Pith. sign in

REVIEW 4 major objections 6 minor 35 references

IU4Rec: Interest Unit-Based Product Organization and Recommendation for E-Commerce Platform

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper proposes IU4Rec, an interest unit-based two-stage recommendation paradigm for C2C platforms, and claims it outperforms product-centric baselines by aggregating user interactions at a persistent cluster level rather than on…

desk verdict A production-scale C2C recommender that aggregates behavior at persistent interest units; worth a serious referee, but one internal inconsistency and a black-box GSID dependency need fixing. read the letter →

arxiv 2502.07658 v1 pith:P6VQ2OGU submitted 2025-02-11 cs.IR

Wenhao Wu , Xiaojie Li , Lin Wang , Jialiang Zhou , Di Wu , Qinye Xie , Qingheng Zhang , Yin Zhang
show 3 more authors
Shuguang Han Fei Huang Junfeng Chen
This is my paper · ORCID
classification cs.IR
keywords interestunitstwo-stagerecommendationlimited-stockproductsCTRpredictionC2Ce-commercecoldstartIU-BoostedNetworkuserbehaviorsequence
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that an interest unit-centric recommendation paradigm beats product-centric recommendation on C2C platforms with limited-stock goods. The core idea is that a product's interactions die with the listing, but interactions aggregated at the level of a persistent cluster continue to accumulate and support new listings. To use this, the paper groups products into Interest Units (IUs) via structured attributes, images, and semantic IDs, then redesigns the recommendation flow into two stages: first surface the IU, then rank products inside it. The IU-Boosted Network feeds IU-level statistics, user-IU cross features, and hierarchical IU click sequences into a DIN-style attention model. On Xianyu production data the paper reports a 1.91% AUC and 2.17% GAUC gain over DIN offline, and an 11.76% CTR and 5.61% Bills gain in online A/B tests for interest-unit recommendations, stating the system is fully deployed.

What carries the argument

The central object is the Interest Unit (IU), a persistent cluster of similar products defined by standardized product attributes (SPU units), image similarity, or a hierarchical semantic ID from the generative GSID system. The IU-Boosted Network is the model mechanism: it constructs IU-level statistical features and user-IU cross features, builds hierarchical IU click sequences from users' historical behaviors, and applies an attention mechanism that scores target products by the distance between their IU ID and the IU IDs of previously clicked products. The key work this machinery does is to let signals accumulate and remain accessible at the IU level even when individual listings are sold out, turning sparse, ephemeral item interactions into denser, persistent cluster-level behavior.

What would settle it

The claim would be falsified if replacing the semantic IU assignments with random clusters of equal size left the offline AUC and GAUC gains essentially unchanged, since then the gains would not come from meaningful persistence at the interest-unit level; it would also be falsified if re-running the online A/B test with the IU-Boosted model but the old single-stage interface produced the same lift in the general product domain.

Watch

Extended reading notes

Core claim

The central claim is that on limited-stock C2C platforms, the item-centric recommendation paradigm wastes accumulated user interactions because each listing disappears once sold. The paper argues the fix is to organize products into persistent Interest Units and make the IU itself the unit of recommendation: stage one recommends interest units to capture broad demand, and stage two guides the user to the best product within the chosen unit. It then claims that the IU-Boosted Network, which aggregates user behaviors under shared IU IDs and models hierarchical IU click sequences, achieves the best offline performance on the production dataset among all compared baselines and shows consistent online gains across interest-unit, general-product, and overall recommendation domains. The paper further claims that this two-stage paradigm changes user navigation so that interactions accumulate on clusters that outlive individual sold-out products, directly mitigating the limited-stock cold-start problem.

Load-bearing premise

The semantic Interest Units come from an unpublished GSID system whose three-level semantic IDs are taken as stable, meaningful clusters that match how users actually group products in their minds.

