Pith. sign in

REVIEW 4 major objections 5 minor 32 references

PaletteID claims that representing each item as a similarity-weighted blend of a small set of real prototype items—rather than as discrete residual codebook tokens—consistently improves multimodal CTR prediction across different backbones a

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

2026-08-03 15:37 UTC pith:FEVG3IZ7

load-bearing objection Squarely novel prototype-composed SID method with a convincing cross-prefix diagnostic; CTR gains are small and the embedding-similarity premise is untested. the 4 major comments →

arxiv 2607.29000 v1 pith:FEVG3IZ7 submitted 2026-07-31 cs.IR cs.LG

PaletteID: Prototype-Composed Semantic Identifiers for Multimodal CTR Prediction

Huanyu Liu , Baining Chen , Hui Liu , Zengyang Li , Ziyi Huang This is my paper
classification cs.IR cs.LG
keywords click-through rate predictionmultimodal recommendationsemantic identifiersdeterminantal point processprototype compositionlong-tail recommendationquantizationembedding robustness
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 paper sets out to fix two weaknesses in semantic-identifier methods for click-through rate prediction: hard codebook assignment discards fine-grained continuous similarity, and hierarchical residual codes suffer from prefix-dependent semantics that don't scale with depth. Its proposed solution, PaletteID (PID), selects a compact palette of real items as semantic anchors using a quality-aware determinantal point process, then represents each item as a top-K weighted composition of those anchors. The authors argue this preserves graded multimodal proximity, provides more robust assignments under embedding perturbation, and yields more interpretable tokens. On two public datasets, PID improves AUC/GAUC over residual SID baselines across both DCNV2 and RankMixer backbones, with tail-item AUC rising from 0.5819 to 0.5947 on TAOBAO-MM. A sympathetic reader would care because the method offers a drop-in item-side feature that helps exactly where ID-based models are weakest.

Core claim

The central claim is that semantic identifiers for multimodal CTR should be prototype-composed rather than codebook-quantized. PaletteID constructs a global palette of M real items selected by Semantic Quality-Aware DPP, which balances local content density (so prototypes sit in representative regions) against global diversity (so the palette covers the semantic space). Each target item is then assigned a PaletteID: an ordered list of its top-K similar prototypes plus continuous similarity scores; the item's semantic representation is a sigmoid-gated, similarity-weighted sum of trainable prototype embeddings. The paper argues this preserves the graded structure of the pretrained multimodal e

What carries the argument

SQ-DPP (Semantic Quality-Aware Determinantal Point Process): a DPP over items with a cosine-RBF kernel L_ij = q_i * exp(-gamma * (1 - cos(x_i, x_j))) * q_j, where q_i is a normalized local-density quality score from ANN neighbors. Its job is to pick the M prototype items once offline so that the palette is both semantically representative (quality term) and non-redundant (determinant diversity), with greedy MAP inference via incremental Cholesky updates. The second piece is the PID aggregation layer: top-K prototype retrieval, similarity-threshold filtering, and sigmoid gating with a monotonic regularizer to produce a weighted mixture embedding that is fed as an item-side feature into any CT

Load-bearing premise

The whole pipeline assumes that cosine similarity in the frozen pretrained multimodal embedding space is a valid, stable proxy for whether items are related in ways that matter for click prediction; if that space misaligns with user behavior, both SQ-DPP prototype selection and top-K retrieval pick anchors that share irrelevant signals.

What would settle it

Train PID with the same hyperparameters but replace the pretrained embeddings with (a) a randomly re-initialized encoder or (b) an encoder fine-tuned on CTR labels; if AUC gains persist under random embeddings, the gains come from the composition/architecture rather than semantic proximity; if they vanish, the mechanism is confirmed. Alternatively, on a dataset where semantically-similar items are intentionally set to have opposite click propensities, PID should fail to beat a pure ID baseline.

