in incremental recommendation
    No falsifiable handle outside this paper: the MNIST correlation study validates One Step Loss Change, not GGscore, and no ablation on recommendation data shows GGscore tracks true influence.
  • One Step Loss Change (Definition 2) no independent evidence
    purpose: Storage-free surrogate for the true Loss Change (Definition 1)
    Validated only in the toy MNIST study (Fig. 2); its transfer to recommender training dynamics is assumed.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of MEGG: Replay via Maximally Extreme GGscore in Incremental Learning for Neural Recommendation Models." pith.science (2026). https://pith.science/paper/DLQIOPYE

@misc{pith2026250907319,
  author       = {Pith},
  title        = {Pith review of: MEGG: Replay via Maximally Extreme GGscore in Incremental Learning for Neural Recommendation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DLQIOPYE}},
  note         = {Machine review of arXiv:2509.07319}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Neural Collaborative Filtering models are widely used in recommender systems but are typically trained under static settings, assuming fixed data distributions. This limits their applicability in dynamic environments where user preferences evolve. Incremental learning offers a promising solution, yet conventional methods from computer vision or NLP face challenges in recommendation tasks due to data sparsity and distinct task paradigms. Existing approaches for neural recommenders remain limited and often lack generalizability. To address this, we propose MEGG, Replay Samples with Maximally Extreme GGscore, an experience replay based incremental learning framework. MEGG introduces GGscore, a novel metric that quantifies sample influence, enabling the selective replay of highly influential samples to mitigate catastrophic forgetting. Being model-agnostic, MEGG integrates seamlessly across architectures and frameworks. Experiments on three neural models and four benchmark datasets show superior performance over state-of-the-art baselines, with strong scalability, efficiency, and robustness. Implementation will be released publicly upon acceptance.

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

63 extracted references · 44 canonical work pages · 9 internal anchors

  1. [1]

    In: Proceedings of the 26th International Con- ference on World Wide Web, pp

    He, X., Liao, L., Zhang, H., Nie, L., Hu, X., Chua, T.-S.: Neural collaborative filtering. In: Proceedings of the 26th International Con- ference on World Wide Web, pp. 173–182 (2017)

    work page 2017

  2. [2]

    In: Proceedings of the 1st Workshop on Deep Learning for Recommender Systems, pp

    Cheng, H.-T., Koc, L., Harmsen, J., Shaked, T., Chandra, T., Aradhye, H., Anderson, G., Corrado, G., Chai, W., Ispir, M.,et al.: Wide & deep learning for recommender systems. In: Proceedings of the 1st Workshop on Deep Learning for Recommender Systems, pp. 7–10 (2016)

    work page 2016

  3. [3]

    In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowl- edge Discovery and Data Mining, pp

    Shan, Y., Hoens, T.R., Jiao, J., Wang, H., Yu, D., Mao, J.: Deep crossing: Web-scale mod- eling without manually crafted combinatorial features. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowl- edge Discovery and Data Mining, pp. 255–262 (2016)

    work page 2016

  4. [4]

    In: Proceedings of the 40th International ACM 14 SIGIR Conference on Research and Develop- ment in Information Retrieval, pp

    He, X., Chua, T.-S.: Neural factorization machines for sparse predictive analytics. In: Proceedings of the 40th International ACM 14 SIGIR Conference on Research and Develop- ment in Information Retrieval, pp. 355–364 (2017)

    work page 2017

  5. [5]

    arXiv preprint arXiv:1703.04247 (2017)

    Guo, H., Tang, R., Ye, Y., Li, Z., He, X.: Deepfm: a factorization-machine based neu- ral network for ctr prediction. arXiv preprint arXiv:1703.04247 (2017)

    Pith/arXiv arXiv 2017

  6. [6]

    CoRRabs/1803.05170 (2018) 1803.05170

    Lian, J., Zhou, X., Zhang, F., Chen, Z., Xie, X., Sun, G.: xdeepfm: Combining explicit and implicit feature interactions for rec- ommender systems. CoRRabs/1803.05170 (2018) 1803.05170

    Pith/arXiv arXiv 2018

  7. [7]

    Zhou, G., Song, C., Zhu, X., Fan, Y., Zhu, H., Ma, X., Yan, Y., Jin, J., Li, H., Gai, K.: Deep Interest Network for Click-Through Rate Prediction (2018)

    work page 2018

  8. [8]

    arXiv preprint arXiv:1906.00091 (2019)

