Pith. sign in

REVIEW 3 major objections 7 minor 88 references

DLF: Enhancing Explicit-Implicit Interaction via Dynamic Low-Order-Aware Fusion for CTR Prediction

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read DLF claims that fusing explicit and implicit interaction streams at every layer, guided by a network-aware attention module and a residual-aware low-order pathway, yields state-of-the-art CTR prediction across four public datasets.

desk verdict A solid but incremental CTR architecture; the layer-wise fusion idea is worth considering, but the equations don't type-check and the empirical claims outrun the reported evidence. read the letter →

arxiv 2505.19182 v1 pith:4C4S2KAH submitted 2025-05-25 cs.IR

Kefan Wang , Hao Wang , Wei Guo , Yong Liu , Jianghao Lin , Defu Lian , Enhong Chen This is my paper · ORCID
classification cs.IR
keywords CTRpredictionrecommendationsystemsfeatureinteractionmodelingexplicit-implicitlow-orderpreservationattentionfusiongatedresidualdeeplearning
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

Click-through rate prediction in ads and recommender systems hinges on modeling feature interactions, but classic two-stream models keep explicit (e.g., inner-product) and implicit (MLP) interaction streams separate until a final fusion, which can lose low-order signals, limit information sharing, and create gradient imbalance. The paper proposes DLF, a stacked architecture that fuses explicit and implicit representations at every layer. Its Residual-Aware Low-Order Interaction Network (RLI) preserves low-order signals through a dedicated low-rank pathway plus a gated residual, while its Network-Aware Attention Fusion Module (NAF) combines block outputs with inter-block cross-attention and intra-block self-attention. The paper's central empirical claim is that DLF achieves the highest AUC and lowest log-loss among all compared baselines on Criteo, Avazu, Movielens, and Frappe, with the largest gains on the sparsest data. If the claim holds, layer-wise dynamic fusion is a more effective default architecture for sparse CTR data than late-fusion two-stream designs.

What carries the argument

The architecture stacks $L$ interaction layers. Each layer runs three parallel blocks: the low-rank block computes $Z_L^{(l+1)} = \sigma(W_{LO}(E^{(1)} \cdot (E^{(l)})^\top W_{LI}) + b_L)$, interacting the original embeddings with the current layer; the high-rank block similarly interacts $E^{(l)}$ with itself; and the implicit block applies a dense ReLU transformation. These outputs $Z_L, Z_C, Z_D$ feed the Network-Aware Attention Fusion Module, which derives query/key/value for each block, applies cross-attention from each block to the other two and self-attention within each block, then concatenates the original block outputs with the attention result: $Z = [Z_L; Z_C; Z_D; Att]$. A gated residual, $E^{(l+1)} = Z + \max(\epsilon, W_{Gate} E^{(l)}) \odot E^{(l)}$, balances the fused signal against the previous layer. The low-rank pathway carries the claimed low-order preservation, and the attention module carries the claimed layer-wise information sharing.

What would settle it

Inspect the released implementation for the projection from the $4d$-dimensional concatenation in Equation (15) to the $d$-dimensional representation required by Equation (9); if no such projection exists, the model as written cannot be trained, and the reported numbers must come from a different architecture than the paper describes.

Watch

Extended reading notes

Core claim

The paper sets out to establish that the long-standing division between explicit and implicit interaction streams is the bottleneck in CTR models, and that fusing the two at every layer—not just at the output—fixes it. Concretely, it claims that combining a residual-aware low-order interaction network with a network-aware attention fusion module yields a model that consistently outperforms every compared baseline on four public datasets, reaching AUC 0.9515 and log-loss 0.2223 on Movielens and AUC 0.9812 and log-loss 0.1679 on Frappe, the sparsest dataset. It also claims the two components are complementary: ablations show removing the attention fusion module costs the most, while the gated residual prevents the low-order pathway from conflicting with residual connections.

Load-bearing premise

The paper assumes that every layer representation, every block output, and the fused vector have the same width, so that Equation (9) can add them directly; but Equation (15) produces a vector four times as wide, and no projection back to the original width is specified.

Editorial extensions

If this is right

  • If DLF's central claim holds, layer-wise attention fusion should outperform late-fusion two-stream designs across sparse CTR benchmarks, making per-layer fusion the default design choice for feature interaction models.
  • The combination of a gated residual and a low-rank interaction pathway offers a reusable recipe for preserving low-order signals without suppressing high-order expressiveness, applicable beyond CTR to other sparse prediction tasks.
  • Because the fusion module is block-agnostic, future interaction blocks (tensor-based, graph-based, or otherwise) could be plugged into DLF's per-layer fusion without redesigning the network.
  • The largest reported gains occur on Frappe, the sparsest dataset, suggesting that dynamic low-order-aware fusion is most valuable exactly where data is scarce.
  • If the reported AUC and log-loss gains reproduce in practice, adopting DLF as a backbone in ad-serving or recommender pipelines would directly improve ranking quality.