Editorial extensions

If this is right

  • If interactions persist at the IU level, new listings can inherit accumulated demand signals without needing a separate traffic channel for cold-start items.
  • The reported gains span both interest-unit and normal-product recommendation domains, suggesting the IU signal helps even when the recommended item is not explicitly organized by IU.
  • The two-stage interface changes how users browse, concentrating clicks within a cluster and potentially shortening purchase decisions, which the paper treats as a product-format benefit beyond model gains.
  • The paradigm is not inherently limited to C2C platforms; the paper argues it can extend to other e-commerce settings with high listing turnover or fragmented supply.

Reading between the lines

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

  • If IU-level behavior is truly persistent, then the economic value of a click or bill is no longer tied to a single listing's lifespan; platforms could plausibly price, allocate, or model traffic at the IU level rather than the item level.
  • A natural decomposition experiment would separate the interface contribution from the model contribution: running the IU-Boosted Network on the old single-stage interface would reveal how much of the 11.76% CTR gain comes from the two-stage format versus the IU features themselves.
  • Because the semantic GSID clusters come from an unpublished external system, a robustness test would vary clustering granularity or inject cluster-ID noise to see how the reported AUC and GAUC gains degrade; the current paper cannot distinguish gains from semantic quality versus mere grouping persistence.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes IU4Rec, a recommendation paradigm for C2C platforms in which individual listings are aggregated into Interest Units (IUs) built from product attributes, images, and query-derived semantic IDs (GSID). The system recommends IUs in a first stage and specific products within a selected IU in a second stage, and the IU-Boosted Network incorporates IU-level statistical features, hierarchical IU click sequences, and IU-target attention into a DIN-style CTR model. The authors report offline experiments on a 10-billion-sample production dataset and online A/B tests, claiming consistent gains in AUC/GAUC and CTR/Clicks/Bills, and state that the system is fully deployed on Alibaba's Xianyu platform.

Significance. If the results hold, the IU-based paradigm is a valuable contribution to limited-stock and C2C recommendation: it converts ephemeral item-level interactions into persistent IU-level signals, addresses the cold-start problem for sold-out items, and includes a product-format redesign. The evidence base is strong for an industrial paper: a multi-day 10B-sample production dataset, a held-out test day, and online A/B tests. At the same time, the paper's central mechanism depends on the semantic GSID grouping, which is delegated to an unpublished system; the load-bearing nature of this dependency and the reported-number inconsistencies prevent the current version from being accepted without revision.

