再探askl元学习

按照论文的说法，landmarking方法因为过于耗时所以不算在元特征中

EXCLUDE_META_FEATURES_CLASSIFICATION
Out[5]: 
{
     'Landmark1NN',
 'LandmarkDecisionNodeLearner',
 'LandmarkDecisionTree',
 'LandmarkLDA',
 'LandmarkNaiveBayes',
 'PCA',
 'PCAFractionOfComponentsFor95PercentVariance',
 'PCAKurtosisFirstPC',
 'PCASkewnessFirstPC'}
EXCLUDE_META_FEATURES_REGRESSION
Out[6]: 
{
     'ClassEntropy',
 'ClassOccurences',
 'ClassProbabilityMax',
 'ClassProbabilityMean',
 'ClassProbabilityMin',
 'ClassProbabilitySTD',
 'Landmark1NN',
 'LandmarkDecisionNodeLearner',
 'LandmarkDecisionTree',
 'LandmarkLDA',
 'LandmarkNaiveBayes',
 'LandmarkRandomNodeLearner',
 'NumberOfClasses',
 'PCA',
 'PCAFractionOfComponentsFor95PercentVariance',
 'PCAKurtosisFirstPC',
 'PCASkewnessFirstPC'}

回归任务因为没有离散标签，所以与class相关的元特征也排除了

'ClassEntropy',
 'ClassOccurences',
 'ClassProbabilityMax',
 'ClassProbabilityMean',
 'ClassProbabilityMin',
 'ClassProbabilitySTD',

    calculate.update(npy_metafeatures)

calculate_all_metafeatures_with_labels 中，在npy_metafeatures的基础上更新不用的元特征。

只有npy_metafeatures需要用DataPreprocessor做transform

LandmarkRandomNodeLearner，

$ASKL$计算数据集元特征的方法其实就放在smbo.py里面，一共有两个函数，一个函数用在计算$X,y$的时候不需要做encode，只获取General的元特征。另一个需要做encode。

# metalearning helpers
def _calculate_metafeatures(data_feat_type, data_info_task, basename,
                            x_train, y_train, watcher, logger):
    # == Calculate metafeatures
    task_name = 'CalculateMetafeatures'
    watcher.start_task(task_name)
    categorical = [True if feat_type.lower() in ['categorical'] else False
                   for feat_type in data_feat_type]

    EXCLUDE_META_FEATURES = EXCLUDE_META_FEATURES_CLASSIFICATION \
        if data_info_task in CLASSIFICATION_TASKS else EXCLUDE_META_FEATURES_REGRESSION

    if data_info_task in [MULTICLASS_CLASSIFICATION, BINARY_CLASSIFICATION,
                          MULTILABEL_CLASSIFICATION, REGRESSION,
                          MULTIOUTPUT_REGRESSION]:
        logger.info('Start calculating metafeatures for %s', basename)
        result = calculate_all_metafeatures_with_labels(
            x_train, y_train, categorical=categorical,
            dataset_name=basename,
            dont_calculate=EXCLUDE_META_FEATURES, )
        for key in list(result.metafeature_values.keys()):
            if result.metafeature_values[key].type_ != 'METAFEATURE':
                del result.metafeature_values[key]

    else:
        result = None
        logger.info('Metafeatures not calculated')
    watcher.stop_task(task_name)
    logger.info(
        'Calculating Metafeatures (categorical attributes) took %5.2f',
        watcher.wall_elapsed(task_name))
    return result


def _calculate_metafeatures_encoded(basename, x_train, y_train, watcher,
                                    task, logger):
    EXCLUDE_META_FEATURES = EXCLUDE_META_FEATURES_CLASSIFICATION \
        if task in CLASSIFICATION_TASKS else EXCLUDE_META_FEATURES_REGRESSION

    task_name = 'CalculateMetafeaturesEncoded'
    watcher.start_task(task_name)
    result = calculate_all_metafeatures_encoded_labels(
        x_train, y_train, categorical=[False] * x_train.shape[1],
        dataset_name=basename, dont_calculate=EXCLUDE_META_FEATURES)
    for key in list(result.metafeature_values.keys()):
        if result.metafeature_values[key].type_ != 'METAFEATURE':
            del result.metafeature_values[key]
    watcher.stop_task(task_name)
    logger.info(
        'Calculating Metafeatures (encoded attributes) took %5.2fsec',
        watcher.wall_elapsed(task_name))
    return result