Reading between the lines

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

  • A natural test would isolate the effect of fusion timing from the attention mechanism: replacing NAF with a simple per-layer concatenation at matched parameter count would tell whether layer-wise fusion itself, rather than the attention design, drives the reported gains.
  • The paper does not discuss the dimensional mismatch between the $4d$-dimensional concatenation in Equation (15) and the $d$-dimensional addition required by Equation (9); an undocumented projection in the released code would mean the written model is under-specified.
  • The low-rank pathway, which interacts each layer with the original embeddings, resembles a parameter-efficient inductive bias that could transfer to sequential recommendation or multi-task CTR/conversion setups where low-order signals also degrade with depth.
  • A further test: replacing RLI's low-rank pathway with a plain copy of the original embeddings concatenated at each layer would check whether the specific functional form of the low-order interaction, rather than mere access to $E^{(1)}$, is what matters.
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

3 major / 7 minor

Summary. The paper proposes DLF, a two-stream CTR prediction model whose two main modules are a Residual-Aware Low-Order Interaction Network (RLI) and a Network-Aware Attention Fusion Module (NAF). RLI combines low-rank, high-rank, and implicit interaction blocks in each layer with a gated residual connection, while NAF fuses these block outputs through inter-block and intra-block attention. The authors report experiments on Criteo, Avazu, Movielens, and Frappe against 15 baselines implemented in FuxiCTR, with 7:2:1 chronological splits and five repetitions, and claim state-of-the-art AUC and log-loss results with p<0.05 significance. The implementation is publicly available.

Significance. If the reported results are reproducible and the architecture is fully specified, the layer-wise fusion idea is a reasonable incremental contribution to CTR prediction, and the public implementation plus standard FuxiCTR benchmarking protocol are strengths. The main significance is empirical: DLF outperforms all baselines on four datasets in Table 2, and the ablations in Table 3 show consistent drops when RLI or NAF is removed. However, the paper's central claim is currently not fully supported because the forward pass as written is dimensionally inconsistent at load-bearing equations, and the statistical significance claim in Table 2 is not verifiable without variance or test details.

major comments (3)
  1. [§4.3.2, Eqs. (15), (9), (16)] There is a load-bearing dimension inconsistency in the forward pass. Equation (15) defines Z = [Z_L; Z_C; Z_D; Att], where each block output and Att are d-dimensional, so Z is 4d-dimensional. Equation (9) then sets E^(l+1) = Z + max(eps, W_Gate E^(l)) ⊙ E^(l), but E^(l) and the gated residual are d-dimensional, so the addition is undefined unless a projection from 4d to d is silently inserted. Equation (16) compounds the problem because W_out ∈ R^d cannot consume a 4d-dimensional E^(L+1). The paper does not specify any projection or reshaping. As written, the model cannot be executed layer-by-layer, and the reported experiments would be testing an implementation that is not fully described. Please state the exact projection or reshaping used, and align Eqs. (9) and (16) with Eq. (15).
  2. [§4.2, Eqs. (5) and (6)] The matrix dimensions in the low-rank and high-rank interaction blocks are incompatible with the stated shapes. If E^(1) and E^(l) are N×d (N feature fields), then E^(1)·(E^(l))^T is N×N, but W_LI is stated to be in R^{R×d}; the product (E^(l))^T W_LI is also dimensionally invalid as d×N times R×d unless N=R. The same issue applies to Eq. (6) with W_CI ∈ R^{R×d}. The paper does not define R=N or give any other consistent tensor interpretation. Please provide correct, consistent dimensions for all matrices in Eqs. (5) and (6), or clarify the intended tensor layout.
  3. [Table 2, §5.2] The caption claims that bold values indicate p<0.05 when comparing DLF with the base models, but the table reports only point estimates with no standard deviations, confidence intervals, or details of the significance test. With five runs, differences as small as 0.0006 AUC (DLF 0.7601 vs. Wukong 0.7580 on Avazu) are not meaningful without variance information. Please report standard deviations or per-run results for all datasets, state which baseline(s) the comparison is against, and describe the paired test and any multiple-comparison correction. Without this, the central claim that DLF 'consistently outperforms all baselines on every dataset' is not verifiable.
minor comments (7)
  1. [§5.5] The text says 'As illustrated in Figure Y' but no Figure Y exists; this placeholder should be replaced with the correct figure reference.
  2. [§5.4] The fusion-method analysis is said to be shown in Figure 3, but the text later refers to 'Figure 4' for the same comparison; the figure numbering is inconsistent.
  3. [Figure 4, §5.5] The hyper-parameter analysis discusses both the number of attention layers and the number of stacked interaction layers L, but the figure appears to show only one panel whose x-axis is labeled 'Attention Layers L'. Please include both panels or clarify the figure content.
  4. [Table 1] The caption says 'Table 1 provides detailed statistics for these two public datasets,' but four datasets are listed. This should read 'four.'
  5. [References] Reference [29] is a duplicate of [28] (both are the AutoFIS paper by Liu et al.); one should be removed and the in-text citations renumbered. Also, the introduction cites DCN as [11] alongside DeepFM [11]; DCN should be [55].
  6. [§4.2] The phrase 'through the LOPL' appears to be a typo for 'through the RLI' or 'through the low-order pathway'; no definition of LOPL is given.
  7. [Abstract and §1] The paper repeatedly claims that NAF alleviates gradient imbalance and that RLI mitigates sparsity issues, but no experiment directly measures gradient norms, embedding quality, or sparsity effects. Consider adding such an analysis or softening these claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DLF is an empirical CTR model evaluated on standard train/test splits, with no fitted parameter renamed as a prediction and no load-bearing self-citation.