    Naumov, M., Mudigere, D., Shi, H.-J.M., Huang, J., Sundaraman, N., Park, J., Wang, X., Gupta, U., Wu, C.-J., Azzolini, A.G., et al.: Deep learning recommendation model for personalization and recommendation systems. arXiv preprint arXiv:1906.00091 (2019)

    Pith/arXiv arXiv 1906

  9. [9]

    In: International Symposium on Ubiquitious Computing Systems, pp

    Lee, J.S., Lee, J.C.: Context awareness by case-based reasoning in a music recommenda- tion system. In: International Symposium on Ubiquitious Computing Systems, pp. 45–58 (2007). Springer

    work page 2007

  10. [10]

    In: Proceedings of the 15th Inter- national Conference on Intelligent User Inter- faces, pp

    Liu, J., Dolan, P., Pedersen, E.R.: Person- alized news recommendation based on click behavior. In: Proceedings of the 15th Inter- national Conference on Intelligent User Inter- faces, pp. 31–40 (2010)

    work page 2010

  11. [11]

    In: Proceedings of the 29th ACM International Conference on Information & Knowledge Management, pp

    Chang, B., Jang, G., Kim, S., Kang, J.: Learning graph-based geographical latent representation for point-of-interest recom- mendation. In: Proceedings of the 29th ACM International Conference on Information & Knowledge Management, pp. 135–144 (2020)

    work page 2020

  12. [12]

    In: 2023 5th Interna- tional Conference on Machine Learning, Big Data and Business Intelligence (MLBDBI), pp

    Gu, S., Wang, C., Zhao, G., Wu, L.: Movie recommendation model based on attention mechanism for dynamically capturing user interest evolution. In: 2023 5th Interna- tional Conference on Machine Learning, Big Data and Business Intelligence (MLBDBI), pp. 319–323 (2023). https://doi.org/10.1109/ MLBDBI60823.2023.10481926

    work page arXiv 2023

  13. [13]

    In: Proceedings of the 2nd ACM Conference on Electronic Commerce, pp

    Sarwar, B., Karypis, G., Konstan, J., Riedl, J.: Analysis of recommendation algorithms for e-commerce. In: Proceedings of the 2nd ACM Conference on Electronic Commerce, pp. 158–167 (2000)

    work page 2000

  14. [14]

    In: Proceedings of the 14th ACM Conference on Recommender Systems

    Mi, F., Lin, X., Faltings, B.: Ader: Adaptively distilled exemplar replay towards continual learning for session-based recommendation. In: Proceedings of the 14th ACM Conference on Recommender Systems. RecSys ’20, pp. 408–413. Association for Computing Machin- ery, New York, NY, USA (2020). https://doi. org/10.1145/3383313.3412218 . https://doi. org/10.11...

    arXiv 2020

  15. [15]

    In: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval

    Zhang, Y., Feng, F., Wang, C., He, X., Wang, M., Li, Y., Zhang, Y.: How to retrain recommender system? a sequential meta- learning method. In: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. SIGIR ’20, pp. 1479–1488. Associ- ation for Computing Machinery, New York, NY, USA (2020). https://do...

    arXiv 2020

  16. [16]

    In: Proceedings of the ACM Web Conference

    Xia, J., Li, D., Gu, H., Liu, J., Lu, T., Gu, N.: Fire: Fast incremental recommen- dation with graph signal processing. In: Proceedings of the ACM Web Conference

    work page

  17. [17]

    2861–2868 (2020)

    Xu, Y., Zhang, Y., Guo, W., Guo, H., Tang, R., Coates, M.: Graphsail: Graph structure aware incremental learning for recommender systems, pp. 2861–2868 (2020). https://doi. org/10.1145/3340531.3412754

    work page arXiv 2020

  18. [18]

    In: Proceedings of the European Conference on Computer Vision 15 (ECCV) (2018)

    Castro, F.M., Marin-Jimenez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: Proceedings of the European Conference on Computer Vision 15 (ECCV) (2018)

    work page 2018

  19. [19]

    arXiv preprint arXiv:1904.07734 (2019)

    Ven, G.M., Tolias, A.S.: Three scenar- ios for continual learning. arXiv preprint arXiv:1904.07734 (2019)

    Pith/arXiv arXiv 1904

  20. [20]

    A Comprehensive Study of Class Incremental Learning Algorithms for Visual Tasks

    Belouadah, E., Popescu, A., Kanellos, I.: A comprehensive study of class incremental learning algorithms for visual tasks. CoRR abs/2011.01844(2020) 2011.01844

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  21. [21]

    CoRRabs/2010.15277 (2020) 2010.15277

    Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A.D., Weijer, J.: Class- incremental learning: survey and perfor- mance evaluation. CoRRabs/2010.15277 (2020) 2010.15277