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

If this is right

  • PID yields consistent AUC/GAUC improvements over VQ-VAE, RQ-VAE, and RQ-KMeans SIDs on both TAOBAO-MM and KuaiRec, under both DCNV2 and RankMixer backbones (e.g., TAOBAO-MM DCNV2+ AUC 0.6275 to 0.6314).
  • Long-tail items benefit most: tail-group AUC rises from 0.5819 to 0.5947 on TAOBAO-MM, more than head or torso groups, indicating effective semantic knowledge transfer from frequent to sparse items.
  • PID assignments are more robust than residual SIDs to perturbations in the multimodal embedding space; assignment cosine similarity degrades smoothly, and deeper SID levels are the least stable.
  • The required palette size grows with the semantic complexity of the item space, not catalog size: the optimal M moves from 200 (on 10.7K items) to 650 (on 1.0M items), suggesting scalability to large catalogs.

Where Pith is reading between the lines

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

  • If cosine similarity in the frozen pretrained embedding space is the load-bearing assumption, replacing the encoder (e.g., a different foundation model) should change PID's gains; the paper never varies this, so an ablation across embedding models would directly probe the mechanism.
  • Because PID's prototypes are real items, the palette itself could be personalized per user or per cluster (the paper lists this as future work); one could expect larger gains in heterogeneous catalogs, but also a risk of overfitting the palette to head items if density is estimated globally.
  • The monotonic regularization and sigmoid gating imply the model learns to treat prototypes as non-competing facets; in domains where item facets are actually mutually exclusive (e.g., strict categories), softmax-style normalization might win — a domain contingency worth testing.
  • PID embeddings could be indexed for approximate nearest-neighbor search, making the semantic anchors useful beyond CTR, e.g., in candidate generation or generative retrieval pipelines.

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 proposes PaletteID (PID), a prototype-composed semantic identifier for multimodal CTR prediction. Instead of quantizing item embeddings into hierarchical residual codes, PID selects a small palette of real prototype items via a Semantic Quality-Aware DPP (SQ-DPP) that balances local density and global diversity, retrieves the top-K prototypes for each item, and produces a weighted combination of prototype embeddings using similarity-derived gates. The method is evaluated on TAOBAO-MM and KuaiRec with DCNV2+ and RankMixer+ backbones, reporting consistent AUC/GAUC improvements over VQ-VAE, RQ-VAE, and RQ-KMeans, larger gains for long-tail items, more stable assignments under embedding perturbation, and interpretable prototype compositions. The paper also includes ablation studies of aggregation, prototype selection, and density-quality thresholds.

Significance. If the reported gains are real, PID is a simple and practical addition to CTR pipelines: it is specified clearly, can be integrated into standard backbones as an extra item-side feature, and has low online overhead because retrieval is offline. The authors provide several useful analyses beyond the headline numbers, including long-tail bucket analysis, assignment robustness under perturbation, semantic facet coverage, and a cross-prefix consistency diagnostic. These are genuine strengths. However, the significance is currently conditional: the main improvements over existing SID methods are on the order of 0.001–0.004 AUC, no uncertainty estimates or significance tests are reported, and the contribution of the proposed SQ-DPP selection itself appears very small relative to random prototype selection (Table 8). The core premise that cosine similarity in a frozen pretrained embedding space is a valid proxy for CTR-relevant semantics is also untested.

