Macro Graph of Experts for Billion-Scale Multi-Task Recommendation

Feiran Huang; Hao Chen; Hongyu Yao; Huan Gong; Qihua Feng; Qijie Shen; Zhiqing Li; Zijin Hong; Zuobin Ying

REVIEW 4 major objections 4 minor 42 references

Macro Graph of Experts for Billion-Scale Multi-Task Recommendation

T0 review · 4 major / 4 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 MGOE claims to be the first graph neural network architecture that brings relational information into billion-scale multi-task recommendation without the computational cost of running GNNs on each task's graph, and reports large gains…

desk verdict A plausible industrial MTL architecture with a real leakage risk and inconsistent tables; deserves review but needs major clarifications. read the letter →

arxiv 2506.10520 v5 pith:JBWHXHFZ submitted 2025-06-12 cs.IR cs.LG

Hongyu Yao , Zijin Hong , Hao Chen , Zhiqing Li , Qijie Shen , Zuobin Ying , Qihua Feng , Huan Gong

show 1 more author

Feiran Huang

This is my paper

classification cs.IRcs.LG

keywords multi-tasklearninggraphneuralnetworksbillion-scalerecommendersystemclick-throughratepredictionmacromixture-of-expertstaskmergingonlineA/Btest

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

This paper tries to show that graph structure can be added to billion-scale multi-task recommendation without the latency that normally comes with graph neural networks. The proposed framework, Macro Graph of Experts (MGOE), merges the separate graphs formed by different user behaviors into one macro task merging graph, then learns task-specific macro embeddings and mixes them through mixture-of-experts towers. On three public benchmark datasets it reports average AUC improvements over the best baseline of 13.42% on Taobao, 7.84% on QB-video, and 3.65% on QK-article, and an online A/B test on a leading e-commerce platform reports a 16.46% GMV gain over the MMoE baseline. The paper's central promise is that macro-level, cluster-based graph information can capture task patterns cheaply, making graph-based multi-task learning deployable at scale.

What carries the argument

The central object is the Macro Task Merging Graph (MTMG), a single macro graph built by merging task-specific interaction matrices into a weighted matrix via a task-combination scoring function, clustering users and items into macro nodes through preference-based grouping, and connecting macro nodes by weighted macro edges. The MGOE model then uses Macro Graph Experts that aggregate macro-neighbor embeddings through log-smoothed attention weights, a Macro Graph Bottom that mixes task-specific experts by gating, and a Macro Prediction Tower with a cascaded readout and a user macro task lookup table. The lookup-table adjustment, shown in Equation (23), multiplies task predictions by a user-specific weighting derived from macro nodes; it is the piece that injects user macro preference directly into the final score.

What would settle it

Re-run MGOE on Taobao with a strict temporal split, rebuilding the user macro task lookup table from training-period interactions only, and compare average AUC to the same baselines; if the 13.42% gap collapses, the lookup table was carrying the signal.

Watch

Extended reading notes

Core claim

The central claim is that a single macro graph built from merged task interactions can carry the relational signal that task-specific billion-scale graphs would provide, at a fraction of the cost. MGOE constructs the Macro Task Merging Graph (MTMG) by scoring each user-item pair's set of coexisting tasks, clustering users and items into weighted macro nodes, and connecting those nodes by macro edges that aggregate the merged scores. Task-specific macro embeddings are then extracted and passed through Macro Graph Experts, which aggregate macro-neighbor information with log-smoothed attention weights rather than expensive graph convolution. A Macro Prediction Tower fuses the macro and micro representations, cascades the task predictions so earlier tasks inform later ones, and applies a user macro task adjustment that scales the final predictions by a user-level weighting derived from the macro task table. The paper reports that this design beats embedding-based multi-task models and micro-graph models on all three public datasets, and that in online A/B tests it outperforms both an MMoE-based ranker and a single-task macro graph model.

