Pith. sign in

REVIEW 4 major objections 5 minor 32 references

A tree-based retrieval model that indexes items by multi-modal similarity and searches 2000–4000-item user histories lifts offline recall by 31–47% over the best baseline.

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 ↗

MISS builds a k-means index tree on interaction-supervised multi-modal embeddings and adds two behavior search units (Co-GSU, MM-GSU) plus ESU/MMoE, reporting ~30-47% relative recall gains over TDM+MMoE on Kuaishou data and +0.248% total app usage time online.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection MISS is a credible industrial system with a plausible design, but the headline gains are confounded by the pre-trained interaction-aligned embeddings that no baseline receives, so the architecture's contribution is not isolated. the 4 major comments →

arxiv 2508.14515 v1 pith:ODYCMDE5 submitted 2025-08-20 cs.IR cs.AI

MISS: Multi-Modal Tree Indexing and Searching with Lifelong Sequential Behavior for Retrieval Recommendation

Chengcheng Guo , Junda She , Kuo Cai , Shiyao Wang , Qigen Hu , Qiang Luo , Kun Gai , Guorui Zhou This is my paper
classification cs.IR cs.AI
keywords recommender systemsmulti-modal recommendationtree-based retrievallifelong sequential behaviorgeneral search unitmulti-task learningbeam searchonline A/B test
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

Large-scale recommendation pipelines lose most candidate quality at the retrieval stage, yet retrieval models rarely get to use either long user histories or multi-modal item content. This paper tries to change that by proposing MISS, a tree-based retriever whose index is built from frozen multi-modal embeddings and whose node estimator searches a user's lifelong behavior sequence, thousands of items long, for behaviors relevant to each candidate. The core claim is that using the same multi-modal embedding space for both tree structure and behavior search, side by side with a collaborative search unit, gives retrieval both content-level and interaction-level interest matching. On industrial offline data, MISS reports recall@K improvements of 31–47% over the strongest baseline, and an online A/B test reports a 0.248% increase in total app usage time with a confidence interval that excludes zero. If this holds, retrieval can be upgraded without abandoning the efficient tree-and-beam-search paradigm.

Core claim

The paper's discovery, on its own terms, is that the tree-retrieval paradigm can absorb multi-modal information and lifelong sequences at the same time, and that the two reinforce each other. The multi-modal index tree is built by recursive k-means over pre-trained multi-modal embeddings, content vectors aligned with item2item interaction signals, so similar items share subtrees. When a user arrives, beam search walks this tree, and at each candidate node the model selects top-K behaviors from up to 4,000 past interactions using Co-GSU, which scores with ID-embedding target attention, and MM-GSU, which scores with dot products in the same frozen embedding space. The selected behaviors are co

What carries the argument

The load-bearing machinery is the pair of general search units coupled with a k-means index tree. Co-GSU scores each behavior from the lifelong sequence against a candidate node using target attention in ID-embedding space; MM-GSU scores them as dot products in the frozen multi-modal embedding space, reusing the same vector geometry that built the tree. Top-K behaviors from each unit form sub-sequences, and ESU applies target attention over them; MMoE turns the fused representation into multiple feedback predictions. The design exploits one trick: because the multi-modal embeddings are aligned to item2item similarity, they can serve simultaneously as index coordinates and as queries for beha

Load-bearing premise

The paper's own Section 3 says it does not train multi-modal embeddings and instead freezes pre-trained vectors aligned to item2item retrieval signals; the whole method depends on those frozen vectors being a stable, unbiased measure of item similarity for both the tree structure and the behavior search. If that space drifts or encodes the wrong notion of similarity, the central gains collapse.

What would settle it

A concrete test: rebuild the same index tree and MM-GSU using randomly permuted multi-modal embeddings of the same dimension, leaving Co-GSU and ESU training intact. If recall stays near the full MISS level, the claimed gains are not caused by multi-modal content structure. A second test: evaluate a tree trained on day-one frozen embeddings against day-seven logs; a sharp drop in MISS's advantage over TDM+MMoE would show the method inherits embedding staleness.

