Pith. sign in

REVIEW 4 major objections 6 minor 57 references

Mini-Game Lifetime Value Prediction in WeChat

T0 review · 4 major / 6 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 Pairing graph-learned embeddings with Pareto-balanced training improves WeChat mini-game lifetime value prediction, beating eleven baselines and lifting online GMV by 8.4%.

desk verdict Serious industrial LTV paper with a real A/B test, but the offline graph likely leaks future payment information and the offline numbers should be read skeptically until the authors specify temporal cutoffs. read the letter →

arxiv 2506.11037 v3 pith:AIQGLTH4 submitted 2025-05-21 cs.LG

Aochuan Chen , Yifan Niu , Ziqi Gao , Yujie Sun , Shoujun Liu , Gong Chen , Yang Liu , Jia Li This is my paper · ORCID
classification cs.LG
keywords lifetimevaluepredictiongraphrepresentationlearningParetooptimizationmulti-horizondatasparsitycomputationaladvertisingmini-gamerecommendationA/Btesting
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 establish that two additions to a single LTV prediction backbone — graph representation learning over the user–game payment graph, and Pareto optimization over the 3-, 7-, and 30-day prediction losses — together make sparse multi-horizon lifetime value prediction both more accurate and more stable. The motivating problems are real: in WeChat mini-game advertising only about 0.1% of registered users purchase, so purchase labels are extremely sparse, and the three horizon tasks conflict because their gradients often point against each other. The proposed GRePO-LTV framework reports the best NMAE, AUC, and N-GINI on all three horizons against eleven baselines on a proprietary six-month dataset, and an online A/B test shows an 8.4% increase in average GMV against the production baseline. A sympathetic reader should care because the method attacks two generic bottlenecks of LTV modeling — data sparsity and task conflict — rather than optimizing a single metric.

What carries the argument

The load-bearing machinery is the pairing of graph representation learning with Pareto optimization. The graph side uses the user–game payment graph to build a user–user graph and a game–game graph via two meta-paths, then trains a GNN encoder by masking random edges and node attributes and reconstructing them; this is what supplies the collaborative signal that softens the 0.1% purchase-rate sparsity. The Pareto side treats the 3-, 7-, and 30-day LTV losses as three objectives, computes a non-dominating descent direction in the convex hull of their gradients by solving a quadratic program, and then searches along the Pareto front with preference weights sampled near equal importance; this is what keeps the three horizons from sabotaging each other. Both pieces feed a common backbone: field-weighted embedding and adaptation layers, a temporal interest network, AdaSparse domain filtering, and zero-inflated lognormal heads.

What would settle it

Inspect the WeChat dataset's graph construction: if any edge between a user and a game has a payment timestamp later than that user's registration for that game — the moment from which 3-, 7-, and 30-day labels are counted — then rebuild the graph from strictly earlier payments and rerun the offline comparison; if the N-GINI and AUC margins over the without-graph ablation collapse, the reported graph gain is label leakage rather than collaborative signal.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that data sparsity and multi-objective conflict in LTV prediction can be handled by the same network if the input features are first enriched with collaborative signals and the training loss is chosen on the Pareto front. Specifically, GRePO-LTV constructs a user–game payment bipartite graph, derives user and game embeddings through a masked edge and attribute reconstruction objective, and feeds these embeddings into a backbone with an encoding, adaptation, temporal interaction, and tower stack that ends in zero-inflated lognormal heads for 3-, 7-, and 30-day values. The Pareto stage computes a non-dominating gradient direction inside the convex hull of the three task gradients and then searches over preference rays near the unit vector to pick a model that is balanced across horizons. The paper reports this combination as improving all three horizons simultaneously over eleven baselines, with average NMAE 0.188, AUC 0.763, N-GINI 0.982, and the online A/B test as raising average LTV and GMV by 8.4%.

Load-bearing premise

The claim stands on the assumption that the graph used for representation learning contains only payments made before a user's 3-, 7-, and 30-day prediction window opens; if it also contains payments from inside those windows, the graph embeddings leak the labels they are supposed to predict.