major comments (4)
  1. [§5.2 (Table 3) and §5.3.1 (Table 4)] The central claim of consistent CTR improvement is supported only by point estimates without any measure of uncertainty. All gains over RQ-VAE are in the 0.0017–0.0041 AUC range on TAOBAO-MM and 0.0012–0.0018 on KuaiRec; the latter dataset is very small (37.9K instances), where such differences can easily be within run-to-run or seed-level noise. The paper should report multiple seeds, standard deviations, and significance tests (e.g., paired bootstrap over users or items) for the main tables and for the long-tail comparisons. Without these, the headline claim is not statistically established.
  2. [§5.4.2 (Table 8)] The ablation designed to demonstrate the value of SQ-DPP actually weakens the paper's central contribution. Random prototype selection achieves AUC 0.6311, KMeans 0.6313, and RBF-DPP 0.6312, while SQ-DPP achieves 0.6318. The difference between SQ-DPP and random selection is only 0.0007 AUC, which is likely within noise given the absence of repeated runs. The paper should report uncertainty and a significance test for this comparison; as written, the claim that SQ-DPP provides a meaningful advantage over simpler selection strategies is not supported.
  3. [§3.2, §4.1.1, §4.2, §5.1.3] Every component of PID—the DPP kernel (Eq. 3), density quality (Eq. 4), and similarity gates (Eq. 9)—depends on cosine similarity in a frozen pretrained multimodal embedding space. The paper itself notes in Sec. 1 that such embeddings are not optimized for CTR behavior, yet it never varies the encoder or compares against task-tuned or behavior-derived embeddings. If this similarity space aligns poorly with user behavior, SQ-DPP may select prototypes that are content-similar but behaviorally irrelevant, and the gating mechanism would inherit that misalignment. This is a correctness-risk concern, not an a priori objection. A concrete test would be to rerun PID with a different pretrained encoder and, where feasible, with an interaction-tuned embedding; without such evidence, the robustness of the core premise is unknown.
  4. [§5.1.3 (Table 2)] Many free parameters—M, K, tau, qmax, eta, and the RQ structure—are tuned per dataset on validation performance. While tuning is common in this area, the paper does not describe the search protocol or report the sensitivity of the main results to the chosen values. Figure 2 shows trends for K and M, but no error bars or statistical comparisons. The concern is that the reported margins are small enough that per-dataset tuning could be selecting favorable settings rather than reflecting a robust algorithmic advantage. The authors should provide the tuning grid and, ideally, a nested-validation or sensitivity analysis for the key hyperparameters.
minor comments (5)
  1. [Table 2] The entry 'K = 12/15' is ambiguous: it presumably refers to the two backbones, but this is not stated. Please clarify the notation.
  2. [§5.3.3 (Fig. 3)] The PID 'weighted change rate' is not directly comparable to the SID 'change rate' because the former is normalized by softmax weights and the latter is a raw per-code change. The authors acknowledge this, but a sentence explaining why the comparison is fair would help.
  3. [Appendix A, Eq. (13)] The notation mixes superscript ℓ and parenthesized (ℓ) for the same level index; please make it consistent.
  4. [Table 9] The columns 'N' and 'Ñ' are not defined in the table or its caption. They appear to denote the average effective neighborhood size and median density, but this should be stated.
  5. [§5.1.1] KuaiRec is described as a multimodal dataset, but the implementation uses only BGE-M3 text embeddings for this dataset. Please clarify whether visual features are intentionally omitted or unavailable.

Circularity Check

0 steps flagged

No circular reasoning found: PID's prototype embeddings are trained on CTR data and evaluated on held-out benchmarks.

full rationale

The paper's derivation chain is not circular. PID's prototype palette is constructed offline from frozen multimodal embeddings via SQ-DPP (Eqs. 3-7); the only trainable components are the prototype lookup embeddings and the affine gate in Eq. (9), optimized by Eq. (11) on CTR labels. The claimed CTR improvements are measured on held-out test instances of TAOBAO-MM and KuaiRec against external baselines, not derived from the selection objective. Hyperparameters (M, K, tau, qmax) are tuned on validation splits, which is standard and does not make the test numbers fitted by construction. The paper does not rely on self-citations: references to DPP/RBF/greedy MAP are external, and the RQ-SID prefix-dependence claim is supported by the in-paper diagnostic in Appendix A rather than by an imported uniqueness theorem. The semantic facet coverage analysis uses the same embedding family as retrieval and is therefore partly self-confirming, but it is an interpretability illustration, not a load-bearing prediction; it does not reduce the central CTR claim to its inputs. No quoted equation or fitted parameter is renamed as a prediction.