    Pith/arXiv arXiv 2010

  22. [22]

    Incremental Learning of Object Detectors without Catastrophic Forgetting

    Shmelkov, K., Schmid, C., Alahari, K.: Incremental learning of object detectors without catastrophic forgetting. CoRR abs/1708.06977(2017) 1708.06977

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  23. [23]

    In: Scott, D., Bel, N., Zong, C

    Biesialska, M., Biesialska, K., Costa- juss` a, M.R.: Continual lifelong learning in natural language processing: A sur- vey. In: Scott, D., Bel, N., Zong, C. (eds.) Proceedings of the 28th International Con- ference on Computational Linguistics, pp. 6523–6541. International Committee on Computational Linguistics, Barcelona, Spain (Online) (2020). https:/...

    doi:10.18653/v1/2020.coling-main.574 2020

  24. [24]

    In: ICLR (2022)

    Jang, J., Ye, S., Yang, S., Shin, J., Han, J., Kim, G., Choi, S.J., Seo, M.: Towards contin- ual knowledge learning of language models. In: ICLR (2022)

    work page 2022

  25. [25]

    CoRR abs/2109.11369(2021) 2109.11369

    Qu, H., Rahmani, H., Xu, L., Williams, B.M., Liu, J.: Recent advances of continual learn- ing in computer vision: An overview. CoRR abs/2109.11369(2021) 2109.11369

    Pith/arXiv arXiv 2021

  26. [26]

    Neurocomputing469, 28–51 (2022)

    Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., Sanner, S.: Online continual learning in image classification: An empirical survey. Neurocomputing469, 28–51 (2022)

    work page 2022

  27. [27]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Wang, Z., Liu, L., Duan, Y., Kong, Y., Tao, D.: Continual learning with lifelong vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 171–181 (2022)

    work page 2022

  28. [29]

    In: Proceedings of the 30th ACM International Conference on Information & Knowledge Manage- ment

    Wang, Y., Zhang, Y., Coates, M.: Graph structure aware contrastive knowledge dis- tillation for incremental learning in rec- ommender systems. In: Proceedings of the 30th ACM International Conference on Information & Knowledge Manage- ment. CIKM ’21, pp. 3518–3522. Associa- tion for Computing Machinery, New York, NY, USA (2021). https://doi.org/10.114...

    work page arXiv 2021

  29. [30]

    Causal Incremental Graph Convolution for Recommender System Retraining

    Ding, S., Feng, F., He, X., Liao, Y., Shi, J., Zhang, Y.: Causal incremental graph con- volution for recommender system retraining. CoRRabs/2108.06889(2021) 2108.06889

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  30. [31]

    In: Proceedings of the 2nd Interna- tional Workshop on Deep Multimodal Gener- ation and Retrieval

    Shi, Y., Xu, M., Zhang, H., Zi, X., Wu, Q.: A learnable agent collaboration network frame- work for personalized multimodal ai search engine. In: Proceedings of the 2nd Interna- tional Workshop on Deep Multimodal Gener- ation and Retrieval. MMGR ’24, pp. 12–20. Association for Computing Machinery, New York, NY, USA (2024). https://doi.org/10. 1145/3689091...

    work page arXiv 2024

  31. [32]

    https://arxiv.org/abs/2405.06683

    Shi, Y., Zi, X., Shi, Z., Zhang, H., Wu, Q., Xu, M.: ERAGent: Enhancing Retrieval- Augmented Language Models with Improved Accuracy, Efficiency, and Personalization (2024). https://arxiv.org/abs/2405.06683

    Pith/arXiv arXiv 2024

  32. [33]

    LLM-based Medical Assistant Personalization with Short- and Long-Term Memory Coordination

    Zhang, K., Zhao, F., Kang, Y., Liu, X.: Memory-augmented llm personalization with short-and long-term memory coordination. arXiv preprint arXiv:2309.11696 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  33. [34]

    In: Proceedings of the AAAI Symposium Series, vol

    Hatalis, K., Christou, D., Myers, J., Jones, 16 S., Lambert, K., Amos-Binks, A., Dannen- hauer, Z., Dannenhauer, D.: Memory mat- ters: The need to improve long-term memory in llm-agents. In: Proceedings of the AAAI Symposium Series, vol. 2, pp. 277–280 (2023)

    work page 2023

  34. [35]