full rationale

The paper's central claim is empirical: DLF outperforms baselines on four public datasets in AUC and LogLoss. The model is trained with binary cross-entropy on standard 7:2:1 time-ordered splits, and the reported numbers are averages over five runs; no parameter is fitted to the test set and then reported as a prediction. The architecture components, RLI and NAF, are defined by their own equations (Eqs. 5-15) rather than imported from the authors' prior work. Self-citations in the reference list (e.g., Refs. 14, 49, 50, 54, 63) appear as related-work context, not as justification for the main result; the comparison baselines are implemented via the external FuxiCTR framework with shared hyperparameter settings. The qualitative claims about gradient imbalance and low-order signal preservation are under-supported rather than circular, since they are explanatory narratives around an empirical ranking and not derived from the equations. The paper does contain a separate specification issue: Eq. (15) concatenates block outputs into a 4d vector while Eq. (9) adds it to a d-dimensional representation without an explicit projection, so the forward pass as written is dimensionally inconsistent. That is a correctness/reproducibility concern, not a circularity concern: it does not make the empirical claim equivalent to its inputs. No step in the paper reduces a derived result to a fitted input or to a self-citation chain, so the appropriate score is 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

All trainable weights are learned from data in a standard supervised setup and are not counted as ad hoc free parameters. The free parameters listed are hyperparameters and an unreported rank. The axioms are the unproved modeling premises underlying the architecture. No independently invented physical or conceptual entity is introduced; RLI and NAF are composite modules built from standard operations and are supported by ablation experiments.

free parameters (5)
  • embedding dimension d = 64
    Fixed to 64 for all models in Section 5.1.4; architecture width depends on it.
  • interaction rank R = not reported
    Appears in W_LI in R^{R x d} and W_CI in R^{R x d} in Equations (5)-(6); the paper never states its value or how it was selected.
  • number of stacked interaction layers L = optimal near 8 (Criteo)
    Section 5.5 tunes L from 2 to 12; final per-dataset values are not reported.
  • number of attention layers = optimal 2
    Section 5.5 tunes attention depth; final configuration per dataset is not reported.
  • optimizer hyperparameters = tuned ranges
    Learning rate, L2, and dropout are tuned from the listed ranges in Section 5.1.4; exact values are not given.
assumptions (6)
  • standard math Standard definitions of softmax attention and binary cross-entropy loss are assumed.
    Sections 4.3 and 4.4 use these without derivation.
  • domain assumption All layer representations E(l), block outputs Z_L, Z_C, Z_D, and fused output share the same dimension d and field count, allowing Eq. (9) to compose with Eq. (15).
    Required by the stacked architecture; contradicted by the 4d concatenation in Eq. (15).
  • domain assumption Interaction order grows linearly when mixing with input embeddings and exponentially when mixing a layer with itself, as in Eq. (2) and Eq. (4).
    Used in Section 3.2 to justify the low-order/high-order gap; no formal definition of interaction order is given.
  • ad hoc to paper Explicitly reusing initial embeddings at each layer preserves useful low-order signals rather than adding redundancy.
    Central design premise of RLI in Section 4.2; the ablation in Section 5.3 notes that RLI can conflict with the residual path when gating is removed.
  • domain assumption Layer-wise attention fusion reduces gradient imbalance between explicit and implicit streams.
    Asserted in Sections 1 and 4.3; no gradient norm measurements or convergence analysis are provided.
  • domain assumption FuxiCTR baseline implementations are faithful and the AFN splitting protocol is suitable for all compared models.
    Section 5.1 relies on FuxiCTR for all implementations; baseline fidelity is not independently verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DLF: Enhancing Explicit-Implicit Interaction via Dynamic Low-Order-Aware Fusion for CTR Prediction." pith.science (2026). https://pith.science/paper/4C4S2KAH

@misc{pith2026250519182,
  author       = {Pith},
  title        = {Pith review of: DLF: Enhancing Explicit-Implicit Interaction via Dynamic Low-Order-Aware Fusion for CTR Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4C4S2KAH}},
  note         = {Machine review of arXiv:2505.19182}
}
read the original abstract

