Investigation Report: “基于集成学习的航班延误等级预测方法” (Flight Delay-Level Prediction Method Based on Ensemble Learning)

Academic fraud report · Geng Detector

Summary

This report rates the paper as highly suspicious, with high confidence regarding an apparent methodological flaw and high confidence regarding a numerical inconsistency. Most importantly, the missing-value procedure states that small-airport METAR values were imputed using “other existing features and labels.” If the delay-level label is the prediction target, using it to fill input features constitutes target leakage unless imputation was performed strictly within each training fold before test prediction. The supplied description does not establish that this safeguard was followed, so the validity of the reported performance is seriously undermined and may require retraction of the conclusions if confirmed. A second clear concern appears in the feature-ablation discussion. The paper reports average decreases of 12.84%, 9.07%, and 9.07% after removing prior-delay level, weather score, and airport-busyness features, respectively. The tabulated accuracies of 88.86%, 80.89%, and 81.47% instead yield absolute differences of 7.97 and 7.39 percentage points for the latter two experiments, neither of which equals 9.07%. Table 3 also contains two identical records, which may indicate accidental duplication or a data-cleaning error. The five closely clustered weights are unusual but are not, by themselves, proof of fabrication. No image-level reuse finding could be made from the available text.

Verdict

Overall assessment: highly suspicious. The strongest concern is the reported use of the target label to impute missing input values, which may constitute severe data leakage and could invalidate the evaluation. The paper also contains a reproducible numerical inconsistency in the ablation analysis. These issues warrant requests for the original dataset, exact code, and complete experimental records. Academic misconduct should not be formally established without an institutional investigation.

Key findings

Potential target leakage: Section 2.6 states that missing METAR values at small airports were imputed using “other existing features and labels.” If “labels” refers to flight-delay classes, this improperly exposes target information during feature construction.
Ablation-result inconsistency: The text reports that removing weather and airport-busyness features each reduced evaluation metrics by 9.07%. Table 7 lists accuracies of 88.86%, 80.89%, and 81.47%; the corresponding absolute decreases from the full-feature result are 7.97% and 7.39%, not 9.07%.
Unreconciled aggregate claim: The text also gives three average decreases—12.84%, 9.07%, and 9.07%—but the underlying per-metric values are not supplied in the report, so the stated 12.84% average cannot be independently verified.
Possible duplicated records: The final two rows of Table 3 are identical: 1316 | FU6620 | NTG | FOC | 21:00 | 21:01 | 22:30 | 22:11 | -19 | -6 | 3 | 4.
Unusually similar model weights: The reported weights are w1=0.2037, w2=0.1973, w3=0.1985, w4=0.1977, and w5=0.2028. Their narrow range is suspicious and may indicate fabricated, rounded, or non-performance-derived values, but it is not conclusive.

Evidence highlights

The reported target-leakage statement appears in Section 2.6, “Overall Prediction Process (1) Missing-Value Handling,” Page 4.
The inconsistent percentage claims appear in Section 3.3, “Classification Experiment,” Page 6, Table 7 and related text.
The allegedly duplicated entries appear in Section 2.5, Table 3, “Flight-Loop-Related Features and Data Examples,” Page 4.
The five model weights appear in Section 3.2, “Prediction Results,” Page 5.
The claimed accuracy values are:
Full-feature experiment: 88.86%
Experiment without weather: 80.89%
Experiment without airport busyness: 81.47%
No pixel-level image-reuse conclusion can be drawn because the available material contains only textual descriptions of Figures 1–3.

Notes

Preservation of a leakage-free pipeline would require fitting every learned imputation procedure exclusively on training-fold data and then applying the frozen model to validation or test features. Fold-level preprocessing, random seeds, data-partitioning code, and the exact sequence of operations must be inspected. The authors should provide Table 7 in machine-readable form, define whether “下降” means percentage points or relative percentage change, and explain the repeated 9.07% value. They should also explain whether the identical Table 3 rows were deliberately retained and whether deduplication was performed. The reported performance should not be considered reliable until the leakage concern and numerical discrepancies are resolved.