    RecSys ’22, pp

    Katsileros, P., Mandilaras, N., Mallis, D., Pitsikalis, V., Theodorakis, S., Chamiel, G.: An incremental learning framework for large- scale ctr prediction. RecSys ’22, pp. 490–493. Association for Computing Machinery, New York, NY, USA (2022). https://doi.org/10. 1145/3523227.3547390 . https://doi.org/10. 1145/3523227.3547390

    work page arXiv 2022

  35. [36]

    A Practical Incremental Method to Train Deep CTR Models

    Wang, Y., Guo, H., Tang, R., Liu, Z., He, X.: A practical incremental method to train deep ctr models. arXiv preprint arXiv:2009.02147 (2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2009

  36. [37]

    Incremental Learning for Personalized Recommender Systems

    Ouyang, Y., Shi, J., Wei, H., Gao, H.: Incremental learning for personalized recommender systems. arXiv preprint arXiv:2108.13299 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  37. [38]

    Incremental Factorization Machines for Persistently Cold-starting Online Item Recommendation

    Kitazawa, T.: Incremental Factorization Machines for Persistently Cold-starting Online Item Recommendation (2016). https://arxiv.org/abs/1607.02858

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  38. [39]

    In: User Modeling, Adaptation, and Personaliza- tion: 22nd International Conference, UMAP 2014, Aalborg, Denmark, July 7-11, 2014

    Vinagre, J., Jorge, A.M., Gama, J.: Fast incremental matrix factorization for recom- mendation with positive-only feedback. In: User Modeling, Adaptation, and Personaliza- tion: 22nd International Conference, UMAP 2014, Aalborg, Denmark, July 7-11, 2014. Proceedings 22, pp. 459–470 (2014). Springer

    work page 2014

  39. [40]

    Interna- tional Journal of Computer Vision129(6), 1789–1819 (2021)

    Gou, J., Yu, B., Maybank, S.J., Tao, D.: Knowledge distillation: A survey. Interna- tional Journal of Computer Vision129(6), 1789–1819 (2021)

    2021

  40. [41]

    In: Com- puter Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pp

    Prabhu, A., Torr, P.H., Dokania, P.K.: Gdumb: A simple approach that questions our progress in continual learning. In: Com- puter Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pp. 524–540 (2020). Springer

    work page 2020

  41. [42]

    CoRRabs/1611.07725 (2016) 1611.07725

    Rebuffi, S., Kolesnikov, A., Lampert, C.H.: icarl: Incremental classifier and represen- tation learning. CoRRabs/1611.07725 (2016) 1611.07725

    Pith/arXiv arXiv 2016

  42. [43]

    Online Continual Learning with Maximally Interfered Retrieval

    Aljundi, R., Caccia, L., Belilovsky, E., Cac- cia, M., Lin, M., Charlin, L., Tuytelaars, T.: Online continual learning with maximally interfered retrieval. CoRRabs/1908.04742 (2019) 1908.04742

    work page internal anchor Pith review Pith/arXiv arXiv 1908

  43. [44]

    In: Psychology of Learning and Motivation vol

    McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of Learning and Motivation vol. 24, pp. 109–

    work page

  44. [45]

    In: International Conference on Machine Learn- ing, pp

    Zenke, F., Poole, B., Ganguli, S.: Contin- ual learning through synaptic intelligence. In: International Conference on Machine Learn- ing, pp. 3987–3995 (2017). PMLR

    work page 2017

  45. [46]

    Advances in neural information processing systems30(2017)

    Shin, H., Lee, J.K., Kim, J., Kim, J.: Con- tinual learning with deep generative replay. Advances in neural information processing systems30(2017)

    work page 2017

  46. [47]

    In: European Conference on Information Retrieval, pp

    Hou, Y., Zhang, J., Lin, Z., Lu, H., Xie, R., McAuley, J., Zhao, W.X.: Large lan- guage models are zero-shot rankers for rec- ommender systems. In: European Conference on Information Retrieval, pp. 364–381 (2024). Springer

    work page 2024

  47. [48]

    In: Duh, K., Gomez, H., Bethard, S

    Lyu, H., Jiang, S., Zeng, H., Xia, Y., Wang, Q., Zhang, S., Chen, R., Leung, C., Tang, J., Luo, J.: LLM-rec: Personalized recommenda- tion via prompting large language models. In: Duh, K., Gomez, H., Bethard, S. (eds.) Find- ings of the Association for Computational Linguistics: NAACL 2024, pp. 583–612. Asso- ciation for Computational Linguistics, Mex- ic...