Axiom & Free-Parameter Ledger

9 free parameters · 7 axioms · 0 invented entities

No new physical or structural entities are introduced: the 'prototype palette' is a subset of real items, and SQ-DPP is a selection algorithm. The main external pull is the assumption that frozen pretrained multimodal embeddings provide CTR-relevant semantics, plus a set of hand-tuned hyperparameters (gamma, R, tau, qmax, M, eta, K) that are selected per dataset on validation performance.

free parameters (9)
  • gamma (Cosine-RBF bandwidth) = median-distance heuristic over sampled item pairs (not numerically reported)
    Controls similarity decay in Eq. 3; chosen from data rather than fixed by theory; affects prototype selection.
  • R (ANN neighborhood size) = 50 (Table 2)
    Neighborhood size for the density score in Eq. 4; hand-set for both datasets.
  • tau (similarity threshold) = 0.5 (TAOBAO-MM), 0.3 (KuaiRec)
    Used in density Eq. 4 and retrieval filtering; selected by validation; Table 9 shows sensitivity.
  • qmax (quality clipping bound) = 1.8 (TAOBAO-MM), 1.2 (KuaiRec)
    Clips density quality in Eq. 5; tuned by validation; Table 9 shows AUC sensitivity.
  • M (palette size) = 650 (TAOBAO-MM), 200 (KuaiRec)
    Prototype budget; tuned by validation; Fig. 2b shows inverted-U sensitivity.
  • eta (retrieval filtering threshold) = 0.2 for both datasets
    Removes low-similarity prototypes before top-K retrieval in Sec. 4.2.
  • K (retrieval size) = 12/15 (TAOBAO-MM), 7 (KuaiRec)
    Number of prototypes composing each PID; tuned by validation; Fig. 2a shows inverted-U sensitivity.
  • W, b (affine gate parameters) = learned
    Eq. 9 maps similarity scores to sigmoid weights; trained end-to-end with CTR objective.
  • L_mono regularization coefficient = not reported (implicitly 1)
    Eq. 11 adds L_mono without a stated weight; no sensitivity analysis is provided.
axioms (7)
  • standard math The Cosine-RBF kernel in Eq. 3 is positive semi-definite and induces an implicit nonlinear feature space.
    Needed so that L = diag(q) K diag(q) in Eq. 6 remains a valid DPP kernel after the congruence transformation.
  • standard math Greedy incremental Cholesky DPP MAP inference (Chen et al. [5]) gives an adequate approximate solution to Eq. 7.
    The paper relies on this approximate solver for global prototype selection and cites it without independent verification here.
  • domain assumption Pretrained multimodal embeddings encode semantic content that is relevant to CTR prediction.
    The entire PID construction uses frozen embeddings for density (Eq. 4), SQ-DPP selection, and retrieval (Eq. 8); no sensitivity analysis is given if the embedding space is poorly aligned with user behavior.
  • domain assumption A small prototype palette (M << N) can cover the multimodal item space.
    Sec. 4.1.3 claims M is governed by semantic covering complexity rather than catalog size; only two dataset-specific M values are offered as empirical support.
  • domain assumption Real items can serve as semantic anchors whose learned embeddings transfer across all items that retrieve them.
    PID assumes shared prototype embeddings trained with the CTR objective generalize to every item sharing a prototype, which is the mechanism behind long-tail gains.
  • ad hoc to paper Median-density normalization and qmax clipping in Eq. 5 are appropriate for prototype quality.
    The quality score is introduced heuristically without derivation; Table 9 shows that AUC changes with these choices.
  • ad hoc to paper The monotonic regularization L_mono in Eq. 10 helps optimization when added with equal weight in Eq. 11.
    No coefficient is specified, and no ablation isolates the effect of this term.

pith-pipeline@v1.3.0-daily-deepseek · 15305 in / 14061 out tokens · 135253 ms · 2026-08-03T15:37:10.873469+00:00 · methodology

