oof_means.py

def get_target_mean(df_train, df_test, var, target, NFOLDS, NOISE):
    """ 
    Creates out of fold averages by the categorical variable passed.
    Decreasing the number of folds and increasing the noise can help 
    prevent over fitting the training set.
    """
    
    df_train['mean_{}'.format(var)] = np.nan
    df_test['mean_{}'.format(var)] = np.nan
    
    values = np.empty((NFOLD))
    values[:] = np.nan
    mean_stats = dict((el,np.copy(values)) for el in df_train[var].unique()) 

    kf = KFold(NFOLD, shuffle=True)
    for i, (train_index, test_index) in enumerate(kf.split(df_train)):

        cv_train = df_train.iloc[train_index]
        cv_test = df_train.iloc[test_index]
        grouped = cv_train.groupby(var)[[target]].mean().reset_index()
        
        for v in grouped[var].unique():
            var_index = cv_test[cv_test[var] == v].index
            impute_mean = grouped[grouped[var] == v][target].mean()
            df_train.loc[var_index, 'mean_{}'.format(var)] = impute_mean * (np.random.uniform(0,0.01)+1)
            mean_stats[v][i] = the_mean
            
    avgDict = {}
    for k,v in mean_stats.items():
        avgDict[k] = sum(v)/ float(len(v))
        
    for v in avgDict:
        var_index = df_test[df_test[var] == v].index
        df_test.loc[var_index, 'mean_{}'.format(var)] = avgDict[v]
        
    df_train.drop(var, axis=1, inplace=True)
    df_test.drop(var, axis=1, inplace=True)
        
    return df_train, df_test