Click-through rate (CTR) prediction is a critical task in online advertising and recommender systems, relying on effective modeling of feature interactions. Explicit interactions capture predefined relationships, such as inner products, but often suffer from data sparsity, while implicit interactions excel at learning complex patterns through non-linear transformations but lack inductive biases for efficient low-order modeling. Existing two-stream architectures integrate these paradigms but face challenges such as limited information sharing, gradient imbalance, and difficulty preserving low-order signals in sparse CTR data. We propose a novel framework, Dynamic Low-Order-Aware Fusion (DLF), which addresses these limitations through two key components: a Residual-Aware Low-Order Interaction Network (RLI) and a Network-Aware Attention Fusion Module (NAF). RLI explicitly preserves low-order signals while mitigating redundancy from residual connections, and NAF dynamically integrates explicit and implicit representations at each layer, enhancing information sharing and alleviating gradient imbalance. Together, these innovations balance low-order and high-order interactions, improving model expressiveness. Extensive experiments on public datasets demonstrate that DLF achieves state-of-the-art performance in CTR prediction, addressing key limitations of existing models. The implementation is publicly available at https://github.com/USTC-StarTeam/DLF.

Figures

Figures reproduced from arXiv: 2505.19182 by the authors.

Figure 1
Figure 1. Drawbacks of deep explicit interaction and deep [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of the proposed DLF model. The left part shows the complete workflow of DLF, with the core [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Interaction fusion methods analysis the model avoids this conflict but loses the ability to effectively model low-order interactions. This observation underscores the critical role of residual gating in retaining meaningful low-order information and integrating it seamlessly into the model. On the other hand, the removal of NAF results in a significant performance drop, which is only marginally better than removing … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Hyper-Parameter Analysis of DLF meaningful feature information compared to traditional shallow models, further validating the design of our attention fusion and residual-aware mechanisms. 6 Conclusion In this work, we addressed the challenges of feature interaction mod…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

88 extracted references · 29 canonical work pages

  1. [1]

    Mathieu Blondel, Akinori Fujino, Naonori Ueda, and Masakazu Ishihata. 2016. Higher-order factorization machines.Advances in neural information processing systems29 (2016)

    2016

  2. [2]

    Bo Chen, Yichao Wang, Zhirong Liu, Ruiming Tang, Wei Guo, Hongkun Zheng, Weiwei Yao, Muyu Zhang, and Xiuqiang He. 2021. Enhancing explicit and implicit SIGIR ’25, July 13–18, 2025, Padua, Italy Kefan Wang et al. feature interactions via information sharing for parallel deep ctr models. In Proceedings of the 30th ACM international conference on information...

    2021

  3. [3]

    Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, et al

  4. [4]

    Weiyu Cheng, Yanyan Shen, and Linpeng Huang. 2020. Adaptive factorization network: Learning adaptive-order feature interactions. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 34. 3609–3616

    2020

  5. [5]

    Yuan Cheng and Yanbo Xue. 2021. Looking at CTR Prediction Again: Is Attention All You Need?. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’21). ACM. https: //doi.org/10.1145/3404835.3462936

    arXiv 2021

  6. [6]

    D. R. Cox. 1958. The Regression Analysis of Binary Sequences.Journal of the Royal Statistical Society. Series B (Methodological)20, 2 (1958), 215–242. http: //www.jstor.org/stable/2983890

    work page arXiv 1958

  7. [8]

    Wei Deng, Junwei Pan, Tian Zhou, Deguang Kong, Aaron Flores, and Guang Lin. 2021. Deeplight: Deep lightweight feature interactions for accelerating ctr predictions in ad serving. InProceedings of the 14th ACM international conference on Web search and data mining. 922–930

    2021

  8. [9]

    Thore Graepel, Joaquin Quinonero Candela, Thomas Borchert, and Ralf Herbrich

Show all 88 references
  1. [10]

    Hongchao Gu, Dexun Li, Kuicai Dong, Hao Zhang, Hang Lv, Hao Wang, Defu Lian, Yong Liu, and Enhong Chen. 2025. RAPID: Efficient Retrieval-Augmented Long Text Generation with Writing Planning and Information Discovery.arXiv preprint arXiv:2503.00751(2025)

    2025 arXiv

  2. [11]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: a factorization-machine based neural network for CTR prediction.arXiv preprint arXiv:1703.04247(2017)

    2017 arXiv

  3. [12]

    Wei Guo, Ruiming Tang, Huifeng Guo, Jianhua Han, Wen Yang, and Yuzhou Zhang. 2019. Order-aware embedding neural network for CTR prediction. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval. 1121–1124

    2019

  4. [13]

    Wei Guo, Hao Wang, Luankang Zhang, Jin Yao Chin, Zhongzhou Liu, Kai Cheng, Qiushi Pan, Yi Quan Lee, Wanqi Xue, Tingjia Shen, et al . 2024. Scal- ing New Frontiers: Insights into Large Recommendation Models.arXiv preprint arXiv:2412.00714(2024)

    2024 arXiv

  5. [14]

    Yongqiang Han, Hao Wang, Kefan Wang, Likang Wu, Zhi Li, Wei Guo, Yong Liu, Defu Lian, and Enhong Chen. 2024. Efficient noise-decoupling for multi-behavior sequential recommendation. InProceedings of the ACM Web Conference 2024. 3297–3306

    2024

  6. [15]

    Yongqiang Han, Likang Wu, Hao Wang, Guifeng Wang, Mengdi Zhang, Zhi Li, Defu Lian, and Enhong Chen. 2023. Guesr: A global unsupervised data- enhancement with bucket-cluster sampling for sequential recommendation. In International conference on database systems for advanced app...

