Confounding-Valid Conformal Inference for Counterfactual KPIs in Wireless Networks
Abdessamed Qchohi, Jessica Moysen Cortes, Matteo Zecchin
cs.LG · cs.NI · eess.SP · stat.ML
Quick verdict
CV-CCI fuses confounded observational and scarce randomized telemetry for valid counterfactual KPI sets; clean GESPI specialization, simulation-only evidence.
Summary
The paper addresses counterfactual what-if KPI prediction in radio access networks when logged telemetry omits variables the controller used — hidden confounding that invalidates standard conformal counterfactual guarantees. It proposes CV-CCI, instantiating the General Synthetic-Powered Inference (GESPI) principle: scarce randomized telemetry (actions assigned independently of network state) serves as the reliable sample, while abundant, possibly confounded observational telemetry serves as auxiliary data. CV-CCI constructs three weighted-conformal sets — a randomized-only set at level alpha, a guardrail set at alpha+epsilon, and a pooled set combining both sources through propensity-weighted calibration — and outputs their sandwich intersection (Eq. 51). Proposition 1 establishes finite-sample marginal coverage of at least 1−alpha−epsilon under arbitrary hidden confounding, regardless of propensity-estimation errors. In scheduling experiments (Nokia simulator, 8 UEs, 7000 observational and 50 randomized samples) and handover experiments (Sionna RT), CCKE's coverage collapses from 0.91 to 0.78 as confounding strengthens, while CV-CCI stays near nominal; wSCP-DR Exact remains valid but yields wider intervals. CV-CCI's fold-level failure rates are consistently below Guardrail's (26% vs 41% at lambda=1, Table I). Limitations: the guarantee holds only at the relaxed level 1−alpha−epsilon; fold-level violations remain frequent with just 50 randomized samples; all evaluation is simulated; action spaces are binary.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- CV-CCI guarantees finite-sample marginal coverage of at least 1−α−ϵ for counterfactual KPI prediction sets under arbitrary hidden confounding and arbitrary propensity-estimation errors (Proposition 1).
- The GESPI aggregation (Eq. 51) yields a sandwich Γguard ⊆ ΓCV-CCI ⊆ Γrnd, so observational data can only shrink — never invalidate — the randomized-only prediction set.
- In two simulated RAN tasks, methods relying on observational telemetry alone (CCKE, wSCP-DR Inexact) lose coverage as confounding grows (CCKE: 0.91→0.78 scheduling; 0.75 at λ=8 handover), while CV-CCI stays near nominal with narrower sets than wSCP-DR Exact.
- CV-CCI's fold-level failure rates are consistently lower than the Guardrail-only variant across all confounding levels (Tables I-II), evidencing variance reduction from the observational component.
Strengths
Validity by construction, independent of propensity quality
The guardrail set (Eq. 43) is calibrated solely on randomized telemetry with uniform weights (Eq. 41), so Prop. 1's coverage ≥ 1−α−ϵ survives arbitrary confounding and arbitrary errors in the observational propensity model — the right division of labor between the two data sources, stated honestly in §IV.
Realistic confounding scenarios for O-RAN
The hidden confounders — an instantaneous PRB budget set by a slice-level application (§V-B, Fig. 4) and neighbor-cell load not shared with the serving BS (§V-C, Fig. 7) — are plausible, operationally motivated mismatches between controller and telemetry, not toy artifacts.
Guardrail ablation plus failure-rate reporting
Tolerance-parameter sensitivity studied
Reproducibility posture
Weaknesses
Guarantee only at the relaxed level, and per-fold reliability is weak at nrnd=50
Thin theoretical novelty
Simulation-only evaluation with prescribed confounding
Binary action spaces only
Narrow baseline set
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper proposes CV-CCI (Confounding-Valid Counterfactual Conformal Inference), a method for building prediction sets for counterfactual KPIs in radio access networks (RANs) when the controller acted on information absent from the logged telemetry (hidden confounding). Three claims are made: (1) a formulation of counterfactual KPI inference under a controller/telemetry information mismatch (§II); (2) a method with finite-sample marginal coverage of at least 1−α−ϵ under arbitrary hidden confounding and arbitrary propensity-estimation errors (Prop. 1, §IV); (3) empirical evidence from MAC-scheduling and handover simulations that CV-CCI stays valid while producing narrower sets than valid baselines (§V). The framing targets an IEEE-style wireless venue; the statistical machinery is imported from the GESPI framework rather than invented here.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does CV-CCI guarantee 90% coverage for counterfactual KPI prediction sets?
No. The proven guarantee (Proposition 1) is coverage of at least 1−α−ϵ — with the paper's settings α=0.1 and ϵ=0.025, that is 87.5%. Empirically the method lands near 90% in simulations, but the rigorous finite-sample bound is at the relaxed level; the extra ϵ is the price paid for robustness to arbitrary hidden confounding.
What is hidden confounding in this wireless-network setting?
It means the RAN controller selected actions using information that was not written into the telemetry logs — for example an instantaneous PRB budget set by a slice-level application, or the neighboring base station's current load. Because such unlogged variables influence both the chosen action and the resulting KPI, prediction sets calibrated only on observational data can systematically miss the true counterfactual outcome.
Is code or data available for this paper?
Yes. The paper links a public GitHub repository (abdessamedqchohi/Confounding-Valid-Conformal-Counterfactual-Inference) and uses two publicly available simulators: the Nokia Wireless Suite for MAC scheduling and Sionna RT for ray-traced handover. All experiments are simulated, so the pipeline is rerunnable without proprietary network data.
How much randomized telemetry does CV-CCI need?
The paper deliberately uses a very small randomized dataset (50 samples vs 7000 observational), which is the realistic regime. However, with only 50 randomized samples, 7-30% of 100-fold repeats fell below the guaranteed guardrail coverage level (Tables I-II), so applications needing per-deployment reliability may require more randomized data or a larger tolerance ϵ.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Quantitative values in Figures 5, 6, 8 and 9 (boxplot medians, spread, curve shapes) are available in the extracted text only as prose approximations quoted by the authors; exact figure-level numbers could not be read.
- PDF extraction flattens mathematical typography (subscripts, superscripts, quantile operators in Eqs. 24-50); formulas were reconstructed from context, so minor notation-level artifacts are possible, though the derivations cited were unambiguous in the prose.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.