Bayes搜索TV (BayesSearchTVSplit)

Java 类名：com.alibaba.alink.pipeline.tuning.BayesSearchTVSplit

Python 类名：BayesSearchTVSplit

功能介绍

BayesSearchTVSplit是通过构造参数的分布，对其中的每一组输入参数的组很分别进行训练，预测，评估。取得评估参数最优的模型，作为最终的返回模型

tv为训练验证，将数据按照比例切分为两份，对其中一份数据做训练，对剩余一份数据做预测和评估，得到一个评估结果。

此函数用tv方法得到每一个grid对应参数的评估结果，得到最优模型

参数说明

名称	中文名称	描述	类型	是否必须？	默认值
trainRatio	训练集比例	训练集与验证集的划分比例，取值范围为(0, 1]。	Double		0.8
ParamDist	参数分布	指定搜索的参数的分布	ParamDist	✓
Estimator	Estimator	用于调优的Estimator	Estimator	✓
TuningEvaluator	评估指标	用于选择最优模型的评估指标	TuningEvaluator	✓

代码示例

** 以下代码仅用于示意，可能需要修改部分代码或者配置环境后才能正常运行！**

Python 代码

def adult(url):
    data = (
        CsvSourceBatchOp()
        .setFilePath('https://alink-test-data.oss-cn-hangzhou.aliyuncs.com/adult_train.csv')
        .setSchemaStr(
            'age bigint, workclass string, fnlwgt bigint,'
            'education string, education_num bigint,'
            'marital_status string, occupation string,'
            'relationship string, race string, sex string,'
            'capital_gain bigint, capital_loss bigint,'
            'hours_per_week bigint, native_country string,'
            'label string'
        )
    )
    return data


def adult_train():
    return adult('https://alink-test-data.oss-cn-hangzhou.aliyuncs.com/adult_train.csv')


def adult_test():
    return adult('https://alink-test-data.oss-cn-hangzhou.aliyuncs.com/adult_test.csv')


def adult_numerical_feature_strs():
    return [
        "age", "fnlwgt", "education_num",
        "capital_gain", "capital_loss", "hours_per_week"
    ]


def adult_categorical_feature_strs():
    return [
        "workclass", "education", "marital_status",
        "occupation", "relationship", "race", "sex",
        "native_country"
    ]


def adult_features_strs():
    feature = adult_numerical_feature_strs()
    feature.extend(adult_categorical_feature_strs())

    return feature


def rf_grid_search_cv(featureCols, categoryFeatureCols, label, metric):
    rf = (
        RandomForestClassifier()
        .setFeatureCols(featureCols)
        .setCategoricalCols(categoryFeatureCols)
        .setLabelCol(label)
        .setPredictionCol('prediction')
        .setPredictionDetailCol('prediction_detail')
    )
    paramDist = (
        ParamDist()
        .addDist(rf, 'NUM_TREES', ValueDist.uniform(1., 10.))
    )
    tuningEvaluator = (
        BinaryClassificationTuningEvaluator()
        .setLabelCol(label)
        .setPredictionDetailCol("prediction_detail")
        .setTuningBinaryClassMetric(metric)
    )
    cv = (
        GridSearchCV()
        .setEstimator(rf)
        .setParamDist(paramDist)
        .setTuningEvaluator(tuningEvaluator)
        .setNumFolds(2)
    )

    return cv


def rf_grid_search_tv(featureCols, categoryFeatureCols, label, metric):
    rf = (
        RandomForestClassifier()
        .setFeatureCols(featureCols)
        .setCategoricalCols(categoryFeatureCols)
        .setLabelCol(label)
        .setPredictionCol('prediction')
        .setPredictionDetailCol('prediction_detail')
    )
    paramDist = (
        ParamDist()
        .addDist(rf, 'NUM_TREES', ValueDist.uniform(1., 10.))
    )
    tuningEvaluator = (
        BinaryClassificationTuningEvaluator()
        .setLabelCol(label)
        .setPredictionDetailCol("prediction_detail")
        .setTuningBinaryClassMetricå(metric)
    )
    cv = (
        GridSearchTVSplit()
        .setEstimator(rf)
        .setParamDist(paramDist)
        .setTuningEvaluator(tuningEvaluator)
    )

    return cv


def tuningcv(cv_estimator, input):
    return cv_estimator.fit(input)


def tuningtv(tv_estimator, input):
    return tv_estimator.fit(input)


def main():
    print('rf cv tuning')
    model = tuningcv(
        rf_grid_search_cv(adult_features_strs(),
                          adult_categorical_feature_strs(), 'label', 'AUC'),
        adult_train()
    )
    
    print(model.getReport())
    
    print('rf tv tuning')
    model = tuningtv(
        rf_grid_search_tv(adult_features_strs(),
                          adult_categorical_feature_strs(), 'label', 'AUC'),
        adult_train()
    )

    print(model.getReport())
main()

运行结果