major comments (4)
  1. [5.4 / Table 3] Table 3 reports for General Product Rec a CTR improvement of +1.58%, Clicks of +1.49%, and Bills of +0.33%, but the text in Section 5.4 states 'For normal product recommendations, CTR improved by 3.72% and Clicks by 5.20%.' Since the Overall row in the same table matches the text's overall numbers, this discrepancy cannot be dismissed as a different segmentation. Please correct the numbers or explain the discrepancy; as written, the reader cannot tell which results are reliable.
  2. [4.1.1 II / Table 1] The Semantic GSID IUs are the dominant IU type (615k IUs, 78.8% product coverage, 77.0% exposure), and every downstream component (IU features in Section 4.2.1, IU click sequences in Section 4.2.2, IU attention in Section 4.2.3) assumes these IDs are semantically coherent and temporally stable. Yet footnote 2 indicates that GSID is unpublished and 'isn't the focus of this paper,' and the manuscript provides no training objective, cluster-quality metric, stability analysis, or comparison against alternative groupings. Please add a direct evaluation of GSID quality (e.g., inter/intra-cluster similarity, manual agreement, temporal ID stability) and, if possible, an ablation that isolates the semantic-IU component from SPU and image IUs; without this, the central claim that persistent IU aggregation drives the gains is not fully supported.
  3. [5.1.1 / 5.3 / 5.4] The offline experiments report only point estimates for AUC, GAUC, and RelaImpr, despite the statement that the testing set was split into 10 parts; no standard deviations or confidence intervals are given, and the stated p<0.05 significance for Table 2 is unverifiable from the text. This is particularly important because the headline AUC improvements are small (0.7411 vs. 0.7366 for DIN). The online A/B results in Table 3 also lack confidence intervals or p-values, although Section 5.4 claims the results are 'statistically validated.' Please provide variance estimates or statistical tests for both offline and online results.
  4. [5.2 / 5.4] The online evaluation appears to involve two separate interventions: Section 5.2 describes an A/B test of the new two-stage product format, while Section 5.4 describes an A/B test of the IU-Boosted model against a DIN-based production model. The manuscript does not state whether the model A/B in Section 5.4 held the product format fixed across control and treatment. If the control arm used the old single-stage format and the treatment arm used the new format plus the new model, the reported overall gains would conflate interface and model effects. Please clarify the experimental design, including which product format was used in each arm.
minor comments (6)
  1. [Abstract / Section 1] There are multiple grammatical errors, such as 'Most of the product on Xianyu posted from individual sellers often have limited stock...' and 'This result in most items...'. Please copyedit the manuscript.
  2. [5.2] The sentence 'The experiment lasted for 7 days, and the new product format achieved a 5% improvement in bills, a 10% improvement in GMV, and even in exposure, there was a minor decrease in clicks' is confusing; please rephrase to distinguish exposure from clicks and to state the direction of each change clearly.
  3. [5.1.2 / Table 2] The baseline named 'GroupID' in Section 5.1.2 is listed as 'GroupEmb' in Table 2 and is attributed to reference [12] (Airbnb embeddings). Please align the name and description of this baseline.
  4. [4.2.2] Equation (6) and the surrounding text define E(IU_Seq) twice with the same notation, and Equation (5) uses E(ItemSeq) inside the IU embedding, making the recursion ambiguous. Please clarify the notation.
  5. [Table 1] The column header 'Numbers' is vague (number of IUs?), and the absolute CTR values are omitted for privacy; please state the unit explicitly and, if possible, give the baseline CTR used to compute the improvement.
  6. [5.1.1] The paper says 'the proposed paradigm and the proposed method are implemented from a industrial practice and there are no suitable public data-set'; please provide a reproducibility statement or at least describe the data-processing pipeline in enough detail to assess potential leakage between IU construction and the offline test.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; the GSID dependency is a reproducibility limitation, not a circular step.

full rationale

Circularity analysis finds no step in which a predicted quantity is equivalent to an input by construction. Interest units are built from product attributes, image clustering, and a query-aware GSID system (Section 4.1.1), none of which is fitted to the CTR labels used in the offline evaluation. The offline experiment trains on seven days and tests on the eighth day (Section 5.1.1), and the online A/B test compares against a DIN baseline (Section 5.4), so the reported AUC/GAUC and CTR gains are empirical comparisons, not algebraic consequences of the IU definitions. The IU-level statistical features and hierarchical IU click sequences in Sections 4.2.1 through 4.2.3 are standard transformations of historical behavior; they do not rename the target label as a feature. The paper's main external dependency is the unpublished GSID system, which the footnote in Section 4.1.1 explicitly says is 'another systematic effort from industry practice, which isn't the focus of this paper' and 'will soon be under review.' That is a reproducibility and support gap for 78.8% of product coverage, but it is not circular: the paper does not derive its results from GSID by definition, from a self-citation, or from a fitted parameter renamed as a prediction. No load-bearing self-citation, imported uniqueness theorem, or ansatz-smuggling citation appears in the derivation chain. Accordingly, no circular step is identified.