0 comments
read the original abstract

Multimodal information can improve the accuracy of click-through rate (CTR) prediction and effectively alleviate item cold-start and long-tail problems. Recent studies commonly discretize pretrained multimodal embeddings into semantic identifiers (SIDs), allowing the model to learn task-specific semantic representations for recommendation. However, existing methods still provide limited gains due to two major limitations. First, codebook assignment fails to preserve semantic relevance and discards fine-grained continuous signals in the original embedding space. Second, the residual code paths are highly dependent on prefix codes, which limits the effective representational scalability of hierarchical identifiers. To address these issues, we propose PaletteID (PID), a prototype-based semantic identifier. Inspired by palette-based color composition, PID uses a compact set of representative prototype items as semantic anchors to bridge pretrained multimodal content space and recommendation models. Specifically, we first construct a prototype palette with Semantic Quality-Aware Determinantal Point Process (SQ-DPP), which jointly considers local content density and global semantic diversity. Then, for each target item, PID retrieves a sequence of semantically related prototypes and aggregates them into an informative PID representation, enabling rich and complementary semantic modeling. Extensive experiments on two public datasets demonstrate that PID consistently improves CTR prediction and yields larger gains for long-tail items. PID also produces more robust identifier assignments and provides more interpretable token semantics than existing residual SID methods.

Figures

Figures reproduced from arXiv: 2607.29000 by Baining Chen, Huanyu Liu, Hui Liu, Zengyang Li, Ziyi Huang.

Figure 1
Figure 1. Figure 1: Overall architecture of PaletteID. The dashed boxes [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: AUC changes under different prototype sequence [PITH_FULL_IMAGE:figures/full_fig_p006_2.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

32 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    Oren Barkan, Noam Koenigstein, Eylon Yogev, and Ori Katz. 2019. CB2CF: a neural multiview content-to-collaborative filtering model for completely cold item recommendations. InACM Conference on Recommender Systems (RecSys). 228–236. https://doi.org/10.1145/3298689.3347038

    arXiv 2019

  2. [2]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin...

    2020

  3. [3]

    Shiteng Cao, Junda She, Ji Liu, Bing Zeng, Chengcheng Guo, Kuo Cai, Qiang Luo, Ruiming Tang, Han Li, Kun Gai, Zhiheng Li, and Cheng Yang. 2026. GLASS: A Generative Recommender for Long-sequence Modeling via SID-Tier and Se- mantic Search.CoRRabs/2602.05663 (2026). https://doi.org/10.48550/arXiv.2602. 05663

    Pith/arXiv arXiv doi:10.48550/arxiv.2602.05663 2026

  4. [4]

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation.CoRRabs/2402.03216 (2024). https://doi.org/10.48550/arXiv.2402.03216

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.03216 2024

  5. [5]

    Laming Chen, Guoxin Zhang, and Eric Zhou. 2018. Fast Greedy MAP Inference for Determinantal Point Process to Improve Recommen- dation Diversity. InAdvances in Neural Information Processing Systems (NeurIPS). 5627–5638. https://proceedings.neurips.cc/paper/2018/hash/ dbbf603ff0e99629dda5d75b6f75f966-Abstract.html

    2018

  6. [6]

    Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. 2025. OneRec: Unifying Retrieve and Rank with Genera- tive Recommender and Iterative Preference Alignment.CoRRabs/2502.18965 (2025). https://doi.org/10.48550/arXiv.2502.18965

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.18965 2025

  7. [7]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. InInternational Joint Conference on Artificial Intelligence (IJCAI). 1725–1731. https://doi.org/10.24963/ijcai.2017/239

    doi:10.24963/ijcai.2017/239 2017

  8. [8]

    Yanhua Huang, Weikun Wang, Lei Zhang, and Ruiwen Xu. 2021. Sliding Spectrum Decomposition for Diversified Recommendation. InACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD). 3041–3049. https://doi.org/10. 1145/3447548.3467108

    arXiv 2021

  9. [9]

    Zhenhua Huang, Xin Xu, Juan Ni, Honghao Zhu, and Cheng Wang. 2019. Multi- modal Representation Learning for Recommendation in Internet of Things.IEEE Internet Things J.6, 6 (2019), 10675–10685. https://doi.org/10.1109/JIOT.2019. 2940709

    doi:10.1109/jiot.2019 2019

  10. [10]