Load-bearing premise

The offline results rest on the assumption that the user macro task lookup table behind the macro task adjustment is constructed from task interactions without leaking the labels being predicted, and that hyperparameters were tuned on a validation split, neither of which the paper explicitly documents.

Editorial extensions

If this is right

Billion-scale multi-task recommenders can incorporate graph structure without GNN-level latency: MGOE's inference time is on par with expert-based models and far below graph-based baselines.
Task graphs can be merged into a single macro graph while still preserving task-specific correlations through task-specific macro embeddings and gating.
The macro prediction tower's cascading readout and macro task adjustment transfer knowledge across tasks, improving sparse-task predictions.
In an industrial A/B test, MGOE raised click, conversion, GMV, and stay-time metrics relative to both an expert-based multi-task model and a single-task macro graph model.

Reading between the lines

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

If the macro task adjustment is rebuilt from a fixed training-time lookup table, the method becomes a clean prior over task tendencies; a strict temporal split would reveal whether the offline gains survive without label leakage.
The same macro graph construction could generalize beyond user-item tasks to other multi-relational prediction problems where different relation types define different tasks.
A natural testable extension is to warm-start new or sparse tasks from the existing macro task table, since the macro nodes encode task-co-occurrence patterns.
The public-dataset evaluations use graphs with tens of thousands of nodes; the billion-scale claim rests on the industrial A/B test, so reproducing the macro-graph construction at true billion-node scale is the decisive external check.

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.

Desk Editor's Note

read the letter

Three things you should know about MGOE. First, the core idea is genuinely useful: instead of running GNNs on billion-scale per-task graphs, it merges all task graphs into one macro graph and learns macro embeddings through lightweight attention-style experts. That makes relational information affordable at serving time, and the efficiency numbers (inference time close to MMoE, parameter count lower than PLE) are credible for the architecture. Second, the empirical claims are too large to take at face value: the main Table 1 reports Taobao average AUC 0.6981 with a 13.42% improvement, while the five-run Table 6 reports 0.6936 with 10.43%. The Favor task jumps between 23.68% and 22.61%. Those are not rounding differences; the paper never explains which table is the headline result. Third, the Macro Task Adjustment (σ(T_u) in Eq. 23) multiplies each task prediction by a user lookup table built from observed task interactions. If the public datasets are split by time and T_u is built only from training-period interactions, fine. But the paper never states the split, and Algorithm 1 line 4 just says "Build macro task table T". Without a temporal boundary, the model can read the sparse-task answer directly from the label-derived lookup table. The ablation says removing adjustment costs about 1% average AUC, which only proves the mechanism is active. That is the load-bearing weak spot. The grid search in Section 4.4 also does not say whether it is done on a validation set or on the test sets, and the online A/B has no significance testing. What the paper does well: the MTMG construction is clearly motivated, the preference-based grouping is a sensible extension of MacGNN's macro idea, and the cascading readout is cheap. The related work is honest about MacGNN and POGCN being predecessors, so the novelty claim is mostly about applying macro graphs to MTL plus the tower machinery. The self-citations are not a problem; MacGNN is the natural baseline. If the leakage concern is resolved and the table inconsistency fixed, this is a solid industrial contribution. As it stands, the evidence does not support the abstract's "breakthrough" language, but the architecture deserves serious evaluation. My recommendation: send to peer review, with major revision requests: state the train/validation/test protocol for each dataset, describe how T_u is built (strictly from pre-evaluation interactions), reconcile Tables 1 and 6, and add significance tests for the A/B metrics. I would not cite it until those questions are answered, but I would bring it to a reading group to discuss the leakage risk.

Referee Report

4 major / 4 minor

Summary. The paper proposes MGOE, a multi-task learning framework for recommendation that constructs a Macro Task Merging Graph (MTMG) by merging task-specific interaction matrices with a preference scoring function, then builds macro nodes through weighted clustering and macro edges via aggregated scores. A Macro Graph Bottom with mixture-of-experts style Macro Graph Experts (MGEs) provides macro-aware representations, and the Macro Prediction Tower combines cascading readout with a user macro task adjustment table to produce final predictions. The authors evaluate MGOE on Taobao, QB-video, and QK-article against thirteen baselines, report ablations, hyperparameter studies, and efficiency comparisons, and present online A/B tests from Alibaba, claiming consistent improvements over embedding-based and graph-based multi-task models.

