Pith. sign in

REVIEW 3 major objections 4 minor 70 references

Player-Team Heterogeneous Interaction Graph Transformer for Soccer Outcome Prediction

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Modeling soccer as typed player-player and team-team interaction graphs improves match outcome prediction, with average accuracy of 52.19 percent.

desk verdict The architecture is interesting, but the team interaction graph likely leaks test outcomes into the features; without a temporal cutoff the main accuracy claim is not trustworthy. read the letter →

arxiv 2507.10626 v1 pith:7FYCU6RX submitted 2025-07-14 cs.LG cs.AI

Lintao Wang , Shiwen Xu , Michael Horton , Joachim Gudmundsson , Zhiyong Wang This is my paper
classification cs.LGcs.AI
keywords socceroutcomepredictionheterogeneousgraphlearningtransformerattentionnetworksmixtureofexpertsplayerinteractionmodelingsportsanalyticsmatch
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that explicitly modeling fine-grained player-player interactions and team-team historical relationships as typed graphs improves pre-match soccer outcome prediction. It argues that prior deep models either flatten features or treat all interactions alike, missing the heterogeneity of event types and the persistence of team strength. On a large real-world soccer dataset, the proposed HIGFormer reports an average accuracy of 52.19 percent, beating the best baseline's 48.33 percent. If true, this supports the broader principle that architecture-level modeling of interaction structure, rather than handcrafted feature engineering, is a productive direction for sports outcome prediction.

What carries the argument

The load-bearing construction is the heterogeneous interaction graph: each historical match is a directed graph whose nodes are the 46 players, whose edges are pass-related and defense-related events with typed features, and whose node types are home or away. A Player Interaction Network encodes this graph twice—once with a local graph attention branch and once with a pure-transformer global branch that uses Laplacian node identifiers and directional edge embeddings—then merges the two with a per-player mixture-of-experts gate. A separate Team Interaction Network builds a directed team graph whose edge weights are historical winning rates and encodes it with graph attention. A Match Comparison Transformer then adds fixed team embeddings to temporally pooled player embeddings, applies self-attention, and predicts the home-team win probability from the difference of average team embeddings.

What would settle it

A reader could rerun the full pipeline with all player interaction edges randomly permuted while preserving the same node features; if average accuracy remains close to 52.19 percent rather than dropping toward the baselines' level, the claim that heterogeneous player interactions drive the improvement is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a two-stream, graph-augmented transformer—one stream for player-event-player graphs per historical match and one for a team-to-team win-rate graph—can learn representations that predict match outcomes better than flat or homogeneous approaches. The paper reports that this design reaches 52.19 percent average accuracy across seven competitions, with the largest gains in loss prediction (68.25 percent), while draw prediction remains weak (24.53 percent), a known hard case. The authors claim the improvement comes specifically from combining local graph convolution with a global graph transformer over heterogeneous edge types, and from fusing those views with a mixture-of-experts gate rather than concatenation.

Load-bearing premise

The load-bearing premise is that pre-training the player encoder on the very match whose events form its graph—predicting that same match's outcome—produces embeddings that transfer to predicting a future match's outcome, and the paper does not verify that this in-match supervision aligns with the pre-match task.

Editorial extensions

If this is right

  • If the reported accuracy holds, explicit heterogeneous interaction modeling gives a concrete accuracy gain over flat-feature and homogeneous-graph baselines on the same event data.
  • The two-stage training procedure shows one practical way to supervise per-match graph encoders when only match-level outcomes are available: pre-train on a single match's outcome, then freeze and pool for history-based prediction.
  • The player-swap experiments imply the learned embeddings can serve as a counterfactual tool: inserting a star player's embedding raises predicted win probability, with smaller effects for low-ranked teams.
  • Draw prediction remains the weak class, suggesting the ordinal MSE setting with a win/draw threshold interval still does not resolve the inherent ambiguity of draws.