Assumptions & free parameters 3 free parameters · 5 assumptions · 2 invented entities

The framework rests on the engineered Interest Unit types (SPU, Image, Semantic), the unpublished GSID semantic encoding, and an explicit behavioral model of two-stage browsing. The model hyperparameters are standard, but the IU clustering granularity and the stability of semantic IDs are load-bearing and untested.

free parameters (3)
  • IU click sequence max length = 5
    Hand-set in Section 5.1.4; caps the hierarchical IU click sequence used by the attention module. The paper does not ablate this choice.
  • GSID codebook sizes = 3 levels x 128 IDs; ~16k level-2; ~2.1M level-3
    Chosen in Section 4.1.1 (II) for the semantic interest unit hierarchy; granularity of IUs directly affects aggregation and persistence benefits, but no sensitivity analysis is provided.
  • Interest unit type mix = SPU 45.9k, Image 287k, Semantic 615k units
    The three IU types (Section 4.1.1) are engineered design choices with different coverage; their relative weighting in the final system is not specified.
assumptions (5)
  • domain assumption The core demands of Xianyu users can be exhaustively identified to some extent
    Opening of Section 4.1.1; the whole IU construction assumes a finite, predefinable set of interest units covers user needs. If not, many products fall outside well-formed IUs.
  • domain assumption Users browse in two stages: broad interest then specific product comparison
    Section 1 states 'users tend to browse through a set of similar products ... and then examine each item in detail'; the two-stage IU interface and two-stage recommendation are built on this behavior pattern. No user-study evidence is given.
  • domain assumption IU-level aggregated behaviors remain valid after individual products are sold
    Section 4.2.1 asserts information aggregated on interest units 'remains permanently accessible'; this persistence property is the core benefit but depends on IUs being stable over time and not polluted by sold-out item drift.
  • ad hoc to paper GSID semantic IDs are meaningful and correctly group products by demand
    Section 4.1.1 (II) delegates semantic IU construction to the unpublished GSID system (footnote 2: 'will soon be under review'). The paper does not evaluate clustering quality or stability.
  • standard math Standard AUC/GAUC/RelaImpr definitions and t-test significance
    Equations (7)-(9) in Section 5.1.3; standard, but statistical significance is asserted without variance.
invented entities (2)
  • Interest Unit (IU) independent evidence
    purpose: Persistent cluster of similar products that aggregates user interactions and is recommended before individual items
    Introduced in Section 4.1; the paper provides online A/B and deployment evidence that IU-based formats lift CTR, bills, and GMV, but only on Xianyu data, with no external replication.
  • GSID hierarchical semantic ID
    purpose: Discrete semantic code (T5 encoder-decoder + CodeBook) assigning each product to a hierarchical demand theme
    Section 4.1.1 (II); the system is unpublished, details are minimal, and no independent validation of the IDs is provided, so its falsifiable handle outside the paper is not available.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IU4Rec: Interest Unit-Based Product Organization and Recommendation for E-Commerce Platform." pith.science (2026). https://pith.science/paper/P6VQ2OGU

@misc{pith2026250207658,
  author       = {Pith},
  title        = {Pith review of: IU4Rec: Interest Unit-Based Product Organization and Recommendation for E-Commerce Platform},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P6VQ2OGU}},
  note         = {Machine review of arXiv:2502.07658}
}
read the original abstract

Most recommendation systems typically follow a product-based paradigm utilizing user-product interactions to identify the most engaging items for users. However, this product-based paradigm has notable drawbacks for Xianyu~\footnote{Xianyu is China's largest online C2C e-commerce platform where a large portion of the product are post by individual sellers}. Most of the product on Xianyu posted from individual sellers often have limited stock available for distribution, and once the product is sold, it's no longer available for distribution. This result in most items distributed product on Xianyu having relatively few interactions, affecting the effectiveness of traditional recommendation depending on accumulating user-item interactions. To address these issues, we introduce \textbf{IU4Rec}, an \textbf{I}nterest \textbf{U}nit-based two-stage \textbf{Rec}ommendation system framework. We first group products into clusters based on attributes such as category, image, and semantics. These IUs are then integrated into the Recommendation system, delivering both product and technological innovations. IU4Rec begins by grouping products into clusters based on attributes such as category, image, and semantics, forming Interest Units (IUs). Then we redesign the recommendation process into two stages. In the first stage, the focus is on recommend these Interest Units, capturing broad-level interests. In the second stage, it guides users to find the best option among similar products within the selected Interest Unit. User-IU interactions are incorporated into our ranking models, offering the advantage of more persistent IU behaviors compared to item-specific interactions. Experimental results on the production dataset and online A/B testing demonstrate the effectiveness and superiority of our proposed IU-centric recommendation approach.