    Clark Mingxuan Ju, Liam Collins, Leonardo Neves, Bhuvesh Kumar, Louis Yufeng Wang, Tong Zhao, and Neil Shah. 2025. Generative Recommendation with Semantic IDs: A Practitioner’s Handbook. InACM International Conference on Information and Knowledge Management (CIKM). 6420–6425. https://doi.org/10. 1145/3746252.3761612

    arXiv 2025

  11. [11]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Optimization. InInternational Conference on Learning Representations (ICLR). http://arxiv.org/abs/1412.6980

    Pith/arXiv arXiv 2015

  12. [12]

    Qijiong Liu, Jieming Zhu, Yanting Yang, Quanyu Dai, Zhaocheng Du, Xiao-Ming Wu, Zhou Zhao, Rui Zhang, and Zhenhua Dong. 2024. Multimodal Pretraining, Adaptation, and Generation for Recommendation: A Survey. InACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD). 6566–6576. https: //doi.org/10.1145/3637528.3671473

    arXiv 2024

  13. [14]

    Odile Macchi. 1975. The coincidence approach to stochastic point processes. Advances in Applied Probability7, 1 (1975), 83–122

    1975

  14. [15]

    Yongheng Mu and Yun Wu. 2023. Multimodal movie recommendation system using deep learning.Mathematics11, 4 (2023), 895

    2023

  15. [16]

    Pingjun Pan, Tingting Zhou, Peiyao Lu, Tingting Fei, Hongxiang Chen, and Chuanjiang Luo. 2026. Hi-SAM: A Hierarchical Structure-Aware Multi-modal Framework for Large-Scale Recommendation.CoRRabs/2602.11799 (2026). https: //doi.org/10.48550/arXiv.2602.11799

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2602.11799 2026

  16. [17]

    Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Jun- jie Bai, and Soumith Chintala

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gre- gory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Z. Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Jun- jie Bai, and Soumith Chintala. 2019. PyTorch: An Imperative Styl...

    2019

  17. [18]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Mod- els From Natural Language Supervision. InInternational Conference on Ma- chine Learning, (ICML), Vol. 139. 8748–8763. http://procee...

    2021

  18. [19]

    Tran, Jonah Samost, Maciej Kula, Ed H

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Mahesh Sathiamoorthy. 2023. Recommender Systems with Generative Retrieval. InAdvances in Neural Information Pro- cessing Systems (NeurIPS). http://papers.nips.cc/paper_files/paper/20...

    2023

  19. [20]

    Xiang-Rong Sheng, Feifan Yang, Litong Gong, Biao Wang, Zhangming Chan, Yujing Zhang, Yueyao Cheng, Yong-Nan Zhu, Tiezheng Ge, Han Zhu, Yuning Jiang, Jian Xu, and Bo Zheng. 2024. Enhancing Taobao Display Advertising with Multimodal Representations: Challenges, Approaches and Insights. InACM International Conference on Information and Knowledge Management (...

    arXiv 2024

  20. [21]

    Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. 2017. Neural Discrete Representation Learning. InAdvances in Neural Information Processing Systems (NeurIPS). 6306–6315. https://proceedings.neurips.cc/paper/2017/hash/ 7a98af17e63a0ac09ce2e96d03992fbc-Abstract.html

    2017

  21. [22]

    Ruoxi Wang, Rakesh Shivanna, Derek Zhiyuan Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed H. Chi. 2021. DCN V2: Improved Deep & Cross Network and Practical Lessons for Web-scale Learning to Rank Systems. InThe Web Conference (WWW). 1785–1797. https://doi.org/10.1145/3442381.3450078

    arXiv 2021

  22. [23]

    Yichao Wang, Xiangyu Zhang, Zhirong Liu, Zhenhua Dong, Xinhua Feng, Ruiming Tang, and Xiuqiang He. 2020. Personalized Re-ranking for Improv- ing Diversity in Live Recommender Systems.CoRRabs/2004.06390 (2020). https://arxiv.org/abs/2004.06390

    Pith/arXiv arXiv 2020

  23. [24]