Reading between the lines

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

  • The same heterogeneous graph formulation could transfer to other team sports with event logs, such as basketball or ice hockey, where player-event-player edges and team win-rate edges are naturally available; the paper only demonstrates soccer.
  • Because the team graph is built from winning rates and is fixed at test time, replacing it with a time-decaying rating curve might yield a fairer evaluation and stronger out-of-sample results than the paper's static graph.
  • A natural stress test is to collapse all edge types into one and rerun the ablation; if accuracy does not drop, the heterogeneity itself, not the graph machinery, is the questionable component.
  • The substitution analysis points toward a player-valuation metric, but the paper leaves it as a percentage shift in predicted win probability rather than a monetary or points-based quantity.
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 / 4 minor

Summary. The paper proposes HIGFormer, a graph-augmented transformer for pre-match soccer outcome prediction. It combines a Player Interaction Network (heterogeneous per-match event graphs encoded by a local GAT and a global TokenGT-style transformer, fused via a mixture-of-experts gate), a Team Interaction Network (a directed graph whose edge weights are historical winning rates, encoded by a GAT), and a Match Comparison Transformer that pools historical player and team embeddings to predict win/draw/lose. The model is trained in two stages: first the player encoder is pre-trained per match with that match's own outcome as supervision, then the remaining components are trained with a temporally split 80/20 WyScout dataset. The paper reports an average accuracy of 52.19% versus 48.33% for the strongest baseline, and includes ablations, attention visualizations, and a player-substitution application.

Significance. If the reported results survive scrutiny, the paper makes a useful architectural contribution to sports analytics: it explicitly models heterogeneous player-player and team-team interactions, combines local and global graph views with a MoE gate, and demonstrates a practical player-evaluation tool. The paper is also transparent about its draw-prediction weakness and provides ablation and attention analyses. However, the significance is conditional on resolving temporal leakage risks and on establishing that the modest accuracy gap is not noise; as written, the central quantitative claim is not yet supported to the standard required by the field.

major comments (3)
  1. [Sec. 3.3 and Sec. 4.1] The Team Interaction Network constructs a directed graph with edge weights equal to 'historical winning rate' between teams, but the paper never specifies the temporal cutoff for computing these rates. Because the dataset is split chronologically (80/20 per competition, Sec. 4.1), if the team graph is built once per competition using all matches including the test-period matches, then for a test match the edge feature already encodes that match's own outcome. This would make the reported 52.19% versus 48.33% comparison circular rather than predictive. The authors must state explicitly that winning rates are computed only from matches strictly before the target match, describe whether the graph is updated per match or frozen at the training cutoff, and rerun the experiments if the original protocol allowed future information.
  2. [Sec. 3.5 and Sec. 4.1] In Stage 1, the Player Interaction Network is pre-trained on each match's event graph using that same match's outcome as supervision, but the paper does not state that this pre-training is restricted to the training split. If the test-period matches are included in Stage 1, the player encoder has been trained on test labels, which is a form of label leakage that inflates the reported accuracy. The authors must clarify which split is used for Stage 1 and, if necessary, re-run the pipeline with pre-training strictly on training-period matches only.
  3. [Sec. 4.2, Table 2] The central claim that HIGFormer 'significantly outperforms' baselines is not supported by any uncertainty quantification. No standard deviations, confidence intervals, or significance tests over multiple seeds are reported; on a test set of 389 matches, the aggregate gap of 3.86 percentage points corresponds to roughly 15 correct predictions and could plausibly arise from run-to-run variation. Furthermore, the outcome decision thresholds [4/7, 5/7] and the history length T=10 are described as 'empirically' chosen with no validation protocol, meaning the test numbers may be optimistically biased by tuning on the test set. The authors should report variance over multiple seeds and either freeze hyperparameters on a separate validation split or describe a proper nested validation procedure.
minor comments (4)
  1. [Sec. 3.3] The sentence explaining that the team interaction graph is 'generally denser' is imprecise, since the construction keeps only one directed edge per team pair; it would be clearer to say the graph is relatively well connected because each team has edges to many opponents within the competition.
  2. [Sec. 4.2, Table 2] The paper should state explicitly how the 'Avg' values in the Total column are computed; the reported 52.19% appears to be a test-size-weighted mean of the division-level averages rather than the overall micro-averaged accuracy, and the per-class Total accuracies likewise need a clear definition (micro versus macro over divisions).
  3. [Sec. 3.5 and Sec. 4.3] The notation for the loss configuration is inconsistent: Section 3.5 defines target values as 1, 0.5, 0 for win, draw, lose, while Section 4.3 refers to 'MSE with [1, 4/7, 5/7, 1] configuration' without explaining how the bracket values map to targets or thresholds; this should be unified.
  4. [Sec. 4.5, Table 4] The columns 'Win', 'Draw', 'Lose' in Table 4 are ambiguous: the first row for each team appears to give predicted probabilities, while the subsequent rows give percentage-point changes; the table should clearly label these as baseline probabilities and differences, and the rounding should be consistent.