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

If this is right

  • Retrieval can consume behavior histories thousands of items long without scanning the candidate corpus; the tree prunes items, and the GSUs prune history.
  • Improvements in pre-trained multi-modal representations should transfer directly to retrieval quality, since the same frozen embeddings drive both tree geometry and MM-GSU scoring.
  • Collaborative and multi-modal search retrieve different behaviors, with the paper reporting only 13% overlap, so combining them gives broader interest coverage than either alone.
  • Building the tree from multi-modal embeddings avoids the TDM step of rebuilding the index from interaction-trained ID embeddings, which the paper argues loses content information.
  • The gains appear at intermediate tree levels, meaning better candidates survive early beam-search pruning, not only at the final leaves.

Where Pith is reading between the lines

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

  • A risk the paper leaves implicit: tree and MM-GSU share one frozen embedding space, so drift or bias in that space degrades both components at once; a time-shifted evaluation would expose this.
  • A testable extension is gating or routing between Co-GSU and MM-GSU per candidate instead of concatenating fixed top-K outputs, exploiting the low overlap to save computation.
  • The sequence-length scaling curve from 1k to 4k suggests an adaptive per-user length policy could trade compute for recall more cheaply than a uniform 4k sequence.
  • Because the embeddings are aligned with item2item supervision, MISS inherits any popularity or co-occurrence bias of that supervision; aligning the embeddings with retrieval-specific feedback is the natural next training target.
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

4 major / 5 minor

Summary. The paper proposes MISS, a retrieval-stage model that combines a multi-modal index tree with lifelong sequential behavior modeling. The tree is built by k-means clustering on frozen multi-modal item embeddings (Eq. 3), and the lifelong sequence module uses two general search units (Co-GSU and MM-GSU) with target attention, followed by an Exact Search Unit and MMoE multi-task learning. Offline experiments on proprietary Kuaishou data report large recall gains over five baselines (Table 1), an ablation study supporting the main components (Fig. 4), and an online A/B test showing +0.248% Total App Usage Time (95% CI [0.16%, 0.34%], Table 3). The authors claim the first use of multi-modal information in tree-based retrieval.

Significance. If the evidence were sufficient, MISS would be a valuable industrial case study of injecting pre-trained multi-modal item embeddings into tree-based retrieval and of using lifelong sequences at the retrieval stage. The paper's strengths are its large-scale deployment, the inclusion of online A/B confidence intervals, the standard and internally consistent pseudo-label/beam-search formulation, and the hierarchical recall metric for intermediate tree levels. However, because the multi-modal embeddings are pre-trained with item2item interaction supervision and are not provided to any baseline, the current experiments do not isolate the contribution of the proposed indexing/searching architecture from the contribution of the embedding itself. No public dataset or code is provided, which limits reproducibility but is not disqualifying for an industrial paper.