Editorial extensions

If this is right

  • A single shared backbone can serve 3-, 7-, and 30-day LTV predictions simultaneously, with the best reported NMAE, AUC, and N-GINI on every horizon, so advertisers no longer need separate models per window.
  • The graph component specifically buys robustness to data sparsity: when 80% of training samples are dropped, the method's N-GINI declines more slowly than all baselines, and without the graph component it falls to roughly second place.
  • The Pareto component buys training stability: across 20 random seeds, runs with Pareto produce highly correlated AUC vectors (average correlation about 0.94) while runs without Pareto are weakly correlated, meaning the deployed model is less likely to swing between refreshes.
  • In production, the full framework adds 8.4% average GMV against the baseline, and the short-horizon (3-day) metric improves most, consistent with Pareto resolving the short-term/long-term conflict.
  • Removing either component visibly hurts online results: without Pareto the 3-day GMV drops by 0.72%, and without the graph component the gains shrink to about 2.4% average, so the reported improvement is attributed to both modules working together.

Reading between the lines

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

  • A testable extension of the paper's causal story would be to rerun the offline comparison with the payment graph constructed strictly from payments made before each user's 3-, 7-, and 30-day prediction window begins; if the N-GINI gains persist, the collaborative-signal explanation holds, and if they shrink, the current graph leaks future labels.
  • The same recipe — masked graph embeddings plus Pareto balancing over horizon-specific losses — should transfer to other sparse, multi-horizon value tasks such as subscription revenue, loan repayment, or insurance premium prediction, where the same funnel sparsity and horizon conflict appear.
  • Because the offline dataset is proprietary and the graph's time cutoffs are not described, the 8.4% GMV number cannot be independently checked; a public benchmark with timestamped edges would let the community test whether the graph gain is causal.
  • The paper's bias table shows systematic overestimation of low-LTV users and underestimation of high-LTV users; if the Pareto search were extended to include a bias-penalizing objective, the same framework could improve calibration rather than ranking alone.
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 GRePO-LTV, a framework for predicting 3-, 7-, and 30-day lifetime value (LTV) of users in WeChat mini-game advertising. The method combines graph representation learning (GRL) over user-game payment graphs with Pareto optimization to balance the three prediction objectives. The authors report offline results on a proprietary WeChat mini-game dataset showing consistent improvements over eleven baselines, and an online A/B test reporting an 8.4% average GMV increase over the production baseline, with ablation studies for both GRL and Pareto components.

Significance. If valid, the result is practically important: a single backbone that improves all three horizons simultaneously, with robustness to data sparsity, is valuable for industrial LTV-based bidding. The paper has genuine strengths: it evaluates on a large proprietary dataset, compares against eleven baselines, includes an online A/B test with ablations, reports a stability analysis, and documents deployment lessons. The Pareto-optimization application to multi-horizon LTV is well motivated by the observed gradient conflict. However, the credibility of the offline numbers depends on a temporal-consistency assumption that is never stated and is in fact contradicted by a passage in Section 5, and the paper lacks repeated-run error bars and statistical detail for the online test.