Circularity Check

1 steps flagged · score 6.0 of 10

Team-interaction edge weights are defined as static per-competition aggregates with no temporal cutoff, so the target match's own outcome can enter the prediction through Eq. 11-14.

  1. self definitional [Section 3.3 (Team Interaction Network) and Section 4.1 (Experimental Settings)]
    "We define the team interaction graph as a directed graph G_team = (V_team,E_team) where each node v_team∈V_team represents a soccer team in competition and edge e_team∈E_team represents the historical winning rate of one team against another. ... For example, if team A has won 20 out of 30 games against team B, we create a directed edge from A to B with a value of 2/3. ... Specifically, the dataset is split into 80% training data and 20% test samples for each division, sorted by time."

    The edge feature X_team_E used in Eq. (11) is defined as a static per-competition winning-rate aggregate ("20 out of 30 games"), not as a per-match statistic computed only from matches strictly before the target match. Section 4.1 splits the dataset chronologically, so a static aggregate over a competition includes the test match's own outcome. The test label y_i then enters X_team_E, and through Eq. (11) the team embeddings Z_team, which are added to player embeddings in Eq. (13) and drive the prediction in Eq. (14). The paper never states that edge weights are recomputed with a temporal cutoff for each target match; the assertion that team embeddings are learned from the training set concerns trainable node embeddings, not the raw edge statistic.

full rationale

Aside from the team-graph issue, the derivation is self-contained: the Player Interaction Network is evaluated with temporally pooled historical embeddings (Eq. 12), and the two-stage pretraining on in-match outcomes is a representation-learning objective, not a claim that the same match's graph predicts its own outcome at test time. There is no load-bearing self-citation chain and no imported uniqueness theorem. However, the Team Interaction Network defines its edge input as a static per-competition winning-rate aggregate, and the chronological 80/20 split is not paired with any statement that edge weights are computed from matches strictly before the target match. Under the literal construction, a test match's own outcome is part of the "historical winning rate" for that pair, and thus enters Eqs. 11-14 as an input. Since the ablation shows removing the Team Interaction Network drops total accuracy from 52.19 to 48.84, the headline gain over the 48.33 baseline is substantially carried by this potentially label-containing feature. That is a partial self-definitional circularity until a temporal cutoff is specified. The absence of confidence intervals or significance tests is a statistical reporting issue, not a circularity, but it compounds the uncertainty in the claimed improvement.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The model rests on standard deep learning components and publicly available data. The main uncharged premises are the ordinal encoding of outcomes, the sufficiency of two event-edge types, and the transferability of in-match pre-training to pre-match prediction. The empirically tuned history length and decision thresholds are the most consequential free parameters.

free parameters (3)
  • Match history length T = 10
    Chosen empirically from {5, 10, 20} in Section 4.1; affects how much history is pooled for player embeddings.
  • Outcome decision thresholds = [0, 4/7, 5/7, 1]
    Set empirically in Section 4.1 'due to difficulty in draw outcome prediction'; used to convert continuous output to win/draw/lose for accuracy, and no validation split is described.
  • Number of event edge types = 2 (pass-related, defense-related)
    Modeling choice in Section 3.2 limiting the player interaction graph to two edge types; if other actions (shots, dribbles) matter, the graph is incomplete.
assumptions (5)
  • domain assumption Soccer match outcomes can be treated as an ordinal variable (lose < draw < win) and optimized with MSE loss.
    Section 3.5 applies MSE to values {0, 0.5, 1}; this ordinal encoding is assumed to capture class relationships better than cross-entropy.
  • domain assumption Historical player event counts and past match outcomes are sufficient input to predict future match outcomes.
    Eq (1) in Section 3.1 defines h as only event counts and outcomes, ignoring injuries, tactics, venue, weather, and other factors.
  • ad hoc to paper Pre-training the player encoder on in-match outcome prediction transfers to pre-match outcome prediction.
    Section 3.5 Stage 1 supervises the player graph encoder with the same match's outcome; the paper does not validate that this in-match supervision aligns with the target pre-match task.
  • ad hoc to paper A single directed edge per team pair, using the higher historical winning rate, captures team-level competition dynamics.
    Section 3.3 drops the lower-winning-rate direction and ignores match margins, venue, and recency.
  • standard math Laplacian eigenvector identifiers from TokenGT remain valid when extended to directed heterogeneous graphs with augmented directional encodings.
    The method in Section 3.2 extends TokenGT's structural identifiers to directed heterogeneous graphs without formal justification of the extension's discriminative power.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Player-Team Heterogeneous Interaction Graph Transformer for Soccer Outcome Prediction." pith.science (2026). https://pith.science/paper/7FYCU6RX