major comments (4)
  1. [§3, §4.1, Table 1] The frozen multi-modal embeddings are trained with item2item retrieval supervision: Eq. (2) uses InfoNCE on item pairs selected from an item2item retrieval model. These embeddings are then used in Eq. (3) to build the tree and in r_mm_i = z_n^T z_i for MM-GSU scoring. They therefore already encode collaborative item similarity, as the paper acknowledges ('both content modality and interaction modality'). Table 1 compares MISS only against baselines (SASRec, NANN, Kuaiformer, TDM, TDM+MMoE) that do not receive these embeddings, so the reported +37.93% recall@800 conflates the benefit of injecting a strong pre-trained collaborative embedding with the benefit of MISS's tree/GSU architecture. The paper needs a control: e.g., a two-tower/ANN retriever or a TDM variant using the same frozen embeddings, and an ablation that keeps the GSUs while replacing only the multi-modal tree with an ID-bas
  2. [§6.3, Fig. 4] The ablation is cumulative: version1 removes MM-GSU, version2 removes Co-GSU, version3 removes the multi-modal index tree. Thus version3 differs from the full model by removing three components simultaneously. Fig. 4 cannot be used to conclude that 'removing each module would lead to performance degradation' for each module individually, because the tree contribution is never isolated while keeping the GSUs. A clean 2x2 (tree type × GSU presence) is needed, e.g., full model, full model with ID-based tree, full model without MM-GSU, full model without Co-GSU, and the combination. This is a load-bearing gap for the ablation claim.
  3. [§6.2.1, Table 1] The sentence 'Compared with the second-best results, our method achieves an improvement of 37.93%, 30.77% and 47.37%' is ambiguous because the second-best baseline differs across metrics. In Table 1, the second-best for Recall@800 is TDM+MMoE (0.29) and for Recall@600 it is TDM+MMoE (0.26), but for Recall@400 the second-best is Kuaiformer (0.19), not TDM+MMoE (0.18). The 47.37% figure is relative to Kuaiformer. Please state the comparison baseline per metric. In addition, no significance tests or confidence intervals are reported for the offline recalls; the reported standard deviations should be accompanied by a statement of whether differences are statistically reliable.
  4. [§6.5, Table 3] The rows labeled 'TDM(w/o MM tree)' and 'TDM(w/o MM searching)' are not defined in the text. If these are ablations of the proposed system, they should be described as such and connected to the offline ablations in Fig. 4; if they are TDM baselines with one component removed, then the online A/B comparison does not isolate the proposed contributions. Clarify the relation between these rows and the proposed method, and report the exact configuration of each online arm.
minor comments (5)
  1. [Throughout] Typos and wording issues: 'relay on' should be 'rely on' (Abstract); 'modulde' should be 'module' (Contributions); 'Coollaborative' should be 'Collaborative' (Sec. 4.2); 'sample multiply nodes' should be 'sample multiple nodes' (Sec. 4.1); 'pxtr' in Fig. 1 caption appears to be a typo; 'mass real industrial data' (Sec. 6.1.3) is informal; Conclusion says 'comparable performance' but the reported results are 'superior'.
  2. [§4.2, Eq. (13)] There is a contradiction about parameter sharing: the text says 'the attention module of MM-SBS, Co-SBS and Co-GSU shares parameters', but the paragraph after Eq. (13) says W_mm_q, W_mm_k, W_mm_v are 'not shared parameters with Co-SBS'. Clarify which modules share parameters and which do not.
  3. [§6.1.3] The offline dataset description is too thin for reproducibility: give the number of users/videos, feature dimensions, the number of testing datasets sampled, and the exact sampling procedure. The paper relies on proprietary data, so additional details matter.
  4. [§6.4.2, Figs. 5-6] The attention maps and histograms lack axis labels and a color scale, which makes the visual claim difficult to verify. Please describe how the maps are aggregated across users and what the color intensity represents.
  5. [References] The ACM reference format is a placeholder ('Conference acronym ’XX, Woodstock, NY', '© 2018', etc.) and should be corrected to the actual venue and year.

Circularity Check

0 steps flagged

No significant circularity: the multi-modal embeddings are an external pre-trained input, the tree and GSU use them as frozen features, and the retrieval predictions are produced by a separately trained node estimator.

full rationale

The paper's derivation chain is not circular. The multi-modal embeddings are pre-trained in Section 3 via item alignment (InfoNCE) on item pairs from an item2item retrieval model; this is an external pretraining step that the paper explicitly states is out of scope ('Since this paper does not focus on the training of multi-modal embeddings, we directly use item alignment[18] to train multi-modal embeddings'). These embeddings are then used as frozen inputs to (a) construct the index tree by k-means (Eq. 3) and (b) compute MM-GSU behavior-selection scores r_mm_i = z_n^T z_i. Neither step fits a parameter to the evaluation labels. The final retrieval decision is made by the node estimator F_theta (Eq. 7), trained with BCE on pseudo labels (Eq. 4-6) derived from user feedback, and the beam search (Eq. 8-10) uses F_theta's predicted probabilities, not the raw embedding dot products. Thus the recall@K results are not equal to the embedding similarity by construction. The self-citation to QARM[18] is transparent and non-load-bearing: the paper takes QARM's embedding training as an external dependency, does not invoke a uniqueness theorem, and does not present QARM's results as the evidence for MISS. The lack of a baseline that receives the same pre-trained embeddings is a fair experimental-control concern, but it is a confounding-variable issue, not a circularity. Consequently no step satisfies the standard of exhibiting Eq. X = Eq. Y by construction or a fitted parameter renamed as a prediction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The method depends on frozen multi-modal embeddings trained elsewhere (QARM), on the TDM tree assumption, and on the effectiveness of GSU-style top-k searching. No new physical or conceptual entities are introduced; Co-GSU and MM-GSU are architectural modules built from standard attention components.

