{
    "model_name": "GradientBoosting_ActiveLearning",
    "approach": "GradientBoosting + Active Learning Simulation",
    "base_models": "LogReg + 3x GradientBoosting",
    "n_labels": 22,
    "labels": [
        "Administra\u00e7\u00e3o Geral, Finan\u00e7as e Recursos Humanos",
        "Ambiente",
        "Atividades Econ\u00f3micas",
        "A\u00e7\u00e3o Social",
        "Ci\u00eancia",
        "Comunica\u00e7\u00e3o e Rela\u00e7\u00f5es P\u00fablicas",
        "Coopera\u00e7\u00e3o Externa e Rela\u00e7\u00f5es Internacionais",
        "Cultura",
        "Desporto",
        "Educa\u00e7\u00e3o e Forma\u00e7\u00e3o Profissional",
        "Energia e Telecomunica\u00e7\u00f5es",
        "Habita\u00e7\u00e3o",
        "Obras Particulares",
        "Obras P\u00fablicas",
        "Ordenamento do Territ\u00f3rio",
        "Outros",
        "Patrim\u00f3nio",
        "Pol\u00edcia Municipal",
        "Prote\u00e7\u00e3o Animal",
        "Prote\u00e7\u00e3o Civil",
        "Sa\u00fade",
        "Tr\u00e2nsito, Transportes e Comunica\u00e7\u00f5es"
    ],
    "feature_dimensions": 10768,
    "tfidf_features": 10000,
    "bert_features": 768,
    "active_samples": 100,
    "n_gb_models": 3,
    "innovations": [
        "Multiple GradientBoosting with different hyperparameters",
        "Active learning uncertainty sampling simulation",
        "Adaptive ensemble weighting by label frequency",
        "Dynamic threshold optimization per label",
        "Dense feature matrix optimization"
    ],
    "performance": {
        "accuracy": 0.45179584120982985,
        "f1_macro": 0.5485386636825255,
        "f1_micro": 0.7362637362637363,
        "hamming_loss": 0.04124420003437017,
        "average_precision_macro": 0.6063436422306382
    }
}