@misc{pith2026250710626,
  author       = {Pith},
  title        = {Pith review of: Player-Team Heterogeneous Interaction Graph Transformer for Soccer Outcome Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7FYCU6RX}},
  note         = {Machine review of arXiv:2507.10626}
}
read the original abstract

Predicting soccer match outcomes is a challenging task due to the inherently unpredictable nature of the game and the numerous dynamic factors influencing results. While it conventionally relies on meticulous feature engineering, deep learning techniques have recently shown a great promise in learning effective player and team representations directly for soccer outcome prediction. However, existing methods often overlook the heterogeneous nature of interactions among players and teams, which is crucial for accurately modeling match dynamics. To address this gap, we propose HIGFormer (Heterogeneous Interaction Graph Transformer), a novel graph-augmented transformer-based deep learning model for soccer outcome prediction. HIGFormer introduces a multi-level interaction framework that captures both fine-grained player dynamics and high-level team interactions. Specifically, it comprises (1) a Player Interaction Network, which encodes player performance through heterogeneous interaction graphs, combining local graph convolutions with a global graph-augmented transformer; (2) a Team Interaction Network, which constructs interaction graphs from a team-to-team perspective to model historical match relationships; and (3) a Match Comparison Transformer, which jointly analyzes both team and player-level information to predict match outcomes. Extensive experiments on the WyScout Open Access Dataset, a large-scale real-world soccer dataset, demonstrate that HIGFormer significantly outperforms existing methods in prediction accuracy. Furthermore, we provide valuable insights into leveraging our model for player performance evaluation, offering a new perspective on talent scouting and team strategy analysis.

Figures

Figures reproduced from arXiv: 2507.10626 by the authors.

Figure 1
Figure 1. A soccer match can be modeled as an interaction [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of HIGFormer, which consists of three major components: (1) Player Interaction Network for historical [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. HIGFormer two-stage training where fire icon indi [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of the attention weights for Player [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 56 canonical work pages

  1. [1]

    Annual Review of Football Finance

    2023. Annual Review of Football Finance. https://www2.deloitte.com/uk/ en/pages/sports-business-group/articles/annual-review-of-football-finance- europe.html. Accessed: 2024-04-29

    work page 2023

  2. [2]

    Jassim AlMulla, Mohammad Tariqul Islam, Hamada RH Al-Absi, and Tanvir Alam

    work page

  3. [3]

    Uri Alon and Eran Yahav. 2020. On the bottleneck of graph neural networks and its practical implications. arXiv preprint arXiv:2006.05205 (2020)

    arXiv 2020

  4. [4]

    Gabriel Anzer, Pascal Bauer, Ulf Brefeld, and Dennis Faßmeyer. 2022. Detection of tactical patterns using semi-supervised graph neural networks. In 16th MIT Sloan Sports Analytics Conference. 1–15

    work page 2022

  5. [5]

    S Mohammad Arabzad, Mohamad Ebrahim Tayebi Araghi, Soheil Sadi-Nezhad, and Nooshin Ghofrani. 2014. Football match results prediction using artificial neural networks; the case of Iran Pro League. Journal of Applied Research on Industrial Engineering 1, 3 (2014), 159–179

    work page 2014

  6. [6]

    Ryan Beal, Stuart E Middleton, Timothy J Norman, and Sarvapali D Ramchurn

    work page

  7. [7]

    Daniel Berrar, Philippe Lopes, and Werner Dubitzky. 2019. Incorporating domain knowledge in machine learning for soccer outcome prediction. Machine Learning 108 (2019), 97–126

    work page 2019

  8. [8]

    Rory Bunker and Teo Susnjak. 2022. The application of machine learning tech- niques for predicting match results in team sport: A review. Journal of Artificial Intelligence Research 73 (2022), 1285–1322

    work page 2022

Show all 70 references
  1. [9]

    Dexiong Chen, Leslie O’Bray, and Karsten Borgwardt. 2022. Structure-aware transformer for graph representation learning. In International Conference on Machine Learning. PMLR, 3469–3489

    2022

  2. [10]