Figures

Figures reproduced from arXiv: 2502.07658 by the authors.

Figure 1
Figure 1. An overview of the IU4Rec framework, which consists of (a) the construction of interest unit: products are organized [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An illustration of Embedding and MLP structure [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. One example of the foundational understanding [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The redesigned product format. Left is stage one [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: An overview of proposed IU-Boosted Network, which consists of three components: (1) the interest unit-level feature [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 19 canonical work pages

  1. [1]

    Bobadilla, F

    J. Bobadilla, F. Ortega, A. Hernando, and A. GutiéRrez. 2013. Recommender systems survey. Know.-Based Syst. 46 (July 2013), 109–132. https://doi.org/10. 1016/j.knosys.2013.03.012

    work page 2013

  2. [2]

    Yi Cao, Sihao Hu, Yu Gong, Zhao Li, Yazheng Yang, Qingwen Liu, and Shoul- ing Ji. 2022. GIFT: Graph-GuIded Feature Transfer for Cold-Start Video Click- Through Rate Prediction. In Proceedings of the 31st ACM International Con- ference on Information & Knowledge Management (Atlanta, GA, USA) (CIKM ’22). Association for Computing Machinery, New York, NY, US...

    arXiv 2022

  3. [3]

    Carbonell and Jade Goldstein-Stewart

    Jaime G. Carbonell and Jade Goldstein-Stewart. 1998. The Use of MMR, Diversity- Based Reranking for Reordering Documents and Producing Summaries. ACM SIGIR Forum 51 (1998), 209 – 210. https://api.semanticscholar.org/CorpusID: 6334682

    work page 1998

  4. [4]

    Zhangming Chan, Yu Zhang, Shuguang Han, Yong Bai, Xiang-Rong Sheng, Siyuan Lou, Jiacen Hu, Baolin Liu, Yuning Jiang, Jian Xu, and Bo Zheng. 2023. Capturing Conversion Rate Fluctuation during Sales Promotions: A Novel Historical Data Reuse Approach. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Long Beach, CA, USA)...

    arXiv 2023

  5. [5]

    Jianxin Chang, Chenbin Zhang, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, and Kun Gai. 2023. PEPNet: Parameter and Embedding Personalized Network for Infusing with Personalized Prior Information. arXiv:2302.01115 [cs.IR]

    work page Pith review arXiv 2023

  6. [6]

    Qiwei Chen, Huan Zhao, Wei Li, Pipei Huang, and Wenwu Ou. 2019. Be- havior Sequence Transformer for E-Commerce Recommendation in Alibaba. In Proceedings of the 1st International Workshop on Deep Learning Practice for High-Dimensional Sparse Data (Anchorage, Alaska) (DLP-KDD ’19). Asso- ciation for Computing Machinery, New York, NY, USA, Article 12, 4 page...

    arXiv 2019

  7. [7]

    Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, Rohan Anil, Zakaria Haque, Lichan Hong, Vihan Jain, Xiaobing Liu, and Hemal Shah

  8. [8]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep Neural Networks for YouTube Recommendations. In Proceedings of the 10th ACM Conference on Recommender Systems (Boston, Massachusetts, USA) (RecSys ’16). Association for Computing Machinery, New York, NY, USA, 191–198. https://doi.org/10.1145/ 2959100.2959190

    arXiv 2016

Show all 35 references
  1. [9]

    Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. 2019. Class- Balanced Loss Based on Effective Number of Samples. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 9260–9269. https://doi.org/ 10.1109/CVPR.2019.00949

    2019

  2. [10]

    Shangfeng Dai, Haobin Lin, Zhichen Zhao, Jianying Lin, Honghuan Wu, Zhe Wang, Sen Yang, and Ji Liu. 2021. POSO: Personalized Cold Start Modules for IU4Rec: Interest Unit-Based Product Organization and Recommendation for E-Commerce Platform WWW ’24 Companion, May 13–17, 2024, S...