Significance. If the claimed results are valid, MGOE would be a valuable industrial contribution: it extends macro-graph ideas from single-task CTR prediction (MacGNN) to the multi-task setting, avoiding the per-task billion-scale GNN cost while still injecting relational structure. The paper provides a code link and reports real-world deployment, which are strengths for reproducibility and practical relevance. However, the significance is conditional on resolving the evaluation concerns raised below; the main empirical claims currently rest on an unspecified data split, a potentially leaking lookup table, and inconsistent reporting between the main and appendix tables.

major comments (4)

[Sec. 3.4, Eq. (23); Algorithm 1, line 4] The macro task adjustment multiplies the prediction vector by sigma(T_u), where T_u is a 'user macro task lookup table' built from task interactions. Algorithm 1 builds this table before training, but the paper never specifies a temporal or hold-out split for the three datasets; Appendix A.1 and Table 5 report only aggregate statistics. If T_u is computed from the full dataset, then for a test sample the table entry for the target task encodes the label (e.g., a user who favored the item has T_u[favor]=1), so Eq. (23) directly scales the prediction by label information. This is especially problematic for sparse tasks where the reported gains are largest (Taobao Favor +23.68%, Table 1). The ablation in Table 2 shows only that removing the adjustment changes predictions, not that the lookup table's influence is legitimate. The authors must state the data split, construct T_u (and the MTMG) using only pre-evaluation interactions, and re-run the offline experiments; otherwise the central empirical claim is not supported.
[Table 1 vs. Table 6 (Sections 4.1.4, 4.2)] The same experiments are reported with different numbers in the main text and the appendix. For example, Taobao average AUC for MGOE is 0.6981 in Table 1 but 0.6936±0.0040 in Table 6; the reported improvement over the best baseline is 13.42% in Table 1 and 10.43% in Table 6; baseline values also differ (MacGNN average AUC 0.6155 vs. 0.6281). Section 4.1.4 states that experiments were run five times with standard deviations and shows the result in Table 6, but Table 1 is presented as the main result without error bars. This makes it unclear which numbers support the abstract's claims and undermines the quantitative improvement statements. The authors should reconcile the two tables or state explicitly which one is the official aggregate result.
[Section 4.4 (hyper-parameter study)] The hyperparameter study is described as a grid search 'on the Taobao and QB-video datasets' with no mention of a validation split; none is described anywhere in Section 4.1 or the appendix. If hyperparameters such as the macro node count em were selected using the same test sets reported in Table 1, the test performance is optimistically biased relative to baselines that did not receive such tuning. The authors should specify the validation protocol and demonstrate that the final test sets were not used for model selection.
[Equations (16)-(17), Section 3.3] The attention formulation in the Macro Graph Expert appears incorrect as written. Msim is defined as a per-neighbor score divided by a sum over neighbors, so it collapses to a scalar, and softmax of a scalar is identically 1; the value eZ_v,ep then reduces to V_v·eE_ep, which is not an attention-weighted aggregation. The standard attention (e.g., softmax over macro neighbors ep with exp(score/sqrt(d)) in the numerator and denominator) is presumably intended, but the equations are inconsistent with the claim of an attention mechanism. This makes the MGE component ambiguous and prevents reproduction from the paper alone. Please correct the equations or provide the precise normalization.

minor comments (4)