free parameters (4)
  • Co-GSU sequence length M_co = 2000
    Chosen in Sec 6.1.4 / 6.4.1 to balance performance and GPU cost; not learned.
  • MM-GSU sequence length M_mm = 4000
    Set to 4000 after trade-off analysis; longer sequences give small recall gains with doubled GPU cost (Sec 6.4.1).
  • ESU Top-K = 50
    Number of behaviors selected by GSUs as ESU input; set in Sec 6.1.4.
  • Tree height H = 22
    Binary tree with 22 levels, chosen in Sec 6.1.4.
axioms (4)
  • domain assumption The pre-trained multi-modal embeddings from QARM item alignment are fixed and capture item similarity.
    Section 3 states 'we directly use item alignment[18] to train multi-modal embeddings' and they are frozen in MM-GSU; the paper does not validate the embedding quality independently.
  • standard math Tree-based max-heap pseudo-label construction (Eq. 4) allows level-wise BCE training and beam-search inference to retrieve relevant leaves.
    Adopted from TDM[28]; the paper relies on the property without re-deriving it.
  • domain assumption Top-K scoring in GSUs selects behaviors sufficiently relevant to the candidate node; the attention scores in ESU can be reused from Co-GSU scores.
    Section 4.2 assumes the dot-product/target-attention scores are a valid relevance measure for lifelong behavior.
  • domain assumption Online A/B metrics (Total App Usage Time, App Usage Time Per User) are a valid proxy for recommendation quality.
    Section 6.5 uses these metrics without discussing potential confounding factors (e.g., novelty effects, other channels).

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of MISS: Multi-Modal Tree Indexing and Searching with Lifelong Sequential Behavior for Retrieval Recommendation." pith.science (2026). https://pith.science/paper/ODYCMDE5

@misc{pith2026250814515,
  author       = {Pith},
  title        = {Pith review of: MISS: Multi-Modal Tree Indexing and Searching with Lifelong Sequential Behavior for Retrieval Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ODYCMDE5}},
  note         = {Machine review of arXiv:2508.14515}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large-scale industrial recommendation systems typically employ a two-stage paradigm of retrieval and ranking to handle huge amounts of information. Recent research focuses on improving the performance of retrieval model. A promising way is to introduce extensive information about users and items. On one hand, lifelong sequential behavior is valuable. Existing lifelong behavior modeling methods in ranking stage focus on the interaction of lifelong behavior and candidate items from retrieval stage. In retrieval stage, it is difficult to utilize lifelong behavior because of a large corpus of candidate items. On the other hand, existing retrieval methods mostly relay on interaction information, potentially disregarding valuable multi-modal information. To solve these problems, we represent the pioneering exploration of leveraging multi-modal information and lifelong sequence model within the advanced tree-based retrieval model. We propose Multi-modal Indexing and Searching with lifelong Sequence (MISS), which contains a multi-modal index tree and a multi-modal lifelong sequence modeling module. Specifically, for better index structure, we propose multi-modal index tree, which is built using the multi-modal embedding to precisely represent item similarity. To precisely capture diverse user interests in user lifelong sequence, we propose collaborative general search unit (Co-GSU) and multi-modal general search unit (MM-GSU) for multi-perspective interests searching.

