feature_analysis.py 文件源码

def response_surface_analysis(df):
    """Perform response surface analysis on df."""

    def tally_results(df):
        features = [f for f in list(df.columns.values)
                    if f not in ['classifier', 'test_accuracy']]
        classifiers = ['Decision Tree', 'Linear SVC', 'Logistic Regression',
                       'Multinomial NB', 'Random Forest', 'Voting Classifier']
        for classifier in classifiers:
            from_class = df['classifier'] == classifier
            class_mean = df[from_class]['test_accuracy'].mean()
            for feature in features:
                with_feature = df[(df[feature] == 1) & from_class]
                wo_feature = df[(df[feature] == 0) & from_class]
                acc_diff = (with_feature['test_accuracy'].sum()
                            - wo_feature['test_accuracy'].sum())
                mean_diff = acc_diff / 255 + class_mean
                yield classifier, feature, mean_diff

    results = pd.DataFrame([res for res in tally_results(df)],
                           columns=['classifier', 'feature', 'effect'])
    class_order = list(df.groupby('classifier')['test_accuracy']
                         .mean()
                         .sort_values(ascending=False)
                         .index)
    results['classifier'] = pd.Categorical(results['classifier'],
                                           categories=class_order)
    feat_order = list(results.groupby('feature')['effect']
                             .mean()
                             .sort_values(ascending=False)
                             .index)
    results['feature'] = pd.Categorical(results['feature'],
                                        categories=feat_order)
    results.sort_values(['feature', 'classifier'], inplace=True)
    print(results)
    sns.stripplot('effect', 'feature', hue='classifier', data=results)
    sns.plt.legend()
    sns.plt.show()