major comments (4)
  1. [Section 5; Section 4.1; Figure 3] The paper does not establish that the graph and features used for prediction are temporally valid. Section 5 states that 'we explore the entire lifespan of each game to gather adequate graph data,' while Section 4.1 says the dataset collects users' cumulative payments over 3-, 7-, and 30-day intervals and lists 'total number of payments' as a user feature; Figure 3 additionally labels 'Target Payments' as an input to the TIN module. If the user-game payment graph or any of these features includes payments made during or after the 3/7/30-day prediction windows, then the GRL reconstruction objective and the backbone encode the prediction targets. Because the offline split is a random 7:2:1 split of LTV samples, test users can also appear as nodes with future edges during GRL training. This would inflate every offline number in Table 1 and Figure 4 without transferring to online serving. The authors must specify the exact temporal construction (e.g., graph edges and features restricted to payments before the prediction start) and either confirm the current setup already respects this restriction or rerun the offline experiments under a strict temporal split with a user-disjoint partition.
  2. [Section 4.1] The offline dataset contains only users who made at least one payment during the six-month collection period; users who registered but never paid are absent. The paper motivates the problem with a purchase rate as low as 0.1%, so a deployed LTV model must rank non-payers against payers. Evaluating AUC and N-GINI on a payer-only population cannot validate that capability, and the reported offline margins may not reflect performance on the full registered-user population that the online system serves. Please report the size of the full population versus the payer subset, and evaluate on a sample that includes non-paying users, or explicitly reframe the claims as conditional LTV prediction among eventual payers.
  3. [Table 1; Section 4.7] The headline offline results and the online A/B test are reported without uncertainty quantification. The '±' in Table 1 is the standard deviation across the three tasks, not across repeated runs; Section 4.6 describes 20 training runs, but only for the correlation analysis of Pareto stability, not for the Table 1 metrics. Table 2 gives relative improvements without confidence intervals, numbers of users, or test duration. Since the offline comparison is the central evidence and the online A/B test is the only externally clean benchmark, both need repeated-run reporting or statistical significance measures before the claimed margins can be assessed.
  4. [Algorithm A1; Section 3.3] The selection of the final Pareto model is unspecified. Algorithm A1 ends with 'Ensure: F* from [F1,...,FK]' but gives no criterion for choosing among the K models, and Section 3.3's Optimal Search only describes how preference vectors are sampled. If F* is selected using test-set metrics, the reported numbers are selection results rather than predictive results. Please specify the selection rule (e.g., best validation NMAE/AUC or a fixed preference vector) and report sensitivity to K and to the sampling range u,v in [1/3, 2/3].
minor comments (6)
  1. [Section 2.1] Definition 2.1 contains a typo: 'Mathmatically' should be 'Mathematically'.
  2. [Section 3.3] Equation (1) is difficult to parse: the definition of the set J* is typeset with KL-divergence conditions whose meaning is not explained, and the role of the hyperparameter epsilon is unclear from the surrounding text.
  3. [Section 4.2] The evaluation metrics are described informally; NMAE is named but its formula is never given, and N-GINI is described only in words. Please provide formal definitions or explicit citations for each metric.
  4. [Figure 3] Figure 3 is overloaded and some labels are ambiguous, particularly 'Target Payments' as an input to the TIN module; please clarify which inputs are available at prediction time and which are labels.
  5. [Section 6] References [50] and [51] appear to be the same WWW 2024 paper by Zhou et al.; please consolidate the duplicate reference.
  6. [Section 4.7] The stability metric Diff is defined on aggregate predictions; please specify whether it is computed per ad or placement and how the 'same sample reflux rate' is controlled across the compared models.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GRePO-LTV's performance claims rest on held-out offline evaluation and an online A/B test, and the method's components are cited to independent prior work.

full rationale

The claimed derivation chain—GRL embeddings from masked edge/attribute reconstruction (Section 3.1), a backbone with FwFM/EPNet/Partitioned Norm/TIN/AdaSparse/ZILN (Section 3.2), and Pareto optimization via MGDA-style QP (Section 3.3)—does not contain any load-bearing step that reduces to its own inputs by construction. The graph embeddings are unsupervised reconstructions, not direct copies of the LTV labels; the Pareto objective is a standard multi-gradient QP cited to Désidéri [11] and Mahapatra and Rajan [25]; ZILN and TIN are cited to independent prior work. The offline evaluation uses a 7:2:1 split of LTV samples (Section 4.1), and the online A/B test (Section 4.7) is an external benchmark with control groups (baseline, w/o GRL, w/o Pareto). Ablations are honest internal comparisons, not fitted parameters renamed as predictions. Some references include works by the same research group (e.g., [21], [45]), but they are background citations for graph representation learning and multi-horizon forecasting, not used to justify the central claim. A genuine validation risk is noted in Section 5 ('we explore the entire lifespan of each game to gather adequate graph data') and Section 4.1, which do not state a temporal cutoff excluding payments inside the 3/7/30-day prediction windows from the graph; if future payments enter the graph, offline gains would be inflated by target leakage. That is a data-consistency and correctness concern, not circular equivalence by construction: no equation in the paper makes the predicted LTV a restatement of its input, and the online A/B test is independent of the graph-construction timestamps. Under the circularity rubric, this paper is self-contained enough to warrant score 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several hyperparameters whose values or selection rules are not reported, and on assumptions about temporal consistency of graph data and the validity of the final-model selection. No new physical or conceptual entities are introduced; the innovation is a configuration of existing techniques.