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 · 11 canonical work pages

  1. [1]

    Xingyan Bin, Jianfei Cui, Wujie Yan, Zhichen Zhao, Xintian Han, Chongyang Yan, Feng Zhang, Xun Zhou, Qi Wu, and Zuotao Liu. 2025. Real-time Indexing for Large-scale Recommendation by Streaming Vector Quantization Retriever. arXiv preprint arXiv:2501.08695 (2025)

    Pith/arXiv arXiv 2025

  2. [2]

    Yue Cao, Xiaojiang Zhou, Jiaqi Feng, Peihao Huang, Yao Xiao, Dayao Chen, and Sheng Chen. 2022. Sampling is all you need on modeling long-term user behaviors for CTR prediction. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management . 2974–2983

    2022

  3. [3]

    Jianxin Chang, Chenbin Zhang, Zhiyi Fu, Xiaoxue Zang, Lin Guan, Jing Lu, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, et al. 2023. TWIN: TWo-stage interest network for lifelong user behavior modeling in CTR prediction at kuaishou. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3785–3794

    2023

  4. [4]

    Qiwei Chen, Changhua Pei, Shanshan Lv, Chao Li, Junfeng Ge, and Wenwu Ou

  5. [5]

    Rihan Chen, Bin Liu, Han Zhu, Yaoxuan Wang, Qi Li, Buting Ma, Qingbo Hua, Jun Jiang, Yunlong Xu, Hongbo Deng, et al. 2022. Approximate nearest neighbor search under neural similarity metric for large-scale recommendation. In Pro- ceedings of the 31st ACM International Conference on Information & Knowledge Management. 3013–3022

    work page 2022

  6. [6]

    Hao Deng, Haibo Xing, Kanefumi Matsuyama, Moyu Zhang, Jinxin Hu, Hong Wen, Yu Zhang, Xiaoyi Zeng, and Jing Zhang. 2025. CSMF: Cascaded Selective Mask Fine-Tuning for Multi-Objective Embedding-Based Retrieval. SIGIR 2025 (2025)

    work page 2025

  7. [7]

    Chen Gao, Xiangnan He, Dahua Gan, Xiangning Chen, Fuli Feng, Yong Li, Tat- Seng Chua, Lina Yao, Yang Song, and Depeng Jin. 2019. Learning to recommend with multiple cascading behaviors. IEEE transactions on knowledge and data engineering 33, 6 (2019), 2588–2601

    work page 2019

  8. [8]

    Weihao Gao, Xiangjun Fan, Chong Wang, Jiankai Sun, Kai Jia, Wenzi Xiao, Ruofan Ding, Xingyan Bin, Hui Yang, and Xiaobing Liu. 2021. Learning an end-to-end structure for retrieval in large-scale recommendations. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management . 524–533

    work page 2021

  9. [9]

    Zhicheng He, Weiwen Liu, Wei Guo, Jiarui Qin, Yingxue Zhang, Yaochen Hu, and Ruiming Tang. 2023. A survey on user behavior modeling in recommender systems. arXiv preprint arXiv:2302.11087 (2023)

    Pith/arXiv arXiv 2023

  10. [10]

    Junjie Huang, Jizheng Chen, Jianghao Lin, Jiarui Qin, Ziming Feng, Weinan Zhang, and Yong Yu. 2024. A Comprehensive Survey on Retrieval Methods in Recommender Systems. arXiv preprint arXiv:2407.21022 (2024)

    work page arXiv 2024

  11. [11]

    Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. 2013. Learning deep structured semantic models for web search using clickthrough data. In Proceedings of the 22nd ACM international conference on Information & Knowledge Management . 2333–2338

    2013

  12. [12]

    Wang-Cheng Kang, Chen Fang, Zhaowen Wang, and Julian McAuley. 2017. Visually-aware fashion recommendation and design with generative image mod- els. In 2017 IEEE international conference on data mining (ICDM) . IEEE, 207–216

    work page 2017

  13. [13]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In 2018 IEEE international conference on data mining (ICDM) . IEEE, 197–206

    2018

  14. [14]

    Chi Liu, Jiangxia Cao, Rui Huang, Kai Zheng, Qiang Luo, Kun Gai, and Guorui Zhou. 2024. KuaiFormer: Transformer-Based Retrieval at Kuaishou.arXiv preprint arXiv:2411.10057 (2024)

    Pith/arXiv arXiv 2024

  15. [15]