    Anthony C Constantinou. 2019. Dolores: a model that predicts football match outcomes from all over the world. Machine Learning 108, 1 (2019), 49–75

    2019

  3. [11]

    Anthony Costa Constantinou and Norman Elliott Fenton. 2013. Determining the level of ability of football teams by dynamic ratings based on the relative discrepancies in scores between adversaries. Journal of Quantitative Analysis in Sports 9, 1 (2013), 37–50

    2013

  4. [12]

    Norbert Danisik, Peter Lacko, and Michal Farkas. 2018. Football match prediction using players attributes. In 2018 World Symposium on Digital Intelligence for Systems and Machines. IEEE, 201–206

    2018

  5. [13]

    Werner Dubitzky, Philippe Lopes, Jesse Davis, and Daniel Berrar. 2019. The open international soccer database for machine learning. Machine Learning 108 (2019), 9–28

    2019

  6. [14]

    Vijay Prakash Dwivedi and Xavier Bresson. 2020. A generalization of transformer networks to graphs. arXiv preprint arXiv:2012.09699 (2020)

    2020 arXiv

  7. [15]

    William Fedus, Barret Zoph, and Noam Shazeer. 2022. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research 23, 120 (2022), 1–39

    2022

  8. [16]

    Xinyu Fu, Jiani Zhang, Ziqiao Meng, and Irwin King. 2020. Magnn: Metap- ath aggregated graph neural network for heterogeneous graph embedding. In Proceedings of the Web Conference 2020 . 2331–2341

    2020

  9. [17]

    Linxia Gong, Xiaochuan Feng, Dezhi Ye, Hao Li, Runze Wu, Jianrong Tao, Changjie Fan, and Peng Cui. 2020. Optmatch: Optimized matchmaking via modeling the high-order interactions on the arena. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery...

    2020

  10. [18]

    Yin Gu, Qi Liu, Kai Zhang, Zhenya Huang, Runze Wu, and Jianrong Tao. 2021. Neuralac: Learning cooperation and competition effects for match outcome prediction. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 35. 4072–4080

    2021

  11. [19]

    Yin Gu, Kai Zhang, Qi Liu, Xin Lin, Zhenya Huang, and Enhong Chen. 2023. MassNE: Exploring Higher-Order Interactions with Marginal Effect for Massive Battle Outcome Prediction. In Proceedings of the ACM Web Conference 2023. 2710– 2718

    2023

  12. [20]

    Shuo Guan and Xiaochen Wang. 2022. Optimization analysis of football match prediction model based on neural network. Neural Computing and Applications (2022), 1–17

    2022

  13. [21]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. Advances in Neural Information Processing Systems 30 (2017)

    2017

  14. [22]

    Ralf Herbrich, Tom Minka, and Thore Graepel. 2006. TrueSkill™: a Bayesian skill rating system. Advances in Neural Information Processing Systems 19 (2006)

    2006

  15. [23]

    Victoria J Hodge, Sam Devlin, Nick Sephton, Florian Block, Peter I Cowling, and Anders Drachen. 2019. Win prediction in multiplayer esports: Live professional match prediction. IEEE Transactions on Games 13, 4 (2019), 368–379

    2019

  16. [24]

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. 2020. Heterogeneous graph transformer. In Proceedings of the Web Conference 2020 . 2704–2710

    2020

  17. [25]

    Kou-Yuan Huang and Wen-Lung Chang. 2010. A neural network method for pre- diction of 2006 world cup football game. InThe 2010 International Joint Conference on Neural Networks. IEEE, 1–8. KDD ’25, August 3–7, 2025, Toronto, ON, Canada Lintao Wang, Shiwen Xu, Michael Horton, Jo...

    2010

  18. [26]

    Tzu-Kuo Huang, Chih-Jen Lin, and Ruby C Weng. 2006. Ranking individuals by group comparisons. In Proceedings of the 23rd International Conference on Machine Learning. 425–432

    2006

  19. [27]

    Ondřej Hubáček, Gustav Šourek, and Filip Železn `y. 2019. Learning to predict soccer results from relational data with gradient boosted trees. Machine Learning 108 (2019), 29–47

    2019

  20. [28]

    Josip Hucaljuk and Alen Rakipović. 2011. Predicting football scores using machine learning techniques. In 2011 Proceedings of the 34th International Convention MIPRO. IEEE, 1623–1627

    2011

  21. [29]

    Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. 1991. Adaptive mixtures of local experts. Neural Computation 3, 1 (1991), 79–87

    1991

  22. [30]

    Anito Joseph, Norman E Fenton, and Martin Neil. 2006. Predicting football results using Bayesian nets and other machine learning techniques. Knowledge-Based Systems 19, 7 (2006), 544–553

    2006

  23. [31]

    Jinwoo Kim, Dat Nguyen, Seonwoo Min, Sungjun Cho, Moontae Lee, Honglak Lee, and Seunghoon Hong. 2022. Pure transformers are powerful graph learners. Advances in Neural Information Processing Systems 35 (2022), 14582–14595

    2022

  24. [32]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  25. [33]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

    2016 arXiv

  26. [34]

    Kezhi Kong, Jiuhai Chen, John Kirchenbauer, Renkun Ni, C Bayan Bruss, and Tom Goldstein. 2023. GOAT: A global transformer on large-scale graphs. In International Conference on Machine Learning . PMLR, 17375–17390

    2023

  27. [35]

    Verica Lazova and Lasko Basnarkov. 2015. PageRank approach to ranking national football teams. arXiv preprint arXiv:1503.01331 (2015)

    2015 arXiv

  28. [36]

    Hojoon Lee, Dongyoon Hwang, Hyunseung Kim, Byungkun Lee, and Jaegul Choo. 2022. Draftrec: personalized draft recommendation for winning in multi- player online battle arena games. In Proceedings of the ACM Web Conference 2022 . 3428–3439

    2022

  29. [37]

    Yao Li, Minhao Cheng, Kevin Fujii, Fushing Hsieh, and Cho-Jui Hsieh. 2018. Learning from group comparisons: exploiting higher order interactions.Advances in Neural Information Processing Systems 31 (2018)

    2018

  30. [38]

    Qiheng Mao, Zemin Liu, Chenghao Liu, and Jianling Sun. 2023. Hinormer: Representation learning on heterogeneous information networks with graph transformer. In Proceedings of the ACM Web Conference 2023 . 599–610

    2023

  31. [39]

    Rodrigo G Martins, Alessandro S Martins, Leandro A Neves, Luciano V Lima, Edna L Flores, and Marcelo Z do Nascimento. 2017. Exploring polynomial clas- sifier to predict match results in football championships. Expert Systems with Applications 83 (2017), 79–93

    2017

  32. [40]

    Karol Odachowski and Jacek Grekow. 2013. Using bookmaker odds to predict the final result of football matches. In Knowledge Engineering, Machine Learning and Lattice Computing with Applications: 16th International Conference, KES 2012, San Sebastian, Spain, September 10-12, 20...

    2013

  33. [41]

    Kenta Oono and Taiji Suzuki. 2019. Graph neural networks exponentially lose expressive power for node classification. arXiv preprint arXiv:1905.10947 (2019)

    2019 arXiv

  34. [42]

    Luca Pappalardo, Paolo Cintia, Alessio Rossi, Emanuele Massucco, Paolo Fer- ragina, Dino Pedreschi, and Fosca Giannotti. 2019. A public data set of spatio- temporal match events in soccer competitions. Scientific Data 6, 1 (2019), 236

    2019

  35. [43]

    Darwin Prasetio et al . 2016. Predicting football match results with logistic regression. In 2016 International Conference On Advanced Informatics: Concepts, Theory And Application. IEEE, 1–5

    2016

  36. [44]

    Md Ashiqur Rahman. 2020. A deep learning framework for football match prediction. SN Applied Sciences 2, 2 (2020), 165

    2020

  37. [45]

    Ladislav Rampášek, Michael Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. 2022. Recipe for a general, powerful, scalable graph transformer. Advances in Neural Information Processing Systems 35 (2022), 14501–14515

    2022

  38. [46]

    Yu Rong, Yatao Bian, Tingyang Xu, Weiyang Xie, Ying Wei, Wenbing Huang, and Junzhou Huang. 2020. Self-supervised graph transformer on large-scale molecular data. Advances in Neural Information Processing Systems 33 (2020), 12559–12571

    2020

  39. [47]

    Dwijen Rudrapal, Sasank Boro, Jatin Srivastava, and Shyamu Singh. 2020. A deep learning approach to predict football match result. In Computational Intelligence in Data Mining: Proceedings of the International Conference on ICCIDM 2018 . Springer, 93–99

    2020

  40. [48]

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538 (2017)