    2023

  7. [16]

    Xiangnan He and Tat-Seng Chua. 2017. Neural Factorization Machines for Sparse Predictive Analytics. arXiv:1708.05027 [cs.LG] https://arxiv.org/abs/1708.05027

    2017 arXiv

  8. [17]

    Xinran He, Junfeng Pan, Ou Jin, Tianbing Xu, Bo Liu, Tao Xu, Yanxin Shi, Antoine Atallah, Ralf Herbrich, Stuart Bowers, et al. 2014. Practical lessons from predicting clicks on ads at facebook. InProceedings of the eighth international workshop on data mining for online advert...

    2014

  9. [18]

    Di Hu, Chengze Wang, Feiping Nie, and Xuelong Li. 2019. Dense multimodal fusion for hierarchically joint representation. InICASSP 2019-2019 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 3941–3945

    2019

  10. [19]

    Jie Hu, Li Shen, and Gang Sun. 2018. Squeeze-and-excitation networks. InProceed- ings of the IEEE conference on computer vision and pattern recognition. 7132–7141

    2018

  11. [20]

    Jin Huang and Charles X Ling. 2005. Using AUC and accuracy in evaluating learning algorithms.IEEE Transactions on knowledge and Data Engineering17, 3 (2005), 299–310

    2005

  12. [21]

    Tongwen Huang, Zhiqi Zhang, and Junlin Zhang. 2019. FiBiNET: combining fea- ture importance and bilinear feature interaction for click-through rate prediction. InProceedings of the 13th ACM conference on recommender systems. 169–177

    2019

  13. [22]

    Yuqing Huang, Rongyang Zhang, Xuesong He, Xuyang Zhi, Hao Wang, Xin Li, Feiyang Xu, Deguang Liu, Huadong Liang, Yi Li, et al . 2024. ChemEval: A Comprehensive Multi-Level Chemical Evaluation for Large Language Models. arXiv preprint arXiv:2409.13989(2024)

    2024 arXiv

  14. [23]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2017. Adam: A Method for Stochastic Opti- mization. arXiv:1412.6980 [cs.LG] https://arxiv.org/abs/1412.6980

    2017 arXiv

  15. [24]

    Honghao Li, Yiwen Zhang, Yi Zhang, Hanwei Li, Lei Sang, and Jieming Zhu

  16. [25]

    Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, and Guangzhong Sun. 2018. xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery &amp; Data...

    2018

  17. [26]

    Chao Liao, Jianchao Tan, Jiyuan Jia, Yi Guo, and Chengru Song. 2023. MaskFusion: Feature Augmentation for Click-Through Rate Prediction via Input-adaptive Mask Fusion. InThe Eleventh International Conference on Learning Representations

    2023

  18. [27]

    Bin Liu, Niannan Xue, Huifeng Guo, Ruiming Tang, Stefanos Zafeiriou, Xiuqiang He, and Zhenguo Li. 2020. AutoGroup: Automatic feature grouping for modelling explicit high-order feature interactions in CTR prediction. InProceedings of the 43rd international ACM SIGIR conference ...

    2020

  19. [29]

    Bin Liu, Chenxu Zhu, Guilin Li, Weinan Zhang, Jincai Lai, Ruiming Tang, Xi- uqiang He, Zhenguo Li, and Yong Yu. 2020. Autofis: Automatic feature interaction selection in factorization models for click-through rate prediction. Inproceedings of the 26th ACM SIGKDD international ...

    2020

  20. [30]

    Weiwen Liu, Wei Guo, Yong Liu, Ruiming Tang, and Hao Wang. 2023. User Behavior Modeling with Deep Learning for Recommendation: Recent Advances. InProceedings of the 17th ACM Conference on Recommender Systems. 1286–1287

    2023

  21. [31]

    Wantong Lu, Yantao Yu, Yongzhe Chang, Zhen Wang, Chenhui Li, and Bo Yuan

  22. [32]

    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. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1930–1939

    2018

  23. [33]

    Xiao Ma, Liqin Zhao, Guan Huang, Zhi Wang, Zelin Hu, Xiaoqiang Zhu, and Kun Gai. 2018. Entire space multi-task model: An effective approach for estimating post-click conversion rate. InThe 41st International ACM SIGIR Conference on Research & Development in Information Retriev...

