COPF: An Online Framework for Deployment-Stable Counterfactual Fairness in Evolving Graphs

Dongmian Zou; Sheng'en Li

arxiv: 2606.00700 · v1 · pith:X5N6YGY4new · submitted 2026-05-30 · 💻 cs.LG · cs.AI

COPF: An Online Framework for Deployment-Stable Counterfactual Fairness in Evolving Graphs

Sheng'en Li , Dongmian Zou This is my paper

Pith reviewed 2026-06-28 19:16 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords counterfactual fairnessonline learningevolving graphslink recommendationperformative predictionmulticalibrationdoubly robust estimationexposure counterfactuals

0 comments

The pith

COPF bounds exposure-counterfactual group gaps in evolving graphs through residual outcome indistinguishability audits on graph-aware doubly robust estimators.

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

The paper presents COPF as a framework that monitors and controls fairness in online link recommendation on graphs whose structure changes because of the recommendations themselves. It defines fairness in terms of group opportunity gaps measured over exposure counterfactuals, where each candidate link is shown or withheld, and makes these gaps estimable by logging show probabilities during explicit exploration. The central technical step is a noisy transfer theorem that converts audits of residual outcome indistinguishability on graph-aware doubly robust residuals into explicit bounds on those exposure gaps, provided the graph process mixes over time and local interference stays bounded. An online multicalibration auditor paired with a primal-dual controller then uses the audited residuals to adjust the recommendation policy in real time. Experiments on two real temporal graph streams and one synthetic bipartite stream indicate that the resulting controller reduces the largest observed spikes in exposure disparities while leaving ranking utility largely unchanged.

Core claim

COPF defines group-level opportunity gaps over exposure counterfactuals and renders them estimable by logging propensities under explicit exploration. It audits and controls fairness via residual outcome indistinguishability over a configurable auditor family that employs graph-aware doubly robust estimators. A noisy transfer theorem establishes that residual-OI evaluated on the estimated GA-DR residuals implies bounds on the exposure-counterfactual group gaps whenever the underlying process satisfies temporal mixing and bounded local interference. The framework instantiates an online multicalibration auditor together with a primal-dual controller that acts on these residuals to enforce the

What carries the argument

The noisy transfer theorem that converts residual outcome indistinguishability on graph-aware doubly robust residuals into bounds on exposure-counterfactual group gaps under temporal mixing and bounded local interference.

If this is right

Fairness estimates remain valid after the policy is updated because the controller acts on residuals that already account for future graph changes.
Worst-case spikes in group exposure disparities can be reduced by online adjustment without separate post-deployment retraining.
Explicit propensity logging makes counterfactual exposure gaps directly estimable even when recommendations alter the observed data distribution.
The multicalibration auditor can be reconfigured for different protected attributes while reusing the same graph-aware doubly robust estimator.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same residual-audit structure could be applied to other performative recommendation settings, such as content or product suggestion, if propensity logging is feasible.
If real-world social graphs exhibit the required mixing rates, the theorem supplies a practical route to deployment-stable fairness without full causal modeling of the entire graph.
Tightening the local-interference bound or relaxing the mixing requirement would directly widen the class of graphs on which the controller guarantees hold.

Load-bearing premise

The translation from residual outcome indistinguishability to bounds on exposure gaps requires that the graph process mixes temporally and that local interference remains bounded.

What would settle it

A controlled stream in which temporal mixing is deliberately broken while the residual-OI auditor still reports low values, yet measured exposure-counterfactual group gaps exceed the claimed bounds.

Figures

Figures reproduced from arXiv: 2606.00700 by Dongmian Zou, Sheng'en Li.

**Figure 1.** Figure 1: Temporal evolution and key policy ablations. (a) Pre/Deploy/Post trajectories for Base vs. Base+COPF, reporting NDCG@10 and counterfactual gaps (g TE gap, g Cal max) as mean ±1 std across seeds. (b) Exploration-policy stability comparing ϵ-greedy and TOPKSTOCHASTIC at the same exploration rate. (c) Coverage-driven exploration as the minimum exposure target ptar varies. Vertical dashed lines mark phase bou… view at source ↗