    2021 arXiv

  3. [13]

    Xiaoqiang Gui, Yueyao Cheng, Xiang-Rong Sheng, Yunfeng Zhao, Guoxian Yu, Shuguang Han, Yuning Jiang, Jian Xu, and Bo Zheng. 2023. Calibration- compatible Listwise Distillation of Privileged Features for CTR Prediction. arXiv preprint arXiv:2312.08727 (2023)

    2023 arXiv

  4. [14]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A Factorization-Machine Based Neural Network for CTR Prediction. In Proceedings of the 26th International Joint Conference on Artificial Intelligence (Melbourne, Australia) (IJCAI’17). AAAI Press...

    2017

  5. [15]

    Jiacen Hu, Zhangming Chan, Yu Zhang, Shuguang Han, Siyuan Lou, Baolin Liu, Han Zhu, Yuning Jiang, Jian Xu, and Bo Zheng. 2023. PS-SA: An Efficient Self- Attention via Progressive Sampling for User Behavior Sequence Modeling. InPro- ceedings of the 32nd ACM International Confer...

    2023

  6. [16]

    Tongwen Huang, Qingyun She, Zhiqiang Wang, and Junlin Zhang. 2020. GateNet: gating-enhanced deep network for click-through rate prediction. arXiv preprint arXiv:2007.03519 (2020)

    2020 arXiv

  7. [17]

    Tongwen Huang, Zhiqi Zhang, and Junlin Zhang. 2019. FiBiNET: Combining Feature Importance and Bilinear Feature Interaction for Click-through Rate Prediction. In Proceedings of the 13th ACM Conference on Recommender Systems (Copenhagen, Denmark) (RecSys ’19). Association for Co...

    2019

  8. [18]

    Tong Man, Huawei Shen, Xiaolong Jin, and Xueqi Cheng. 2017. Cross-Domain Recommendation: An Embedding and Mapping Approach. In Proceedings of the 26th International Joint Conference on Artificial Intelligence (Melbourne, Australia) (IJCAI’17). AAAI Press, 2464–2470

    2017

  9. [19]

    Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. 2021. Long-tail learning via logit adjustment. In International Conference on Learning Representations . https://openreview.net/ forum?id=37nvvqkCo5

    2021

  10. [20]

    Kaixiang Mo, Bo Liu, Lei Xiao, Yong Li, and Jie Jiang. 2015. Image Feature Learning for Cold Start Problem in Display Advertising. In Proceedings of the 24th International Conference on Artificial Intelligence (Buenos Aires, Argentina) (IJCAI’15). AAAI Press, 3728–3734

    2015

  11. [21]

    Feiyang Pan, Shuokai Li, Xiang Ao, Pingzhong Tang, and Qing He. 2019. Warm Up Cold-Start Advertisements: Improving CTR Predictions via Learning to Learn ID Embeddings. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retr...

    2019

  12. [22]

    Xiang-Rong Sheng, Liqin Zhao, Guorui Zhou, Xinyao Ding, Binding Dai, Qiang Luo, Siran Yang, Jingshan Lv, Chi Zhang, Hongbo Deng, and Xiaoqiang Zhu. 2021. One Model to Serve All: Star Topology Adaptive Recommender for Multi-Domain CTR Prediction. In Proceedings of the 30th ACM ...

    2021

  13. [23]

    Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W

    Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W. Cohen, and Don- ald Metzler. 2022. Transformer Memory as a Differentiable Search Index. ArXiv abs/2202.06991 (2022). https://api.semanti...