    Qidong Liu, Jiaxi Hu, Yutian Xiao, Xiangyu Zhao, Jingtong Gao, Wanyu Wang, Qing Li, and Jiliang Tang. 2024. Multimodal recommender systems: A survey. Comput. Surveys 57, 2 (2024), 1–17

    2024

  16. [16]

    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. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 6566–6576

    work page 2024

  17. [17]

    Yifan Liu, Kangning Zhang, Xiangyuan Ren, Yanhua Huang, Jiarui Jin, Yingjie Qin, Ruilong Su, Ruiwen Xu, Yong Yu, and Weinan Zhang. 2024. AlignRec: Aligning and Training in Multimodal Recommendations. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management . 1503–1512

    2024

  18. [18]

    Xinchen Luo, Jiangxia Cao, Tianyu Sun, Jinkai Yu, Rui Huang, Wei Yuan, Hezheng Lin, Yichen Zheng, Shiyao Wang, Qigen Hu, et al . 2024. QARM: Quantita- tive Alignment Multi-Modal Recommendation at Kuaishou. arXiv preprint arXiv:2411.11739 (2024)

    Pith/arXiv arXiv 2024

  19. [19]

    Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. 2018. Modeling task relationships in multi-task learning with multi-gate mixture-of- experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining . 1930–1939

    work page 2018

  20. [20]

    Yu A Malkov and Dmitry A Yashunin. 2018. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE transactions on pattern analysis and machine intelligence 42, 4 (2018), 824–836

    2018

  21. [21]

    Qi Pi, Weijie Bian, Guorui Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Practice on long sequential user behavior modeling for click-through rate prediction. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining . 2671–2679

    2019

  22. [22]

    Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. 2020. Search-based user interest modeling with lifelong sequential behavior data for click-through rate prediction. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management . 2685–2692

    2020

  23. [23]

    Jiarui Qin, Weinan Zhang, Xin Wu, Jiarui Jin, Yuchen Fang, and Yong Yu. 2020. User behavior retrieval for click-through rate prediction. SIGIR 2020

    work page 2020

  24. [24]

    Xiang-Rong Sheng, Feifan Yang, Litong Gong, Biao Wang, Zhangming Chan, Yujing Zhang, Yueyao Cheng, Yong-Nan Zhu, Tiezheng Ge, Han Zhu, et al

  25. [25]

    Guorui Zhou, Xiaoqiang Zhu, Chenru 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. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining . 1059–1068

    2018

  26. [26]

    Xin Zhou, Hongyu Zhou, Yong Liu, Zhiwei Zeng, Chunyan Miao, Pengwei Wang, Yuan You, and Feijun Jiang. 2023. Bootstrap latent representations for multi- modal recommendation. In Proceedings of the ACM web conference 2023. 845–854

    work page 2023

  27. [27]

    Han Zhu, Daqing Chang, Ziru Xu, Pengye Zhang, Xiang Li, Jie He, Han Li, Jian Xu, and Kun Gai. 2019. Joint optimization of tree-based index and deep model for recommender systems. Advances in Neural Information Processing Systems 32 (2019)

    2019

  28. [28]

    Han Zhu, Xiang Li, Pengye Zhang, Guozheng Li, Jie He, Han Li, and Kun Gai

  29. [29]

    Jingwei Zhuo, Ziru Xu, Wei Dai, Han Zhu, Han Li, Jian Xu, and Kun Gai. 2020. Learning optimal tree models under beam search. In International Conference on Machine Learning. PMLR, 11650–11659

    work page 2020

  30. [2018]

    In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining

    Learning tree-based deep model for recommender systems. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1079–1088

  31. [2021]

    arXiv preprint arXiv:2108.04468 (2021)

    End-to-end user behavior retrieval in click-through rateprediction model. arXiv preprint arXiv:2108.04468 (2021)

    Pith/arXiv arXiv 2021

  32. [2024]

    InProceedings of the 33rd ACM International Conference on Information and Knowledge Management

    Enhancing Taobao Display Advertising with Multimodal Representations: Challenges, Approaches and Insights. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management . 4858–4865

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.