**Figure 2.** Figure 2: Certificate alignment and sensitivity. (a) Full-window vs. certificate-aligned reporting using the Residual-OI slice set for fairness metrics (g TE gap, g Cal max, g Min). (b) The same comparison for utility metrics (NDCG@10, MRR, Hits@10). (c) Sensitivity to TGB protected-attribute construction, comparing ID-mod2 with degree/activity splits. Shaded bands denote ±1 std across seeds; vertical dashed lines m… view at source ↗

**Figure 3.** Figure 3: Prequential logging protocol (OPP). At each round we build the candidate set online, apply a stochastic Top-K exposure policy with explicit exploration and propensity logging, observe banditized feedback on exposed candidates, and update the backbone and nuisance models sequentially. All reported counterfactual metrics and certificates are computed prequentially from the logged stream. B. Protocol [PITH_F… view at source ↗

**Figure 4.** Figure 4: Spike statistics and certificate slack (Base vs. COPF). For each setting, the top panel reports phase-wise spikes via the per-seed maximum of the rolling maximum of g TE gap over w = 5 logged checkpoints. The bottom panel reports median certificate slack, BbTE/(g TE gap + εnum), per seed and phase, where BbTE is the logged Residual-OI bound and εnum = 10−12 is used only for numerical stability. I. Propensi… view at source ↗

**Figure 5.** Figure 5: Sensitivity to propensity estimation under matched-run probes. Panel (a) fixes the estimator to IPS and varies oracle, lowerbudget Monte Carlo, stale, and noisy logged propensities. Panel (b) reports the available GA-DR runs over oracle/Monte Carlo and propensity-clipping variants. Lines show Deploy and Post phase means; error bars show standard deviation across seeds when multiple seeds are available. Th… view at source ↗

**Figure 6.** Figure 6: Target-pair feasibility diagnostic on Synthetic-Bipartite with GraphMixer at offset scale 1.0. Panel (a) reports the Deploy/Post feasible-window rate. Panel (b) reports the mean treatment-effect slack g TE gap − ρTE; the dashed zero line denotes exact satisfaction of the TE-gap target, and negative values satisfy the target. Panel (c) reports the mean minimum-effect violation g Min, with zero denoting no v… view at source ↗

read the original abstract

Online link recommendation on evolving graphs is performative: by choosing which candidate links to show users, the system changes which links form and what feedback it later observes. Consequently, fairness estimates from logged outcomes can be misleading and may drift after deployment when the recommendation policy is updated. We introduce COPF (Counterfactual Online Performative Fairness), a decision-layer framework for deployment-stable fairness monitoring and control in online link recommendation. COPF (i) defines group-level opportunity gaps over exposure (shown vs. not shown) counterfactuals, (ii) makes them estimable by explicit exploration and by logging the probability (propensity) that each candidate is shown, and (iii) audits and controls fairness using residual outcome indistinguishability (OI) over a configurable auditor family with graph-aware doubly robust (GA-DR) estimators. We provide a noisy transfer theorem showing that Residual-OI on estimated GA-DR residuals implies bounds on exposure-counterfactual group gaps under temporal mixing and bounded local interference, and we instantiate an online multicalibration auditor together with a primal-dual controller. Experiments on two TGB streams and a controlled synthetic bipartite stream show that COPF reduces worst-case spikes in exposure-counterfactual group disparities with modest impact on ranking utility. Our code is available at https://github.com/lsnnnnnnnn/COPF.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

COPF gives a concrete framework for exposure-counterfactual fairness in performative graph recs, with a transfer theorem and online controller, but the key assumptions look unverified on the TGB streams.

read the letter

COPF targets the problem that link recommendations on evolving graphs change the data they later observe, so standard fairness checks can drift after deployment. The paper defines opportunity gaps over exposure counterfactuals, makes them estimable with graph-aware doubly robust estimators that log propensities, and supplies a noisy transfer theorem: residual outcome indistinguishability on the GA-DR residuals implies bounds on the group gaps when temporal mixing and bounded local interference hold. They pair this with an online multicalibration auditor and primal-dual controller, then run experiments on two TGB streams plus a synthetic bipartite case that show lower worst-case disparity spikes at modest cost to ranking utility. Code is released.

The combination of counterfactual exposure fairness, explicit exploration for estimation, and the online control layer is the actual new piece; prior work on static fairness or non-performative settings does not cover this setup. The experiments are straightforward and the public code helps.