free parameters (4)
  • GRL loss coefficients xi_e, xi_a, zeta
    These scaling hyperparameters in Section 3.1 control the relative weight of edge and attribute reconstruction; values are not stated, and the final LTV predictions depend on the resulting embeddings.
  • Pareto epsilon (KL constraint)
    Hyperparameter in Eq. (1) that controls when the preference constraint J* is active; no value or sensitivity analysis is given.
  • Pareto search hyperparameters (K, T, eta, sampling range u,v in [1/3, 2/3])
    Algorithm A1 lists these but reports no specific values or selection rule for the final model F*, which is central to the balanced-performance claim.
  • Graph masking ratio and GNN architecture dimensions
    The GRL setup is referenced to HGMAE [33] but no masking ratio, layer count, or hidden size is reported, so exact reproduction is impossible.
assumptions (4)
  • domain assumption User and game embeddings from meta-path graphs (user-game-user and game-user-game) carry collaborative signal useful for LTV prediction.
    Section 3.1 asserts this follows from prior GRL work; the paper provides only indirect evidence via ablation, not a mechanism.
  • domain assumption LTV values follow a zero-inflated lognormal distribution (ZILN).
    Appendix B adopts ZILN from [37] as the predictive head; if the true conditional distribution departs strongly from this, the NMAE rankings could be affected.
  • standard math Desideri's multiple-gradient descent theorem guarantees that the non-dominating gradient direction d_nd = G beta* descends toward the Pareto front.
    Used in Section 3.3 to justify the QP in Eq. (1); this is an established mathematical result, assumed without re-derivation.
  • ad hoc to paper The final model F* chosen from K random preference vectors is selected without using test-set information.
    Algorithm A1 returns F* from [F1,...,FK] but does not state the selection criterion; the paper's reported performance assumes this selection is valid and does not leak test labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mini-Game Lifetime Value Prediction in WeChat." pith.science (2026). https://pith.science/paper/AIQGLTH4

@misc{pith2026250611037,
  author       = {Pith},
  title        = {Pith review of: Mini-Game Lifetime Value Prediction in WeChat},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AIQGLTH4}},
  note         = {Machine review of arXiv:2506.11037}
}
read the original abstract

The LifeTime Value (LTV) prediction, which endeavors to forecast the cumulative purchase contribution of a user to a particular item, remains a vital challenge that advertisers are keen to resolve. A precise LTV prediction system enhances the alignment of user interests with meticulously designed advertisements, thereby generating substantial profits for advertisers. Nonetheless, this issue is complicated by the paucity of data typically observed in real-world advertising scenarios. The purchase rate among registered users is often as critically low as 0.1%, resulting in a dataset where the majority of users make only several purchases. Consequently, there is insufficient supervisory signal for effectively training the LTV prediction model. An additional challenge emerges from the interdependencies among tasks with high correlation. It is a common practice to estimate a user's contribution to a game over a specified temporal interval. Varying the lengths of these intervals corresponds to distinct predictive tasks, which are highly correlated. For instance, predictions over a 7-day period are heavily reliant on forecasts made over a 3-day period, where exceptional cases can adversely affect the accuracy of both tasks. In order to comprehensively address the aforementioned challenges, we introduce an innovative framework denoted as Graph-Represented Pareto-Optimal LifeTime Value prediction (GRePO-LTV). Graph representation learning is initially employed to address the issue of data scarcity. Subsequently, Pareto-Optimization is utilized to manage the interdependence of prediction tasks.

Figures

Figures reproduced from arXiv: 2506.11037 by the authors.

Figure 1
Figure 1. An illustration for the advertising conversion fun [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Gradient conflict in multi-period prediction. We [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An overview of GRePO-LTV. User and game embeddings are first obtained through GRL. Subsequently, GRePO-LTV is [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Effectiveness of the graph representation learning. We present the changes in N-GINI values of different methods [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Effectiveness of the Pareto optimization. The left [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 44 canonical work pages

  1. [1]

    Shaojie Bai, J Zico Kolter, and Vladlen Koltun. 2018. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271 (2018)

    arXiv 2018

  2. [2]

    Josef Bauer and Dietmar Jannach. 2021. Improved customer lifetime value pre- diction with sequence-to-sequence learning and feature-based models. TKDD 15, 5 (2021), 1–37

    work page 2021

  3. [3]