[Section 3.1, Eq. (9)] The text says 'if the sum of the merging preference score values exceeding s0 surpasses a preset threshold' but the equation counts the number of entries d_vj >= s0, not their sum. Please align the wording with the formula.
[Abstract and Section 1] The claims 'first approach capable of leveraging macro graph embeddings...' and 'first graph neural network architecture for billion-scale multi-task recommendation' should be moderated in light of existing graph-based MTL methods such as MoGENet and MMoCEG and the macro-graph work of MacGNN; the novelty appears to lie in the specific multi-task macro-graph construction and expert design, not in being the first graph-based MTL method.
[Section 4.1.4] The sentence 'We run the experiments five times with different seeds with standard deviation to prevent extreme cases, the result is shown in Table 6 in Appendix A.3' is a run-on sentence and should be split; it also leaves unclear whether Table 1 is the best-run or the mean.
[Section 4.6 and Table 4] The online A/B test description does not state the duration of the test beyond 'two weeks' and does not report sample sizes, confidence intervals, or statistical significance for the relative improvements; adding this information would strengthen the deployment claim.

Circularity Check

1 steps flagged · score 6.0 of 10

Offline macro task adjustment makes predictions a function of the labels: Eq. 23 multiplies by sigma(T_u), and T_u is built from the task interaction (label) matrix, with no stated temporal split.

self definitional [Section 3.4 (Macro Task Adjustment), Eq. (23); Algorithm 1 (lines 1-4); Section 2.1]
"First, based on MTMG, we construct a user macro task lookup table according to different task interactions. Specifically, we associate various user's task interactions with the corresponding macro nodes. For a certain user u, the final prediction Y_ui can be expressed as: Y_ui = W_u ⊙ Y_cat = sigma(T_u) ⊙ [y_hat(t1)_ui ; y_hat(t2)_ui ; ...], (23) where T_u is the users' macro task lookup table"

Section 2.1 states that r^(t)_ui and y^(t)_ui are 'numerical equivalent,' so the task interaction matrices used to build MTMG are the label matrices. Algorithm 1 builds S (Eq. 7), then macro nodes/edges, then the macro task table T, before any training. Eq. 23 then defines the final prediction as sigma(T_u)⊙Y_cat. Thus the predicted vector is, by construction, a Hadamard product with a monotone function of the user's task labels via T_u. If (as the paper does not rule out, since Appendix A.1 gives only aggregate date ranges and no temporal/hold-out split) T includes the evaluation-period interactions, the prediction is partly determined by the labels it is supposed to predict.

full rationale

The core MGOE construction (Eqs. 7-22) is a standard supervised architecture: macro embeddings, gating, and towers are trained with cross-entropy, and no equation in that chain reduces to the label it predicts. The complexity analysis in A.5 is an algebraic derivation from the stated module sizes and is self-contained. The self-citations (notably MacGNN [3], which shares authors with this paper) are used as prior work and as a baseline, not as a uniqueness theorem or as the sole justification for MGOE's design, so they are not load-bearing circularity. The one self-definitional step is in the offline evaluation path: Eq. (23) defines the final prediction as sigma(T_u)⊙Y_cat, and T_u is built (Algorithm 1, line 4) from the MTMG interaction matrix S, which is computed from the task interaction matrices R^(t); Section 2.1 states R^(t) and the labels y^(t) are numerically equivalent. Because the paper never states a train/validation/test or temporal split, nothing in the text excludes evaluation-period labels from T_u. Under that construction, the offline AUC gain is partially forced: the prediction is multiplied by a monotone function of the user's observed task labels, so the reported 13-24% gains on sparse tasks are not evidence that the macro representation itself learned the task structure. The online A/B results and the efficiency study are independent evidence and keep the paper from being entirely circular, hence the score of 6 rather than 8-10.