    Chi, and Jennifer Gillenwater

    Mark Wilhelm, Ajith Ramanathan, Alexander Bonomo, Sagar Jain, Ed H. Chi, and Jennifer Gillenwater. 2018. Practical Diversified Recommendations on YouTube with Determinantal Point Processes. InACM International Confer- ence on Information and Knowledge Management (CIKM). 2165–2173. https: //doi.org/10.1145/3269206.3272018

    arXiv 2018

  24. [25]

    Bin Wu, Feifan Yang, Zhangming Chan, Yu-Ran Gu, Jiawei Feng, Chao Yi, Xiang- Rong Sheng, Han Zhu, Jian Xu, Mang Ye, and Bo Zheng. 2025. MUSE: A Simple Yet Effective Multimodal Search-Based Framework for Lifelong User Interest Modeling.CoRRabs/2512.07216 (2025). https://doi.org/10.48550/arXiv.2512. 07216

    doi:10.48550/arxiv.2512.07216 2025

  25. [26]

    Yi Xu, Chaofan Fan, Jinxin Hu, Yu Zhang, Xiaoyi Zeng, and Jing Zhang. 2025. STORE: Semantic Tokenization, Orthogonal Rotation and Efficient Attention for Scaling Up Ranking Models.CoRRabs/2511.18805 (2025). https://doi.org/10. 48550/arXiv.2511.18805

    doi:10.48550/arxiv.2511.18805 2025

  26. [27]

    Yining Yao, Ziwei Li, Shuwen Xiao, Boya Du, Jialin Zhu, Junjun Zheng, Xiangheng Kong, and Yuning Jiang. 2025. SaviorRec: Semantic-Behavior Alignment for Cold-Start Recommendation.CoRRabs/2508.01375 (2025). https://doi.org/10. 48550/arXiv.2508.01375

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.01375 2025

  27. [28]

    Wencai Ye, Mingjie Sun, Shaoyun Shi, Peng Wang, Wenjin Wu, and Peng Jiang

  28. [29]

    Yu Ye, Junchen Fu, Yu Song, Kaiwen Zheng, and Joemon M. Jose. 2026. Are Multimodal Embeddings Truly Beneficial for Recommendation? A Deep Dive into Whole vs. Individual Modalities. InEuropean Conference on Information Retrieval (ECIR). 66–81. https://doi.org/10.1007/978-3-032-21324-2_5

    work page doi:10.1007/978-3-032-21324-2_5 2026

  29. [30]

    Carolina Zheng, Minhui Huang, Dmitrii Pedchenko, Kaushik Rangadurai, Siyu Wang, Fan Xia, Gaby Nahum, Jie Lei, Yang Yang, Tao Liu, Zutian Luo, Xiaohan Wei, Dinesh Ramasamy, Jiyan Yang, Yiping Han, Lin Yang, Hangjun Xu, Rong Jin, and Shuang Yang. 2025. Enhancing Embedding Representation Stability in Recommendation Systems with Semantic ID. InACM Conference ...

    arXiv 2025

  30. [31]

    Jiawei Zheng, Hao Gu, Lingling Yi, Jie Wen, and Chuan Chen. 2025. Personalized Multi Modal Alignment Encoding for CTR-Recommendation in WeChat. InACM International Conference on Information and Knowledge Management (CIKM). 6301–6308. https://doi.org/10.1145/3746252.3761525

    arXiv 2025

  31. [32]

    Guorui Zhou, Xiaoqiang Zhu, Chengru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep Interest Network for Click-Through Rate Prediction. InACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD). 1059–1068. https://doi.org/10. 1145/3219819.3219823

    arXiv 2018

  32. [33]

    Jie Zhu, Zhifang Fan, Xiaoxie Zhu, Yuchen Jiang, Hangyu Wang, Xintian Han, Haoran Ding, Xinmin Wang, Wenlin Zhao, Zhen Gong, Huizhi Yang, Zheng Chai, Zhe Chen, Yuchao Zheng, Qiwei Chen, Feng Zhang, Xun Zhou, Peng Xu, Xiao Yang, Di Wu, and Zuotao Liu. 2025. RankMixer: Scaling Up Ranking Models in Industrial Recommenders. InACM International Conference on I...

    doi:10.1145/3746252 2025