    Albert C Bemmaor and Nicolas Glady. 2012. Modeling purchasing behavior with sudden “death”: A flexible customer lifetime model. Management Science 58, 5 (2012), 1012–1021

    work page 2012

  4. [4]

    George EP Box, Gwilym M Jenkins, Gregory C Reinsel, and Greta M Ljung. 2015. Time series analysis: forecasting and control . John Wiley & Sons

    2015

  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. In KDD. 3795–3804

    work page 2023

  6. [6]

    Hao Chen, Yu Yang, Yuanchen Bei, Zefan Wang, Yue Xu, and Feiran Huang

    work page

  7. [7]

    Pei Pei Chen, Anna Guitart, Ana Fernández del Río, and Africa Periánez. 2018. Customer lifetime value in video games using deep learning and parametric models. In 2018 IEEE international conference on big data (big data) . IEEE, 2134– 2140

    work page 2018

  8. [8]

    Edwin L Crow and Kunio Shimizu. 1987. Lognormal distributions. Marcel Dekker New York

    work page 1987

Show all 57 references
  1. [9]

    Kalyanmoy Deb, Karthik Sindhya, and Jussi Hakanen. 2016. Multi-objective optimization. In Decision sciences. CRC Press, 161–200

    2016

  2. [10]

    Anders Drachen, Mari Pastor, Aron Liu, Dylan Jack Fontaine, Yuan Chang, Julian Runge, Rafet Sifa, and Diego Klabjan. 2018. To be or not to be... social: Incorpo- rating simple social features in mobile game customer lifetime value predictions. In proceedings of the australasia...

    2018

  3. [11]

    Jean-Antoine Désidéri. 2012. Multiple-gradient descent algorithm (MGDA) for multiobjective optimization. Comptes Rendus Mathematique 350, 5 (2012), 313– 318

    2012

  4. [12]

    Peter S Fader, Bruce GS Hardie, and Ka Lok Lee. 2005. RFM and CLV: Using iso-value curves for customer base analysis. Journal of marketing research 42, 4 (2005), 415–430

    2005

  5. [13]

    Counting your customers

    Peter S Fader, Bruce GS Hardie, and Ka Lok Lee. 2005. “Counting your customers” the easy way: An alternative to the Pareto/NBD model. Marketing science 24, 2 (2005), 275–284

    2005

  6. [14]

    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

  7. [15]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In SIGIR. 639–648

    2020

  8. [16]

    Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short- Term Memory. Neural Computation 9, 8 (11 1997), 1735–1780. doi:10.1162/neco.1997.9.8.1735 arXiv:https://direct.mit.edu/neco/article- pdf/9/8/1735/813796/neco.1997.9.8.1735.pdf

    1997 doi

  9. [17]

    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

  10. [18]

    Jipeng Jin, Zhaoxiang Zhang, Zhiheng Li, Xiaofeng Gao, Xiongwen Yang, Lei Xiao, and Jie Jiang. 2024. Pareto-based Multi-Objective Recommender System with Forgetting Curve. In CIKM. 4603–4611

    2024

  11. [19]

    Bart Larivière and Dirk Van den Poel. 2005. Predicting customer retention and profitability by using random forests and regression forests techniques. Expert systems with applications 29, 2 (2005), 472–484

    2005

  12. [20]

    Kunpeng Li, Guangcui Shao, Naijun Yang, Xiao Fang, and Yang Song. 2022. Billion-user customer lifetime value prediction: an industrial-scale solution from Kuaishou. In CIKM. 3243–3251

    2022

  13. [21]

    Yuhan Li, Xinni Zhang, Linhao Luo, Heng Chang, Yuxiang Ren, Irwin King, and Jia Li. 2025. G-Refer: Graph Retrieval-Augmented Large Language Model for Explainable Recommendation. In Proceedings of the ACM on Web Conference 2025 . 240–251

    2025

  14. [22]

    Xiao Lin, Hongjie Chen, Changhua Pei, Fei Sun, Xuanji Xiao, Hanxiao Sun, Yongfeng Zhang, Wenwu Ou, and Peng Jiang. 2019. A pareto-efficient algo- rithm for multiple objective optimization in e-commerce recommendation. In Proceedings of the 13th ACM Conference on recommender sy...