Assumptions & free parameters 9 free parameters · 5 assumptions · 2 invented entities

The method depends on many tuned hyperparameters and domain assumptions about collapsing multi-task interactions into a macro graph. The key unreported choices are beta, thresholds, w+/w-, gamma, macro node counts, task weights, and regularization strengths. The potential label-derived lookup table is the most consequential assumption.

free parameters (9)

beta (transformation parameter) = Not reported
Controls the rank-to-score transform in Eq. 7 for merging task combinations.
delta1, delta2 (activity thresholds) = Not reported
Classify micro nodes as active or inactive in Eq. 9.
s0 (merging preference score baseline) = Not reported
Threshold for counting active neighbors in Eq. 9.
w+, w- (active/inactive group weights) = Grid searched over (2,0.5), (2,1), (1,0.5), (1,1); final values not explicitly fixed
Weights in preference-based grouping, Eq. 9 and Eq. 10.
gamma (log smoothing temperature) = Best 1.1 on Taobao, 1.5 on QB-video
Temperature in Eq. 15 controlling macro neighbor weighting.
macro user/item node counts (em, en) = Grid searched 5 to 30; best 20 macro user nodes
Number of clusters determines macro graph granularity.
task weights alpha_t = Not reported
Relative importance of each task in the loss, Eq. 2.
lambda1, lambda2 (regularization strengths) = Searched over {1e-4, 5e-5, 1e-5}; final values not reported
L1 and L2 penalties in Eq. 24.
learning rate = Searched over {1e-2, 5e-2, 1e-3, 5e-3}; final value not reported
Optimizer step size for Adam.

assumptions (5)

domain assumption Clustering on S_U and S_I yields stable and meaningful macro nodes.
Section 3.1 Macro Node Computation relies on k-means-like clustering over preference-scored matrices; no convergence or stability analysis is provided.
domain assumption The merging score function preserves information needed for multi-task prediction.
Eq. 7 collapses multi-task interaction combinations into scalar scores; assumes this rank-based reduction does not discard signal needed for CTR/CVR.
domain assumption The user macro task lookup table is a legitimate feature, not a label leak.
Section 3.4 Eq. 23 assumes T_u built from user task interactions can be used at prediction time without contaminating offline labels; no temporal split or exclusion rule is described.
domain assumption Attention over macro neighbors substitutes for GNN message passing.
Section 3.3 uses attention over macro nodes rather than graph convolution; assumes this captures high-order graph information.
standard math Standard ML background: softmax, attention, BCE loss, Adam optimizer.
These components are used as standard building blocks without proof in this paper.

invented entities (2)

MTMG (Macro Task Merging Graph)
purpose: A single merged macro graph representing multiple task-specific interaction graphs for billion-scale MTL.
The graph is only evaluated inside the MGOE pipeline; no independent benchmark validates that the merged graph structure itself is meaningful.
Macro Prediction Tower with cascading readout and macro task adjustment
purpose: Dynamically integrates macro knowledge across tasks and adjusts final predictions using a user macro task lookup table.
The tower is jointly trained with the rest of the model and no isolated falsifiable prediction is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Macro Graph of Experts for Billion-Scale Multi-Task Recommendation." pith.science (2026). https://pith.science/paper/JBWHXHFZ

@misc{pith2026250610520,
  author       = {Pith},
  title        = {Pith review of: Macro Graph of Experts for Billion-Scale Multi-Task Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JBWHXHFZ}},
  note         = {Machine review of arXiv:2506.10520}
}

read the original abstract

Graph-based multi-task learning at billion-scale presents a significant challenge, as different tasks correspond to distinct billion-scale graphs. Traditional multi-task learning methods often neglect these graph structures, relying solely on individual user and item embeddings. However, disregarding graph structures overlooks substantial potential for improving performance. In this paper, we introduce the Macro Graph of Experts (MGOE) framework, the first approach capable of leveraging macro graph embeddings to capture task-specific macro features while modeling the correlations between task-specific experts. Specifically, we propose the concept of a Macro Graph Bottom, which, for the first time, enables multi-task learning models to incorporate graph information effectively. We design the Macro Prediction Tower to dynamically integrate macro knowledge across tasks. MGOE has been deployed at scale, powering multi-task learning for a leading billion-scale recommender system, Alibaba. Extensive offline experiments conducted on three public benchmark datasets demonstrate its superiority over state-of-the-art multi-task learning methods, establishing MGOE as a breakthrough in multi-task graph-based recommendation. Furthermore, online A/B tests confirm the superiority of MGOE in billion-scale recommender systems.