    2017 arXiv

  41. [49]

    Hamed Shirzad, Ameya Velingker, Balaji Venkatachalam, Danica J Sutherland, and Ali Kemal Sinop. 2023. Exphormer: Sparse transformers for graphs. In International Conference on Machine Learning . PMLR, 31613–31632

    2023

  42. [50]

    Ian Simpson, Ryan J Beal, Duncan Locke, and Timothy J Norman. 2022. Seq2event: Learning the language of soccer using transformer-based match event prediction. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3898–3908

    2022

  43. [51]

    Niek Tax and Yme Joustra. 2015. Predicting the Dutch football competition using public data: A machine learning approach. Transactions on Knowledge and Data Engineering 10, 10 (2015), 1–13

    2015

  44. [52]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in Neural Information Processing Systems 30 (2017)

    2017

  45. [53]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)

    2017 arXiv

  46. [54]

    Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu

  47. [55]

    Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. 2019. Simplifying graph convolutional networks. In International Conference on Machine Learning . PMLR, 6861–6871

    2019

  48. [56]

    Qitian Wu, Wentao Zhao, Chenxiao Yang, Hengrui Zhang, Fan Nie, Haitian Jiang, Yatao Bian, and Junchi Yan. 2024. Simplifying and empowering transformers for large-graph representations. Advances in Neural Information Processing Systems 36 (2024)

    2024

  49. [57]

    Zhanghao Wu, Paras Jain, Matthew Wright, Azalia Mirhoseini, Joseph E Gonzalez, and Ion Stoica. 2021. Representing long-range context for graph neural networks with global attention. Advances in Neural Information Processing Systems 34 (2021), 13266–13279

    2021

  50. [58]

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2018. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826 (2018)

    2018 arXiv

  51. [59]

    Pengyang Yu, Chaofan Fu, Yanwei Yu, Chao Huang, Zhongying Zhao, and Junyu Dong. 2022. Multiplex heterogeneous graph convolutional network. In Proceed- ings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2377–2387

    2022

  52. [60]

    Seongjun Yun, Minbyul Jeong, Raehyun Kim, Jaewoo Kang, and Hyunwoo J Kim

  53. [61]

    Chuxu Zhang, Dongjin Song, Chao Huang, Ananthram Swami, and Nitesh V Chawla. 2019. Heterogeneous graph neural network. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining . 793–803

    2019

  54. [62]

    Daokun Zhang, Jie Yin, Xingquan Zhu, and Chengqi Zhang. 2018. Network representation learning: A survey. IEEE Transactions on Big Data 6, 1 (2018), 3–28

    2018

  55. [63]

    Jiawei Zhang, Haopeng Zhang, Congying Xia, and Li Sun. 2020. Graph-bert: Only attention is needed for learning graph representations. arXiv preprint arXiv:2001.05140 (2020)

    2020 arXiv

  56. [64]

    Advances in Neural Information Processing Systems 32 (2019)

    Graph transformer networks. Advances in Neural Information Processing Systems 32 (2019)

    2019

  57. [65]

    Kai Zhao, Chunjie Du, and Guangxin Tan. 2023. Enhancing basketball game outcome prediction through fused graph convolutional networks and random forest algorithm. Entropy 25, 5 (2023), 765

    2023

  58. [66]

    Chenyi Zhuang and Qiang Ma. 2018. Dual graph convolutional networks for graph-based semi-supervised classification. In Proceedings of the 2018 World Wide Web Conference. 499–508. Player-Team Heterogeneous Interaction Graph Transformer for Soccer Outcome Prediction KDD ’25, Aug...

    2018

  59. [68]

    Chuang Zhao, Hongke Zhao, Yong Ge, Runze Wu, and Xudong Shen. 2022. Winning tracker: a new model for real-time winning prediction in MOBA games. In Proceedings of the ACM Web Conference 2022 . 3387–3395

    2022

  60. [2019]

    InThe World Wide Web Conference

    Heterogeneous graph attention network. InThe World Wide Web Conference. 2022–2032

    2022

  61. [2021]

    In Proceedings of the AAAI Conference on Artificial Intelligence, Vol

    Combining machine learning and human experts to predict match outcomes in football: A baseline model. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 15447–15451

  62. [2023]

    Plos One 18, 8 (2023), e0288933

    SoccerNet: A Gated Recurrent Unit-based model to predict soccer match winners. Plos One 18, 8 (2023), e0288933

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.