The soft spot is the transfer theorem itself. It only connects observable residuals to the desired counterfactual bounds under temporal mixing and bounded local interference. The stress-test note flags that these are not obviously satisfied on TGB data with rapid structural change, and nothing in the abstract or description shows they were checked or stress-tested for sensitivity. If the assumptions slip, the deployment-stability guarantee reduces to the empirical plots. The configurable auditor family adds flexibility but also means results can shift with the choice of auditors.

This is for researchers working on fairness in live recommendation or dynamic graph systems. A reader who cares about performative effects or online auditing would find the setup and the empirical results useful.

It deserves peer review. The problem is practical, the formal link is stated explicitly, and the experiments are on real streams even if the assumptions need closer examination.

Referee Report

2 major / 2 minor

Summary. The paper introduces COPF, a decision-layer framework for deployment-stable counterfactual fairness in online link recommendation on evolving graphs. It defines group-level opportunity gaps over exposure counterfactuals, makes them estimable via explicit exploration and propensity logging, and audits/controls fairness using residual outcome indistinguishability (OI) over a configurable auditor family with graph-aware doubly robust (GA-DR) estimators. A noisy transfer theorem is stated showing that Residual-OI on estimated GA-DR residuals implies bounds on exposure-counterfactual group gaps under temporal mixing and bounded local interference; an online multicalibration auditor and primal-dual controller are instantiated. Experiments on two TGB streams and a synthetic bipartite stream report reduced worst-case spikes in disparities with modest impact on ranking utility; code is released.

Significance. If the noisy transfer theorem is valid, the framework supplies a theoretically grounded approach to fairness monitoring and control that remains stable under policy updates in performative, evolving-graph settings. This is relevant for real-world recommendation systems. Strengths include the explicit statement of assumptions in the theorem, the release of code, and the use of real TGB streams alongside controlled synthetic data.

major comments (2)

[Abstract / Theorem] Abstract / noisy transfer theorem: The theorem asserts that Residual-OI on estimated GA-DR residuals implies bounds on exposure-counterfactual group gaps under temporal mixing and bounded local interference. No verification, sensitivity analysis, or discussion of whether these assumptions hold on the TGB streams (e.g., estimated mixing times or interference ranges) is provided, which is load-bearing for converting observable Residual-OI into the claimed counterfactual guarantees.
[Experiments] Experiments: The reported reductions in worst-case spikes are presented as supporting deployment stability, yet without checks on the theorem's assumptions the empirical plots alone cannot substantiate the transfer from Residual-OI to group-gap bounds; a direct comparison of observed vs. theorem-predicted gaps under varying mixing/interference regimes would be needed.

minor comments (2)

[Abstract] The abstract introduces GA-DR without expansion; ensure the term is defined on first use in the main text.
[Method] Notation for the auditor family and primal-dual controller could be introduced with a short table of symbols for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, particularly the focus on strengthening the link between the noisy transfer theorem and the empirical results. We address each major comment below.

read point-by-point responses

Referee: [Abstract / Theorem] Abstract / noisy transfer theorem: The theorem asserts that Residual-OI on estimated GA-DR residuals implies bounds on exposure-counterfactual group gaps under temporal mixing and bounded local interference. No verification, sensitivity analysis, or discussion of whether these assumptions hold on the TGB streams (e.g., estimated mixing times or interference ranges) is provided, which is load-bearing for converting observable Residual-OI into the claimed counterfactual guarantees.

Authors: We agree that the assumptions of temporal mixing and bounded local interference are load-bearing for the theorem, and that the original manuscript provides no empirical discussion or verification of these on the TGB streams. In the revision we will add a dedicated subsection (in the theory/experiments bridge) that (i) states how mixing times can be estimated from the observed temporal streams using standard autocorrelation-based methods, (ii) reports rough estimates for the two TGB datasets, and (iii) includes a sensitivity analysis that varies the assumed mixing horizon and interference radius while recomputing the implied group-gap bounds. This will make the transfer from Residual-OI to counterfactual guarantees more transparent. revision: yes
Referee: [Experiments] Experiments: The reported reductions in worst-case spikes are presented as supporting deployment stability, yet without checks on the theorem's assumptions the empirical plots alone cannot substantiate the transfer from Residual-OI to group-gap bounds; a direct comparison of observed vs. theorem-predicted gaps under varying mixing/interference regimes would be needed.