    2018

  24. [34]

    Kelong Mao, Jieming Zhu, Liangcai Su, Guohao Cai, Yuru Li, and Zhenhua Dong

  25. [35]

    Wentao Ouyang, Xiuwu Zhang, Shukui Ren, Chao Qi, Zhaojie Liu, and Yanlong Du. 2019. Representation learning-assisted click-through rate prediction.arXiv preprint arXiv:1906.04365(2019)

    2019 arXiv

  26. [36]

    Junwei Pan, Jian Xu, Alfonso Lobos Ruiz, Wenliang Zhao, Shengjun Pan, Yu Sun, and Quan Lu. 2018. Field-weighted factorization machines for click-through rate prediction in display advertising. InProceedings of the 2018 world wide web conference. 1349–1357

    2018

  27. [37]

    Yanru Qu, Han Cai, Kan Ren, Weinan Zhang, Yong Yu, Ying Wen, and Jun Wang

  28. [38]

    Steffen Rendle. 2010. Factorization machines. In2010 IEEE International conference on data mining. IEEE, 995–1000

    2010

  29. [39]

    Steffen Rendle, Walid Krichene, Li Zhang, and John Anderson. 2020. Neural collaborative filtering vs. matrix factorization revisited. InProceedings of the 14th ACM Conference on Recommender Systems. 240–248

    2020

  30. [40]

    Matthew Richardson, Ewa Dominowska, and Robert Ragno. 2007. Predicting clicks: estimating the click-through rate for new ads. InProceedings of the 16th international conference on World Wide Web. 521–530

    2007

  31. [41]

    Lei Sang, Qiuze Ru, Honghao Li, Yiwen Zhang, Qian Cao, and Xindong Wu. 2024. Feature Interaction Fusion Self-Distillation Network For CTR Prediction.arXiv preprint arXiv:2411.07508(2024)

    2024 arXiv

  32. [42]

    Tingjia Shen, Hao Wang, Chuhan Wu, Jin Yao Chin, Wei Guo, Yong Liu, Huifeng Guo, Defu Lian, Ruiming Tang, and Enhong Chen. 2024. Optimizing Sequential Recommendation Models with Scaling Laws and Approximate Entropy.arXiv preprint arXiv:2412.00430(2024)

    2024 arXiv

  33. [43]

    In2016 IEEE 16th international conference on data mining (ICDM)

    Product-based neural networks for user response prediction. In2016 IEEE 16th international conference on data mining (ICDM). IEEE, 1149–1154

  34. [44]

    Qingquan Song, Dehua Cheng, Hanning Zhou, Jiyan Yang, Yuandong Tian, and Xia Hu. 2020. Towards automated neural interaction discovery for click-through rate prediction. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 945–955

    2020

  35. [45]

    Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. 2019. Autoint: Automatic feature interaction learning via self- attentive neural networks. InProceedings of the 28th ACM international conference DLF: Enhancing Explicit-Implicit Interac...

    2019

  36. [46]

    Junxiong Tong, Mingjia Yin, Hao Wang, Qiushi Pan, Defu Lian, and Enhong Chen

  37. [47]

    Fangye Wang, Hansu Gu, Dongsheng Li, Tun Lu, Peng Zhang, and Ning Gu

  38. [48]

    Hao Wang, Wei Guo, Luankang Zhang, Jin Yao Chin, Yufei Ye, Huifeng Guo, Yong Liu, Defu Lian, Ruiming Tang, and Enhong Chen. 2025. Generative Large Recommendation Models: Emerging Trends in LLMs for Recommendation.arXiv preprint arXiv:2502.13783(2025)

    2025 arXiv

  39. [49]

    Tingjia Shen, Hao Wang, Jiaqing Zhang, Sirui Zhao, Liangyue Li, Zulong Chen, Defu Lian, and Enhong Chen. 2024. Exploring user retrieval integration towards large language models for cross-domain sequential recommendation.arXiv preprint arXiv:2406.03085(2024)

    2024 arXiv

  40. [50]

    Hao Wang, Defu Lian, Hanghang Tong, Qi Liu, Zhenya Huang, and Enhong Chen. 2021. Decoupled representation learning for attributed networks.IEEE Transactions on Knowledge and Data Engineering35, 3 (2021), 2430–2444

    2021

  41. [51]

    Hao Wang, Defu Lian, Hanghang Tong, Qi Liu, Zhenya Huang, and Enhong Chen. 2021. Hypersorec: Exploiting hyperbolic user and item representations with multiple aspects for social-aware recommendation.ACM Transactions on Information Systems (TOIS)40, 2 (2021), 1–28

    2021

  42. [52]

    Hao Wang, Tong Xu, Qi Liu, Defu Lian, Enhong Chen, Dongfang Du, Han Wu, and Wen Su. 2019. MCNE: An end-to-end framework for learning multiple conditional network representations of social network. InProceedings of the 25th ACM SIGKDD international conference on knowledge disco...

