QML Meta-Dataset: When Does Quantum Machine Learning Outperform Classical Machine Learning? A Systematic Meta-Analysis across 92 Benchmark Datasets

This dataset supports a systematic meta-analysis investigating the conditions under which Quantum Machine Learning (QML) outperforms classical state-of-the-art (SOTA) models. ## Dataset OverviewThe meta-dataset comprises 92 benchmark datasets extracted from 35 peer-reviewed studies (2025), following the PRISMA 2020 protocol. Source databases include SCOPUS (428), IEEE (312), and arXiv (189). ## Domain Coverage- Software Defect Prediction (29%)- Healthcare / Medical (16%)- Energy / Power (13%)- Weather / Climate (11%)- Others (31%) ## VariablesEach record contains 19 variables: [Independent Variables - 18]- Sample Size: N, SCR, MCS- Structure & Quality: K, IR, D, ED, CF, MV- Complexity & Noise: NLS, DBC, Max_Corr, NL, LLI- QML Suitability: NG, SG, FDR, NCE [Dependent Variable - 1]- Gap_from_SOTA (%): Relative performance improvement of best QML model over classical SOTA baseline. Positive = QML advantage, Negative = Classical advantage. ## Key Findings- Noise_level < 1.5 → QML advantage maximized- Nonlinearity_Gap > -1 → QML advantage increases- Max_Corr < 0.6 → QML advantage present- Best QML architecture: QLSTM family (QSA-QConvLSTM: +40.2%)- Best domains: Energy/Power (+20.4%), Weather/Climate (+18.1%) ## QML Models CoveredQSVM (37%), Hybrid (37%), VQC (26%), QNN (11%) ## Classical SOTA BaselinesRandomForest, LSTM, Transformer, iTransformer, PatchTST, LightAutoML ## Analysis MethodsMultiple regression, Ensemble models (XGBoost, LightGBM, CatBoost, RandomForest, GradientBoosting), SHAP analysis