Authors: We concur that the current experiments demonstrate practical disparity reduction but do not directly test the theorem's predictive transfer. We will therefore extend the experimental section with an additional controlled study on the synthetic bipartite stream. In this study we will explicitly modulate the temporal mixing rate and local interference radius, compute both the observed exposure-counterfactual group gaps and the bounds predicted by the noisy transfer theorem from the measured Residual-OI, and plot the two quantities against each other across regimes. This will provide the direct comparison requested and will be reported alongside the existing TGB results. revision: yes

Circularity Check

0 steps flagged

No significant circularity; theorem is a standard implication under stated assumptions

full rationale

The paper's central claim is a noisy transfer theorem establishing that Residual-OI on GA-DR residuals implies bounds on exposure-counterfactual group gaps, conditional on temporal mixing and bounded local interference. This is presented as a derived mathematical result rather than a self-definition, fitted prediction, or reduction to prior self-citations. The auditor and primal-dual controller are instantiated separately, with empirical evaluation on TGB streams providing independent content. No load-bearing step reduces by construction to its own inputs, and the derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The ledger is based on elements explicitly mentioned in the abstract. Full paper would likely reveal more free parameters in the estimators and controller.

free parameters (1)

configurable auditor family
The auditor family is configurable, implying parameters chosen for the application.

axioms (2)

domain assumption Temporal mixing
Assumed in the noisy transfer theorem for bounding the gaps.
domain assumption Bounded local interference
Assumed in the noisy transfer theorem.

pith-pipeline@v0.9.1-grok · 5767 in / 1390 out tokens · 34344 ms · 2026-06-28T19:16:31.023259+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 7 canonical work pages · 1 internal anchor

[1]

Coston, A., Mishler, A., Kennedy, E

URL https://openreview.net/ forum?id=ayPPc0SyLv1. Coston, A., Mishler, A., Kennedy, E. H., and Chouldechova, A. Counterfactual risk assessments, evaluation, and fair- ness. InProceedings of the 2020 conference on fairness, accountability, and transparency, pp. 582–593,

2020
[2]

N., Sculley, D., and Halpern, Y

D’Amour, A., Srinivasan, H., Atwood, J., Baljekar, P. N., Sculley, D., and Halpern, Y . Fairness is not static: deeper understanding of long term fairness via simulation stud- ies.Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency,

2020
[3]

Dud´ık, M., Langford, J., and Li, L

doi: 10.1007/s11257-023-09364-z. Dud´ık, M., Langford, J., and Li, L. Doubly robust policy evaluation and learning. InProceedings of the 28th In- ternational Conference on International Conference on Machine Learning, ICML’11, pp. 1097–1104, Madison, WI, USA,

work page doi:10.1007/s11257-023-09364-z
[4]

ISBN 9781450306195

Omnipress. ISBN 9781450306195. Dwork, C., Kim, M. P., Reingold, O., Rothblum, G. N., and Yona, G. Outcome indistinguishability. InProceedings of the 53rd Annual ACM SIGACT Symposium on Theory of Computing, STOC 2021, pp. 1095–1108,

2021
[5]

Dwork, C., Hays, C., Immorlica, N., Perdomo, J

doi: 10.1145/3406325.3451064. Dwork, C., Hays, C., Immorlica, N., Perdomo, J. C., and Tankala, P. From fairness to infinity: Outcome- indistinguishable (omni)prediction in evolving graphs. InProceedings of Thirty Eighth Conference on Learning Theory, pp. 1564–1637,

work page doi:10.1145/3406325.3451064
[6]

E., Karimi, F., and Wagner, C

Ferrara, A., Noboa, L. E., Karimi, F., and Wagner, C. Link recommendations: Their impact on network structure and minorities.Proceedings of the 14th ACM Web Science Conference 2022,

2022
[7]

Temporal graph benchmark for ma- chine learning on temporal graphs.Advances in Neural Information Processing Systems, 36:2056–2073,

Huang, S., Poursafaei, F., Danovitch, J., Fey, M., Hu, W., Rossi, E., Leskovec, J., Bronstein, M., Rabusseau, G., and Rabbany, R. Temporal graph benchmark for ma- chine learning on temporal graphs.Advances in Neural Information Processing Systems, 36:2056–2073,

2056
[8]

N., Zeng, Z., Pan, S., and Foulds, J

Islam, R., Keya, K. N., Zeng, Z., Pan, S., and Foulds, J. Debiasing career recommendations with neural fair col- laborative filtering. InProceedings of the Web Conference 2021, pp. 3779–3790, New York, NY , USA,