    2019

  43. [53]

    InInternational Conference on Web Information Systems Engineering

    MDAP: A Multi-view Disentangled and Adaptive Preference Learning Framework for Cross-Domain Recommendation. InInternational Conference on Web Information Systems Engineering. Springer, 164–178

  44. [54]

    Kefan Wang, Hao Wang, Kenan Song, Wei Guo, Kai Cheng, Zhi Li, Yong Liu, Defu Lian, and Enhong Chen. 2025. A Universal Framework for Compressing Embeddings in CTR Prediction.arXiv preprint arXiv:2502.15355(2025)

    2025 arXiv

  45. [55]

    InProceedings of the 32nd ACM International Conference on Information and Knowledge Management

    Towards deeper, lighter and interpretable cross network for ctr prediction. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 2523–2533

  46. [56]

    Ruoxi Wang, Rakesh Shivanna, Derek Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed Chi. 2021. Dcn v2: Improved deep & cross network and practical lessons for web-scale learning to rank systems. InProceedings of the web conference 2021. 1785–1797

    2021

  47. [57]

    Hao Wang, Yongqiang Han, Kefan Wang, Kai Cheng, Zhen Wang, Wei Guo, Yong Liu, Defu Lian, and Enhong Chen. 2024. Denoising Pre-Training and Customized Prompt Learning for Efficient Multi-Behavior Sequential Recommendation.arXiv preprint arXiv:2408.11372(2024)

    2024 arXiv

  48. [58]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al . 2024. A survey on large language models for recommendation.World Wide Web27, 5 (2024), 60

    2024

  49. [59]

    Shu Wu, Zekun Li, Yunyue Su, Zeyu Cui, Xiaoyu Zhang, and Liang Wang. 2021. GraphFM: Graph factorization machines for feature interaction modeling.arXiv preprint arXiv:2105.11866(2021)

    2021 arXiv

  50. [60]

    Jun Xiao, Hao Ye, Xiangnan He, Hanwang Zhang, Fei Wu, and Tat-Seng Chua

  51. [61]

    Hao Wang, Mingjia Yin, Luankang Zhang, Sirui Zhao, and Enhong Chen. 2025. MF-GSLAE: A Multi-Factor User Representation Pre-training Framework for Dual-Target Cross-Domain Recommendation.ACM Transactions on Information Systems43, 2 (2025), 1–28

    2025

  52. [62]

    Wenjia Xie, Rui Zhou, Hao Wang, Tingjia Shen, and Enhong Chen. 2024. Bridging User Dynamics: Transforming Sequential Recommendations with Schrödinger Bridge and Diffusion Models. InProceedings of the 33rd ACM International Con- ference on Information and Knowledge Management. ...

    2024

  53. [63]

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & cross network for ad click predictions. InProceedings of the ADKDD’17. 1–7

    2017

  54. [64]

    Yi Yang, Baile Xu, Shaofeng Shen, Furao Shen, and Jian Zhao. 2020. Operation- aware neural networks for user response prediction.Neural Networks121 (2020), 161–168

    2020

  55. [65]

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. InProceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval. 165–174

    2019

  56. [66]

    Mingjia Yin, Hao Wang, Wei Guo, Yong Liu, Zhi Li, Sirui Zhao, Zhen Wang, Defu Lian, and Enhong Chen. 2024. Learning partially aligned item representation for cross-domain sequential recommendation.arXiv preprint arXiv:2405.12473 (2024)

    2024 arXiv

  57. [67]

    Mingjia Yin, Hao Wang, Wei Guo, Yong Liu, Suojuan Zhang, Sirui Zhao, Defu Lian, and Enhong Chen. 2024. Dataset regeneration for sequential recommendation. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3954–3965

    2024

  58. [68]

    Mingjia Yin, Hao Wang, Xiang Xu, Likang Wu, Sirui Zhao, Wei Guo, Yong Liu, Ruiming Tang, Defu Lian, and Enhong Chen. 2023. Apgl4sr: A generic framework with adaptive and personalized global collaborative information in sequential recommendation. InProceedings of the 32nd ACM i...

    2023

  59. [69]

    Mingjia Yin, Chuhan Wu, Yufei Wang, Hao Wang, Wei Guo, Yasheng Wang, Yong Liu, Ruiming Tang, Defu Lian, and Enhong Chen. 2024. Entropy law: The story behind data compression and llm performance.arXiv preprint arXiv:2407.06645 (2024)

    2024 arXiv

  60. [70]

    Wenjia Xie, Hao Wang, Luankang Zhang, Rui Zhou, Defu Lian, and Enhong Chen

  61. [71]

    Breaking determinism: Fuzzy modeling of sequential recommendation using discrete state space diffusion model.Advances in Neural Information Processing Systems37 (2024), 22720–22744

    2024

  62. [72]