    2022 arXiv

  14. [24]

    Manos Tsagkias, Tracy Holloway King, Surya Kallumadi, Vanessa Murdock, and Maarten de Rijke. 2021. Challenges and research opportunities in eCommerce search and recommendations. SIGIR Forum 54, 1, Article 2 (Feb. 2021), 23 pages. https://doi.org/10.1145/3451964.3451966

    2021

  15. [25]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is All You Need. In Proceedings of the 31st International Conference on Neural Information Processing Systems (Long Beach, California, ...

    2017

  16. [26]

    Maksims Volkovs, Guangwei Yu, and Tomi Poutanen. 2017. DropoutNet: Address- ing Cold Start in Recommender Systems. In Proceedings of the 31st International Conference on Neural Information Processing Systems (Long Beach, California, USA) (NIPS’17). Curran Associates Inc., Red ...

    2017

  17. [27]

    Hongwei Wang, Fuzheng Zhang, Jialin Wang, Miao Zhao, Wenjie Li, Xing Xie, and Minyi Guo. 2018. RippleNet: Propagating User Preferences on the Knowledge Graph for Recommender Systems. In Proceedings of the 27th ACM International Conference on Information and Knowledge Managemen...

    2018

  18. [28]

    Kailun Wu, Weijie Bian, Zhangming Chan, Lejian Ren, Shiming Xiang, Shu- Guang Han, Hongbo Deng, and Bo Zheng. 2022. Adversarial gradient driven exploration for deep click-through rate prediction. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data M...

    2022

  19. [29]

    Qitian Wu, Hengrui Zhang, Xiaofeng Gao, Peng He, Paul Weng, Han Gao, and Guihai Chen. 2019. Dual Graph Attention Networks for Deep Latent Representation of Multifaceted Social Effects in Recommender Systems. In The World Wide Web Conference (San Francisco, CA, USA) (WWW ’19) ....

    2019

  20. [30]

    Wenhao Wu, Jialiang Zhou, Ailong He, Shuguang Han, Jufeng Chen, and Bo Zheng. 2024. MetaSplit: Meta-Split Network for Limited-Stock Product Recom- mendation. In Companion Proceedings of the ACM Web Conference 2024(Singapore, Singapore) (WWW ’24). Association for Computing Mach...

    2024

  21. [31]

    Dongo Xi, Fuzhen Zhuang, Yanchi Liu, Jingjing Gu, Hui Xiong, and Qing He

  22. [32]

    Ruobing Xie, Zhijie Qiu, Jun Rao, Yi Liu, Bo Zhang, and Leyu Lin. 2021. Internal and Contextual Attention Network for Cold-Start Multi-Channel Matching in Recommendation. In Proceedings of the Twenty-Ninth International Joint Confer- ence on Artificial Intelligence (Yokohama, ...

    2021

  23. [33]

    Yujing Zhang, Zhangming Chan, Shuhao Xu, Weijie Bian, Shuguang Han, Hongbo Deng, and Bo Zheng. 2022. KEEP: An industrial pre-training framework for online recommendation via knowledge extraction and plugging. In Proceedings of the 31st ACM International Conference on Informati...

    2022

  24. [34]

    Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep Interest Evolution Network for Click-through Rate Prediction. In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelli- gence and Thirty-First Innovative Appli...

    2019

  25. [35]

    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...

    2018

  26. [2016]

    In Proceedings of the 1st Workshop on Deep Learning for Recommender Systems (Boston, MA, USA) (DLRS 2016)

    Wide & Deep Learning for Recommender Systems. In Proceedings of the 1st Workshop on Deep Learning for Recommender Systems (Boston, MA, USA) (DLRS 2016). Association for Computing Machinery, New York, NY, USA, 7–10. https://doi.org/10.1145/2988450.2988454

    2016

  27. [2019]

    Modelling of Bi-Directional Spatio-Temporal Dependence and Users’ Dy- namic Preferences for Missing POI Check-in Identification. In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence and Thirty-First Inno- vative Applications of Artificial Intelligence ...

    doi

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.