    2019

  15. [23]

    Yang Liu, Liang Chen, Xiangnan He, Jiaying Peng, Zibin Zheng, and Jie Tang

  16. [24]

    Eugenio Lomurno, Stefano Samele, Matteo Matteucci, and Danilo Ardagna. 2021. Pareto-optimal progressive neural architecture search. In Proceedings of the Ge- netic and Evolutionary Computation Conference Companion . 1726–1734

    2021

  17. [25]

    Debabrata Mahapatra and Vaibhav Rajan. 2020. Multi-task learning with user preferences: Gradient descent with controlled ascent in pareto optimization. In ICML. PMLR, 6597–6607

    2020

  18. [26]

    Erxue Min, Da Luo, Kangyi Lin, Chunzhen Huang, and Yang Liu. 2023. Scenario- adaptive feature interaction for click-through rate prediction. InKDD. 4661–4672

    2023

  19. [27]

    James R Norris. 1998. Markov chains. Number 2. Cambridge university press

    1998

  20. [28]

    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. In WWW. 1349–1357

    2018

  21. [29]

    Xiang-Rong Sheng, Liqin Zhao, Guorui Zhou, Xinyao Ding, Binding Dai, Qiang Luo, Siran Yang, Jingshan Lv, Chi Zhang, Hongbo Deng, et al. 2021. One model to serve all: Star topology adaptive recommender for multi-domain ctr prediction. In CIKM. 4104–4113

    2021

  22. [30]

    Lavneet Singh, Nancy Kaur, and Girija Chetty. 2018. Customer Life Time Value Model Framework Using Gradient Boost Trees with RANSAC Response Regular- ization. 1–8. doi:10.1109/IJCNN.2018.8489710

    2018

  23. [31]

    Hongzu Su, Zhekai Du, Jingjing Li, Lei Zhu, and Ke Lu. 2023. Cross-domain adaptative learning for online advertisement customer lifetime value prediction. In AAAI, Vol. 37. 4605–4613

    2023

  24. [32]

    Jianing Sun, Yingxue Zhang, Chen Ma, Mark Coates, Huifeng Guo, Ruiming Tang, and Xiuqiang He. 2019. Multi-graph convolution collaborative filtering. In ICDM. IEEE, 1306–1311

    2019

  25. [33]

    Yijun Tian, Kaiwen Dong, Chunhui Zhang, Chuxu Zhang, and Nitesh V Chawla

  26. [34]

    Ali Vanderveld, Addhyan Pandey, Angela Han, and Rajesh Parekh. 2016. An engagement-based customer lifetime value system for e-commerce. In KDD. 293–302

    2016

  27. [35]

    Ruize Wang, Hui Xu, Ying Cheng, Qi He, Xing Zhou, Rui Feng, Wei Xu, Lei Huang, and Jie Jiang. 2024. ADSNet: Cross-Domain LTV Prediction with an Adaptive Siamese Network in Advertising. In KDD. 5872–5881

    2024

  28. [36]

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

    2019

  29. [37]

    Xiaojing Wang, Tianqi Liu, and Jingang Miao. 2019. A deep probabilistic model for customer lifetime value prediction. arXiv preprint arXiv:1912.07753 (2019)

    2019 arXiv

  30. [38]

    Yunpeng Weng, Xing Tang, Zhenhao Xu, Fuyuan Lyu, Dugang Liu, Zexu Sun, and Xiuqiang He. 2024. OptDist: Learning Optimal Distribution for Customer Lifetime Value Prediction. In CIKM. 2523–2533

    2024

  31. [39]

    Haolun Wu, Chen Ma, Bhaskar Mitra, Fernando Diaz, and Xue Liu. 2022. A multi-objective optimization framework for multi-stakeholder fairness-aware recommendation. ACM Transactions on Information Systems 41, 2 (2022), 1–29

    2022

  32. [40]

    Ruobing Xie, Yanlei Liu, Shaoliang Zhang, Rui Wang, Feng Xia, and Leyu Lin

  33. [41]