Figures

Figures reproduced from arXiv: 2506.10520 by the authors.

**Figure 2.** In this section, we firstly introduce the utilized MTMG. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 2.** An illustration of the construction of Macro Task Merging Graph and the overview of Macro Graph of Experts. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Figures from the paper (2 more)

**Figure 3.** Figure 3: Parameter study of 𝑤𝑣 on Taobao and QB-video. 0.5 1.0 1.5 0.680 0.685 0.690 0.695 Average AUC Taobao 0.5 1.0 1.5 0.74 0.75 0.76 0.77 0.78 0.79 Average AUC QB-video 5 10 15 20 25 30 m 0.675 0.680 0.685 0.690 0.695 Average AUC 5 10 15 20 25 30 m 0.72 0.74 0.76 0.78 Avera…

**Figure 4.** Figure 4: Parameter study of 𝛾 and𝑚e on Taobao and QB-video. While examining 𝑤𝑣,𝛾,𝑚e, we keep all other hyperparameters fixed. The results are shown in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 34 canonical work pages

[1]

Yuanchen Bei, Weizhi Zhang, Siwen Wang, Weizhi Chen, Sheng Zhou, Hao Chen, Yong Li, Jiajun Bu, Shirui Pan, Yizhou Yu, et al. 2025. Graphs Meet AI Agents: Taxonomy, Progress, and Future Opportunities.arXiv preprint arXiv:2506.18019 (2025)

arXiv 2025
[2]

Rich Caruana. 1997. Multitask learning.Machine Learning(1997)

work page 1997
[3]

Hao Chen, Yuanchen Bei, Qijie Shen, Yue Xu, Sheng Zhou, Wenbing Huang, Feiran Huang, Senzhang Wang, and Xiao Huang. 2024. Macro graph neural networks for online billion-scale recommender systems. InInternational World Wide Web Conference (WWW)

work page 2024
[4]

Junxuan Chen, Baigui Sun, Hao Li, Hongtao Lu, and Xian-Sheng Hua. 2016. Deep ctr prediction in display advertising. InInternational Conference on Multimedia (MM)

work page 2016
[5]

Jiawei Chen, Junkang Wu, Jiancan Wu, Xuezhi Cao, Sheng Zhou, and Xiangnan He. 2023. Adap-𝜏: Adaptively modulating embedding magnitude for recommen- dation. InInternational World Wide Web Conference (WWW)

work page 2023
[6]

Tom Fawcett. 2006. An introduction to ROC analysis.Pattern Recognition Letters (2006)

work page 2006
[7]

Xavier Glorot and Yoshua Bengio. 2010. Understanding the difficulty of training deep feedforward neural networks

work page 2010
[8]

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)

arXiv 2017

Show all 42 references

[9]

Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. InConference on Information Retrieval (SIGIR)

2020
[10]

Binbin Hu, Bin Shen, Ruize Wu, Zhiqiang Zhang, Yuetian Cao, Yong He, Liang Zhang, Linjian Mo, and Jun Zhou. 2022. Mixture of Graph Enhanced Expert Networks for Multi-task Recommendation. InPacific Rim International Conference on Artificial Intelligence (PRICAI)

2022
[11]

Feiran Huang, Yuanchen Bei, Zhenghang Yang, Junyi Jiang, Hao Chen, Qijie Shen, Senzhang Wang, Fakhri Karray, and Philip S Yu. 2025. Large Language Model Simulator for Cold-Start Recommendation. InInternational Conference on Web Search and Data Mining (WSDM)

2025
[12]

Jacobs, Michael I

Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton
[13]

Yuchin Juan, Yong Zhuang, Wei-Sheng Chin, and Chih-Jen Lin. 2016. Field- aware factorization machines for CTR prediction. InConference on Recommender Systems (RecSys)

2016
[14]

Diederik P Kingma. 2014. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980(2014)

2014 arXiv
[15]