    Jiaqing Zhang, Mingjia Yin, Hao Wang, Yawen Li, Yuyang Ye, Xingyu Lou, Junping Du, and Enhong Chen. 2025. TD3: Tucker Decomposition Based Dataset Distilla- tion Method for Sequential Recommendation.arXiv preprint arXiv:2502.02854 (2025)

    2025 arXiv

  63. [73]

    Xiang Xu, Hao Wang, Wei Guo, Luankang Zhang, Wanshan Yang, Runlong Yu, Yong Liu, Defu Lian, and Enhong Chen. 2024. Multi-granularity Interest Retrieval and Refinement Network for Long-Term User Behavior Modeling in CTR Prediction.arXiv preprint arXiv:2411.15005(2024)

    2024 arXiv

  64. [74]

    Luankang Zhang, Kenan Song, Yi Quan Lee, Wei Guo, Hao Wang, Yawen Li, Huifeng Guo, Yong Liu, Defu Lian, and Enhong Chen. 2025. Killing Two Birds with One Stone: Unifying Retrieval and Ranking with a Single Generative Rec- ommendation Model. arXiv:2504.16454 [cs.IR] https://arx...

    2025 arXiv

  65. [75]

    Yufei Ye, Wei Guo, Jin Yao Chin, Hao Wang, Hong Zhu, Xi Lin, Yuyang Ye, Yong Liu, Ruiming Tang, Defu Lian, et al. 2025. FuXi-𝛼: Scaling Recommendation Model with Feature Interaction Enhanced Transformer.arXiv preprint arXiv:2502.03036 (2025)

    2025 arXiv

  66. [76]

    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. InProceedings of the AAAI conference on artificial intelligence, Vol. 33. 5941–5948

    2019

  67. [77]

    Guorui Zhou, Chengru Song, Xiaoqiang Zhu, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep Interest Network for Click-Through Rate Prediction. arXiv:1706.06978 [stat.ML] https://arxiv.org/ abs/1706.06978

    2018 arXiv

  68. [78]

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

  69. [79]

    Jieming Zhu, Qinglin Jia, Guohao Cai, Quanyu Dai, Jingjie Li, Zhenhua Dong, Ruiming Tang, and Rui Zhang. 2023. Final: Factorized interaction layer for ctr prediction. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrie...

    2023

  70. [80]

    Buyun Zhang, Liang Luo, Yuxin Chen, Jade Nie, Xi Liu, Daifeng Guo, Yanli Zhao, Shen Li, Yuchen Hao, Yantao Yao, et al. 2024. Wukong: Towards a Scaling Law for Large-Scale Recommendation.arXiv preprint arXiv:2403.02545(2024)

    2024 arXiv

  71. [81]

    Junlin Zhang, Tongwen Huang, and Zhiqi Zhang. 2019. FAT-DeepFFM: Field attentive deep field-aware factorization machine.arXiv preprint arXiv:1905.06336 (2019)

    2019 arXiv

  72. [83]

    Kexin Zhang, Fuyuan Lyu, Xing Tang, Dugang Liu, Chen Ma, Kaize Ding, Xi- uqiang He, and Xue Liu. 2024. Fusion Matters: Learning Fusion in Deep Click- through Rate Prediction Models.arXiv preprint arXiv:2411.15731(2024)

    2024 arXiv

  73. [85]

    Luankang Zhang, Hao Wang, Suojuan Zhang, Mingjia Yin, Yongqiang Han, Jiaqing Zhang, Defu Lian, and Enhong Chen. 2024. A Unified Framework for Adaptive Representation Enhancement and Inversed Learning in Cross-Domain Recommendation. InInternational Conference on Database System...

    2024

  74. [2010]

    Omnipress

    Web-scale bayesian click-through rate prediction for sponsored search advertising in microsoft’s bing search engine. Omnipress

  75. [2016]

    InProceedings of the 1st workshop on deep learning for recommender systems

    Wide & deep learning for recommender systems. InProceedings of the 1st workshop on deep learning for recommender systems. 7–10

  76. [2017]

    arXiv:1708.04617 [cs.LG] https://arxiv.org/abs/ 1708.04617

    Attentional Factorization Machines: Learning the Weight of Feature Inter- actions via Attention Networks. arXiv:1708.04617 [cs.LG] https://arxiv.org/abs/ 1708.04617

    arXiv

  77. [2018]

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

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

  78. [2021]

    InProceedings of the twenty-ninth international conference on international joint conferences on artificial intelligence

    A dual input-aware factorization machine for CTR prediction. InProceedings of the twenty-ninth international conference on international joint conferences on artificial intelligence. 3139–3145

  79. [2023]

    In Proceedings of the AAAI Conference on Artificial Intelligence, Vol

    FinalMLP: an enhanced two-stream MLP model for CTR prediction. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 4552–4560

  80. [2024]

    arXiv preprint arXiv:2407.13349(2024)

    DCNv3: Towards Next Generation Deep Cross Network for CTR Prediction. arXiv preprint arXiv:2407.13349(2024)

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.