2021
[9]

doi: 10.1145/3442381.3449904

ACM. doi: 10.1145/3442381.3449904. Khajehnejad, A., Khajehnejad, M., Babaei, M., Gummadi, K. P., Weller, A., and Mirzasoleiman, B. Crosswalk: Fairness-enhanced node representation learning. InPro- ceedings of the AAAI Conference on Artificial Intelli- gence, volume 36, pp. 11963–11970,

work page doi:10.1145/3442381.3449904
[10]

Fairgan: Gans-based fairness- aware learning for recommendations with implicit feed- back

Li, J., Ren, Y ., and Deng, K. Fairgan: Gans-based fairness- aware learning for recommendations with implicit feed- back. InProceedings of the ACM Web Conference 2022, pp. 297–307,

2022
[11]

doi: 10.1145/3485447.3511958. Liu, L. T., Dean, S., Rolf, E., Simchowitz, M., and Hardt, M. Delayed impact of fair machine learning. InInternational Conference on Machine Learning,

work page doi:10.1145/3485447.3511958
[12]

and Dalmasso, N

Mishler, A. and Dalmasso, N. Fair when trained, un- fair when deployed: Observable fairness measures are unstable in performative prediction settings.CoRR, abs/2202.05049,

work page arXiv
[13]

Temporal Graph Networks for Deep Learning on Dynamic Graphs

Rossi, E., Chamberlain, B. P., Frasca, F., Eynard, D., Monti, F., and Bronstein, M. M. Temporal graph networks for deep learning on dynamic graphs.ArXiv, abs/2006.10637,

work page internal anchor Pith review Pith/arXiv arXiv 2006
[14]

Santos, F

doi: 10.1609/aaai.v36i7.20789. Santos, F. P., Lelkes, Y ., and Levin, S. A. Link recommen- dation algorithms and dynamics of polarization in online social networks.Proceedings of the National Academy of Sciences, 118,

work page doi:10.1609/aaai.v36i7.20789
[15]

Off-policy learning over heterogeneous information for recommendation

Wang, X., Li, Q., Yu, D., and Xu, G. Off-policy learning over heterogeneous information for recommendation. In Proceedings of the ACM Web Conference 2022, pp. 2348– 2359,

2022
[16]

group on = src; outcome = bandit audit off audit=1000; pre-cal; PD+cov in Deploy/Post GraphMixer: tokens=neighbors

Dataset Groups / outcome Base Base+COPF Backbone notes Synth-bip. group on = src; outcome = bandit audit off audit=1000; pre-cal; PD+cov in Deploy/Post GraphMixer: tokens=neighbors. TGBL-Wiki group on = dst; attr = nodemod(2) audit off audit=200; cf upd=200; pre-cal; PD+cov in Deploy/Post GraphMixer: neighbors=10, tokens=20 (stable). TGBL-Review group on ...

1935

[1] [1]

Coston, A., Mishler, A., Kennedy, E

URL https://openreview.net/ forum?id=ayPPc0SyLv1. Coston, A., Mishler, A., Kennedy, E. H., and Chouldechova, A. Counterfactual risk assessments, evaluation, and fair- ness. InProceedings of the 2020 conference on fairness, accountability, and transparency, pp. 582–593,

2020

[2] [2]

N., Sculley, D., and Halpern, Y

D’Amour, A., Srinivasan, H., Atwood, J., Baljekar, P. N., Sculley, D., and Halpern, Y . Fairness is not static: deeper understanding of long term fairness via simulation stud- ies.Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency,

2020

[3] [3]

Dud´ık, M., Langford, J., and Li, L

doi: 10.1007/s11257-023-09364-z. Dud´ık, M., Langford, J., and Li, L. Doubly robust policy evaluation and learning. InProceedings of the 28th In- ternational Conference on International Conference on Machine Learning, ICML’11, pp. 1097–1104, Madison, WI, USA,

work page doi:10.1007/s11257-023-09364-z

[4] [4]

ISBN 9781450306195

Omnipress. ISBN 9781450306195. Dwork, C., Kim, M. P., Reingold, O., Rothblum, G. N., and Yona, G. Outcome indistinguishability. InProceedings of the 53rd Annual ACM SIGACT Symposium on Theory of Computing, STOC 2021, pp. 1095–1108,

2021

[5] [5]