Iasonas Kokkinos. 2017. Ubernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. InConference on Computer Vision and Pattern Recognition (CVPR)

2017
[16]

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. InConference on Knowledge Discovery and Data Mining (KDD)

2018
[17]

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. InConference on Information Retrieval (SIGIR)

2018
[18]

Erxue Min, Yu Rong, Tingyang Xu, Yatao Bian, Da Luo, Kangyi Lin, Junzhou Huang, Sophia Ananiadou, and Peilin Zhao. 2022. Neighbour interaction based click-through rate prediction via graph-masked transformer. InConference on Information Retrieval (SIGIR)

2022
[19]

Sebastian Ruder, Joachim Bingel, Isabelle Augenstein, and Anders Søgaard. 2019. Latent multi-task architecture learning. InConference on Artificial Intelligence (AAAI)

2019
[20]

Ozan Sener and Vladlen Koltun. 2018. Multi-task learning as multi-objective optimization. InAdvances in Neural Information Processing Systems (NeurIPS)

2018
[21]

Qijie Shen, Yuanchen Bei, Zihong Huang, Jialin Zhu, Keqin Xu, Boya Du, Jiawei Tang, Yuning Jiang, Feiran Huang, Xiao Huang, et al. 2025. AliBoost: Ecological Boosting Framework in Alibaba Platform. InConference on Knowledge Discovery and Data Mining (KDD)

2025
[22]

Liangcai Su, Junwei Pan, Ximei Wang, Xi Xiao, Shijie Quan, Xihua Chen, and Jie Jiang. 2024. STEM: Unleashing the Power of Embeddings for Multi-task Recommendation. InConference on Artificial Intelligence (AAAI)

2024
[23]

Huinan Sun, Guangliang Yu, Pengye Zhang, Bo Zhang, Xingxing Wang, and Dong Wang. 2022. Graph Based Long-Term And Short-Term Interest Model for Click-Through Rate Prediction. InInternational Conference on Information and Knowledge Management (CIKM)

2022
[24]

Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. 2020. Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations. InConference on Recommender Systems (RecSys)

2020
[25]

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. InAdvances in Neural Information Processing Systems (NeurIPS)

2017
[26]

Weixun Wang, Junqi Jin, Jianye Hao, Chunjie Chen, Chuan Yu, Weinan Zhang, Jun Wang, Xiaotian Hao, Yixi Wang, Han Li, et al. 2019. Learning adaptive display exposure for real-time advertising. InInternational Conference on Information and Knowledge Management (CIKM)

2019
[27]

Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. InConference on Information Retrieval (SI- GIR)

2019
[28]

Yuhao Wang, Ha Tsz Lam, Yi Wong, Ziru Liu, Xiangyu Zhao, Yichao Wang, Bo Chen, Huifeng Guo, and Ruiming Tang. 2023. Multi-task deep recommender systems: A survey.arXiv preprint arXiv:2302.03525(2023)

2023 arXiv
[29]

Dongbo Xi, Zhen Chen, Peng Yan, Yinger Zhang, Yongchun Zhu, Fuzhen Zhuang, and Yu Chen. 2021. Modeling the sequential dependence among audience multi- step conversions with multi-task learning in targeted display advertising. In Conference on Knowledge Discovery and Data Mining (KDD)

2021
[30]

Jiahui Xu, Lu Sun, and Dengji Zhao. 2024. MoME: Mixture-of-Masked-Experts for Efficient Multi-Task Recommendation. InConference on Information Retrieval (SIGIR)

2024
[31]

Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. 2018. Graph convolutional neural networks for web-scale recommender systems. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 974–983

2018
[32]

Guanghu Yuan, Fajie Yuan, Yudong Li, Beibei Kong, Shujie Li, Lei Chen, Min Yang, Chenyun YU, Bo Hu, Zang Li, Yu Xu, and Xiaohu Qie. 2022. Tenrec: A Large-scale Multipurpose Benchmark Dataset for Recommender Systems. In Advances in Neural Information Processing Systems (NeurIPS)