    doi:10.18653/v1/2024.findings-naacl.39 2024

  48. [49]

    https:// arxiv.org/abs/2405.17890

    Xu, W., Wu, Q., Liang, Z., Han, J., Ning, X., Shi, Y., Lin, W., Zhang, Y.: SLM- Rec: Empowering Small Language Models for Sequential Recommendation (2024). https:// arxiv.org/abs/2405.17890

    Pith/arXiv arXiv 2024

  49. [50]

    In: Proceedings of the AAAI Conference on Artificial Intelli- gence, vol

    Kemker, R., McClure, M., Abitino, A., Hayes, 17 T., Kanan, C.: Measuring catastrophic for- getting in neural networks. In: Proceedings of the AAAI Conference on Artificial Intelli- gence, vol. 32 (2018)

    work page 2018

  50. [51]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(7), 3366–3385 (2022) https://doi.org/10.1109/ TPAMI.2021.3057446

    De Lange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learn- ing survey: Defying forgetting in classifica- tion tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence44(7), 3366–3385 (2022) https://doi.org/10.1109/ TPAMI.2021.3057446

    work page arXiv 2022

  51. [52]

    arXiv preprint arXiv:2303.10158 (2023)

    Zha, D., Bhat, Z.P., Lai, K.-H., Yang, F., Jiang, Z., Zhong, S., Hu, X.: Data- centric artificial intelligence: A survey. arXiv preprint arXiv:2303.10158 (2023)

    Pith/arXiv arXiv 2023

  52. [53]

    Transactions on Machine Learning Research (2023)

    Daxberger, E., Swaroop, S., Osawa, K., Yokota, R., Turner, R.E., Hern´ andez-Lobato, J.M., Khan, M.E.: Improving continual learn- ing by accurate gradient reconstructions of the past. Transactions on Machine Learning Research (2023)

    work page 2023

  53. [54]

    CoRR abs/1807.02802(2018) 1807.02802

    Javed, K., Shafait, F.: Revisiting distillation and incremental classifier learning. CoRR abs/1807.02802(2018) 1807.02802

    Pith/arXiv arXiv 2018

  54. [55]

    In: International Conference on Machine Learn- ing, pp

    Koh, P.W., Liang, P.: Understanding black- box predictions via influence functions. In: International Conference on Machine Learn- ing, pp. 1885–1894 (2017). PMLR

    work page 2017

  55. [56]

    Advances in Neural Information Processing Systems32(2019)

    Hara, S., Nitanda, A., Maehara, T.: Data cleansing for models trained with sgd. Advances in Neural Information Processing Systems32(2019)

    work page 2019

  56. [57]

    arXiv preprint arXiv:2205.09329 (2022)

    Yang, S., Xie, Z., Peng, H., Xu, M., Sun, M., Li, P.: Dataset pruning: Reducing training data by examining generalization influence. arXiv preprint arXiv:2205.09329 (2022)

    Pith/arXiv arXiv 2022

  57. [58]

    In: Proceedings of the 2022 International Conference on Manage- ment of Data

    Pradhan, R., Zhu, J., Glavic, B., Salimi, B.: Interpretable data-based explanations for fairness debugging. In: Proceedings of the 2022 International Conference on Manage- ment of Data. SIGMOD ’22, pp. 247–261. Association for Computing Machinery, New York, NY, USA (2022). https://doi.org/10. 1145/3514221.3517886 . https://doi.org/10. 1145/3514221.3517886

    arXiv 2022

  58. [59]

    Proceedings of the IEEE86(11), 2278–2324 (1998) https://doi.org/10.1109/5

    Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE86(11), 2278–2324 (1998) https://doi.org/10.1109/5. 726791

    doi:10.1109/5 1998

  59. [60]

    Selective and Collaborative Influence Function for Efficient Recommendation Unlearning

    Li, Y., Chen, C., Zheng, X., Zhang, Y., Gong, B., Wang, J.: Selective and col- laborative influence function for efficient recommendation unlearning. arXiv preprint arXiv:2304.10199 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  60. [61]

    Springer, ??? (2010)

    Celma, O.: Music Recommendation and Dis- covery in the Long Tail. Springer, ??? (2010)

    work page 2010

  61. [62]

    Wang, R., Fu, B., Fu, G., Wang, M.: Deep & Cross Network for Ad Click Predictions (2017) 18

    work page 2017

  62. [165]

    Elsevier, ??? (1989)

    work page 1989

  63. [2022]

    2360–2369

    WWW ’22, pp. 2360–2369. Associa- tion for Computing Machinery, New York, NY, USA (2022). https://doi.org/10.1145/ 3485447.3512108 . https://doi.org/10.1145/ 3485447.3512108

    work page arXiv 2022

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.