    Mingzhe Xing, Shuqing Bian, Wayne Xin Zhao, Zhen Xiao, Xinji Luo, Cunxiang Yin, Jing Cai, and Yancheng He. 2021. Learning reliable user representations from volatile and sparse data to accurately predict customer lifetime value. In KDD. 3806–3816

    2021

  34. [42]

    Xuejiao Yang, Binfeng Jia, Shuangyang Wang, and Shijie Zhang. 2023. Fea- ture Missing-aware Routing-and-Fusion Network for Customer Lifetime Value Prediction in Advertising. In WSDM. 1030–1038

    2023

  35. [43]

    Xuanhua Yang, Xiaoyu Peng, Penghui Wei, Shaoguo Liu, Liang Wang, and Bo Zheng. 2022. Adasparse: Learning adaptively sparse structures for multi-domain click-through rate prediction. In CIKM. 4635–4639

    2022

  36. [44]

    Junwoo Yun, Wonryeol Kwak, and Joohyun Kim. 2023. Multi Datasource LTV User Representation (MDLUR). In KDD. New York, NY, USA, 5500–5508. doi:10. 1145/3580305.3599871

    2023

  37. [45]

    Jiawen Zhang, Shun Zheng, Xumeng Wen, Xiaofang Zhou, Jiang Bian, and Jia Li. 2024. ElasTST: Towards Robust Varied-Horizon Forecasting with Elastic Time-Series Transformer. arXiv preprint arXiv:2411.01842 (2024)

    2024 arXiv

  38. [46]

    Mingwang Zhang, Kun Huang, and Yanli Lv. 2021. A wide neighborhood arc- search interior-point algorithm for convex quadratic programming with box constraints and linear constraints. Optimization and Engineering (2021), 1–21

    2021

  39. [47]

    Shijie Zhang, Xin Yan, Xuejiao Yang, Binfeng Jia, and Shuangyang Wang. 2023. Out of the Box Thinking: Improving Customer Lifetime Value Modelling via Expert Routing and Game Whale Detection. In CIKM. 3206–3215

    2023

  40. [48]

    Haolin Zhou, Junwei Pan, Xinyi Zhou, Xihua Chen, Jie Jiang, Xiaofeng Gao, and Guihai Chen. 2023. Temporal interest network for click-through rate prediction. arXiv preprint arXiv:2308.08487 (2023)

    2023 arXiv

  41. [49]

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond efficient transformer for long se- quence time-series forecasting. In AAAI, Vol. 35. 11106–11115

    2021

  42. [50]

    Zhiyuan Zhou, Li Lin, Hai Wang, Xiaolei Zhou, Wei Gong, and Shuai Wang

  43. [51]

    Zhiyuan Zhou, Li Lin, Hai Wang, Xiaolei Zhou, Gong Wei, and Shuai Wang. 2024. A Cross Domain Method for Customer Lifetime Value Prediction in Supply Chain Platform. In WWW. 4037–4046. KDD ’25, August 3–7, 2025, Toronto, ON, Canada Aochuan Chen et al

    2024

  44. [52]

    Zhipeng Zhou, Liu Liu, Peilin Zhao, and Wei Gong. 2024. Pareto Deep Long-Tailed Recognition: A Conflict-Averse Solution. In ICLR. OpenReview.net. A Pareto Optimization for Multi-horizon LTV Algorithm A1 Pareto optimization for multi-horizon LTV Require: Step size𝜂 > 0. 1: for𝑘...

    2024

  45. [2020]

    TKDE 34, 9 (2020), 4385–4397

    Modelling high-order social relations for item recommendation. TKDE 34, 9 (2020), 4385–4397

    2020

  46. [2021]

    Personalized approximate pareto-efficient recommendation. In WWW. 3839–3849

  47. [2023]

    In AAAI, Vol

    Heterogeneous graph masked autoencoders. In AAAI, Vol. 37. 9997–10005

  48. [2024]

    In The Web Conference 2024

    A Cross Domain Method for Customer Lifetime Value Prediction in Supply Chain Platform. In The Web Conference 2024. https://openreview.net/forum?id= NwpvPL66Ps

    2024

  49. [2025]

    In Australasian Database Conference

    Graph Neural Patching for Cold-Start Recommendations. In Australasian Database Conference. Springer, 334–346

Pith tools

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