2022
[33]

Cong Zhang, Dongyang Liu, Lin Zuo, Junlan Feng, Chao Deng, Jian Sun, Haitao Zeng, and Yaohong Zhao. 2023. Multi-gate Mixture-of-Contrastive-Experts with Graph-based Gating Mechanism for TV Recommendation. InInternational Conference on Information and Knowledge Management (CIKM)

2023
[34]

Weizhi Zhang, Yuanchen Bei, Liangwei Yang, Henry Peng Zou, Peilin Zhou, Aiwei Liu, Yinghui Li, Hao Chen, Jianling Wang, Yu Wang, et al. 2025. Cold-start rec- ommendation towards the era of large language models (llms): A comprehensive survey and roadmap.arXiv preprint arXiv:25...

2025 arXiv
[35]

Yijie Zhang, Yuanchen Bei, Hao Chen, Qijie Shen, Zheng Yuan, Huan Gong, Senzhang Wang, Feiran Huang, and Xiao Huang. 2024. Multi-behavior collabora- tive filtering with partial order graph convolutional networks. InConference on Knowledge Discovery and Data Mining (KDD)

2024
[36]

Yijie Zhang, Yuanchen Bei, Shiqi Yang, Hao Chen, Zhiqing Li, Lijia Chen, and Feiran Huang. 2023. Alleviating Behavior Data Imbalance for Multi-Behavior Graph Collaborative Filtering. InICDM 2023 First Learning with Knowledge Graphs Workshop (ICDM)

2023
[37]

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. InConference on Artificial Intelligence (AAAI)

2019
[38]

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. InConference on Knowledge Discovery and Data Mining (KDD)

2018
[39]

Han Zhu, Xiang Li, Pengye Zhang, Guozheng Li, Jie He, Han Li, and Kun Gai
[40]

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. InConference on Information Retrieval (SIGIR). A Experimental Details Table 5: Statistics of the experimental ...

2023
[1991]

Adaptive Mixtures of Local Experts.Neural Computation(1991)

1991
[2018]

InConference on Knowledge Discovery and Data Mining (KDD)

Learning tree-based deep model for recommender systems. InConference on Knowledge Discovery and Data Mining (KDD)

Pith tools

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

[1] [1]

Yuanchen Bei, Weizhi Zhang, Siwen Wang, Weizhi Chen, Sheng Zhou, Hao Chen, Yong Li, Jiajun Bu, Shirui Pan, Yizhou Yu, et al. 2025. Graphs Meet AI Agents: Taxonomy, Progress, and Future Opportunities.arXiv preprint arXiv:2506.18019 (2025)

arXiv 2025

[2] [2]

Rich Caruana. 1997. Multitask learning.Machine Learning(1997)

work page 1997

[3] [3]

Hao Chen, Yuanchen Bei, Qijie Shen, Yue Xu, Sheng Zhou, Wenbing Huang, Feiran Huang, Senzhang Wang, and Xiao Huang. 2024. Macro graph neural networks for online billion-scale recommender systems. InInternational World Wide Web Conference (WWW)

work page 2024

[4] [4]

Junxuan Chen, Baigui Sun, Hao Li, Hongtao Lu, and Xian-Sheng Hua. 2016. Deep ctr prediction in display advertising. InInternational Conference on Multimedia (MM)

work page 2016

[5] [5]

Jiawei Chen, Junkang Wu, Jiancan Wu, Xuezhi Cao, Sheng Zhou, and Xiangnan He. 2023. Adap-𝜏: Adaptively modulating embedding magnitude for recommen- dation. InInternational World Wide Web Conference (WWW)

work page 2023

[6] [6]

Tom Fawcett. 2006. An introduction to ROC analysis.Pattern Recognition Letters (2006)

work page 2006

[7] [7]

Xavier Glorot and Yoshua Bengio. 2010. Understanding the difficulty of training deep feedforward neural networks

work page 2010

[8] [8]

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)

arXiv 2017