Dwork, C., Hays, C., Immorlica, N., Perdomo, J

doi: 10.1145/3406325.3451064. Dwork, C., Hays, C., Immorlica, N., Perdomo, J. C., and Tankala, P. From fairness to infinity: Outcome- indistinguishable (omni)prediction in evolving graphs. InProceedings of Thirty Eighth Conference on Learning Theory, pp. 1564–1637,

work page doi:10.1145/3406325.3451064

[6] [6]

E., Karimi, F., and Wagner, C

Ferrara, A., Noboa, L. E., Karimi, F., and Wagner, C. Link recommendations: Their impact on network structure and minorities.Proceedings of the 14th ACM Web Science Conference 2022,

2022

[7] [7]

Temporal graph benchmark for ma- chine learning on temporal graphs.Advances in Neural Information Processing Systems, 36:2056–2073,

Huang, S., Poursafaei, F., Danovitch, J., Fey, M., Hu, W., Rossi, E., Leskovec, J., Bronstein, M., Rabusseau, G., and Rabbany, R. Temporal graph benchmark for ma- chine learning on temporal graphs.Advances in Neural Information Processing Systems, 36:2056–2073,

2056

[8] [8]

N., Zeng, Z., Pan, S., and Foulds, J

Islam, R., Keya, K. N., Zeng, Z., Pan, S., and Foulds, J. Debiasing career recommendations with neural fair col- laborative filtering. InProceedings of the Web Conference 2021, pp. 3779–3790, New York, NY , USA,

2021

[9] [9]

doi: 10.1145/3442381.3449904

ACM. doi: 10.1145/3442381.3449904. Khajehnejad, A., Khajehnejad, M., Babaei, M., Gummadi, K. P., Weller, A., and Mirzasoleiman, B. Crosswalk: Fairness-enhanced node representation learning. InPro- ceedings of the AAAI Conference on Artificial Intelli- gence, volume 36, pp. 11963–11970,

work page doi:10.1145/3442381.3449904

[10] [10]

Fairgan: Gans-based fairness- aware learning for recommendations with implicit feed- back

Li, J., Ren, Y ., and Deng, K. Fairgan: Gans-based fairness- aware learning for recommendations with implicit feed- back. InProceedings of the ACM Web Conference 2022, pp. 297–307,

2022

[11] [11]

doi: 10.1145/3485447.3511958. Liu, L. T., Dean, S., Rolf, E., Simchowitz, M., and Hardt, M. Delayed impact of fair machine learning. InInternational Conference on Machine Learning,

work page doi:10.1145/3485447.3511958

[12] [12]

and Dalmasso, N

Mishler, A. and Dalmasso, N. Fair when trained, un- fair when deployed: Observable fairness measures are unstable in performative prediction settings.CoRR, abs/2202.05049,

work page arXiv

[13] [13]

Temporal Graph Networks for Deep Learning on Dynamic Graphs

Rossi, E., Chamberlain, B. P., Frasca, F., Eynard, D., Monti, F., and Bronstein, M. M. Temporal graph networks for deep learning on dynamic graphs.ArXiv, abs/2006.10637,

work page internal anchor Pith review Pith/arXiv arXiv 2006

[14] [14]

Santos, F

doi: 10.1609/aaai.v36i7.20789. Santos, F. P., Lelkes, Y ., and Levin, S. A. Link recommen- dation algorithms and dynamics of polarization in online social networks.Proceedings of the National Academy of Sciences, 118,

work page doi:10.1609/aaai.v36i7.20789

[15] [15]

Off-policy learning over heterogeneous information for recommendation

Wang, X., Li, Q., Yu, D., and Xu, G. Off-policy learning over heterogeneous information for recommendation. In Proceedings of the ACM Web Conference 2022, pp. 2348– 2359,

2022

[16] [16]

group on = src; outcome = bandit audit off audit=1000; pre-cal; PD+cov in Deploy/Post GraphMixer: tokens=neighbors

Dataset Groups / outcome Base Base+COPF Backbone notes Synth-bip. group on = src; outcome = bandit audit off audit=1000; pre-cal; PD+cov in Deploy/Post GraphMixer: tokens=neighbors. TGBL-Wiki group on = dst; attr = nodemod(2) audit off audit=200; cf upd=200; pre-cal; PD+cov in Deploy/Post GraphMixer: neighbors=10, tokens=20 (stable). TGBL-Review group on ...

1935