bindings/python-compute/feature

import pandas as pd import numpy as np from scipy.stats.kde import gaussian_kde from scipy.stats import entropy from numpy.linalg import LinAlgError from .data_manager import DUMMY_PREFIX_SEP from .constants import RANGE_FILTER CLUSTER_COL = 'clusters' GROUP_ID_COL = 'clusterGroupId' NUMERICAL_DOMAIN_INTERVAL = 100 # index of nearest element in array to value nearest_index = lambda arr, val: (np.abs(arr - val)).argmin() # get a distribution where all densities are 0 except for the position of unique value (=1) def get_single_value_distribution(domain, raw_values): unique_value = raw_values.unique()[0] distribution = np.zeros(domain.shape) distribution[nearest_index(domain, unique_value)] = 1. return distribution class FeatureDifferentiation(object): def __init__(self, feature_df, categorical_features=None): self.feature_df = feature_df.copy() self.categorical_features = categorical_features self.cat_dict = self.compute_categorical_features_dict() self.features_meta_data = self.compute_features_meta_data() def set_params(self, segment_group_0, segment_group_1, segment_ids): self.target = pd.Series([0 if i in segment_group_0 else (1 if i in segment_group_1 else 2) for i in segment_ids]) def compute_categorical_features_dict(self): if self.categorical_features is None: # if data type is non-number or has small number of unique values self.categorical_features = [c for c in self.feature_df.columns if len(self.feature_df[c].unique()) < 7 or self.feature_df.dtypes[c] == 'object'] cat_dict = {} # create a dict of list, dict fields are categorical parent features, # and list elements are parent feature names suffixed with category names for c in self.categorical_features: # in case of features are already one-hot encoded c_split = c.split(DUMMY_PREFIX_SEP) try: cat_dict[c_split[0]] except: cat_dict[c_split[0]] = [] cat_dict[c_split[0]] += [c] return cat_dict # todo: merge histogram computation in compute_features_meta_data, compute_split_cat_count, compute_split_kde def compute_features_meta_data(self): features_list = [] for feature_name in self.feature_df.columns: col = self.feature_df[feature_name] if feature_name in self.cat_dict: type = 'categorical' value_counts = col.value_counts(dropna=False) distribution = np.stack([value_counts.index.values, value_counts.values]).tolist() else: x = np.linspace(np.min(col), np.max(col), num=NUMERICAL_DOMAIN_INTERVAL) try: kde = gaussian_kde(col, bw_method=0.1) except LinAlgError: kde = lambda x: get_single_value_distribution(x, col) distribution = np.stack([x, 10000*kde(x)]).tolist() if len(distribution[0]) <= 1 or \ len(distribution[0]) > max(len(col) / 10., 100): continue feature_dict = { 'name': feature_name, 'type': 'categorical' if feature_name in self.cat_dict else 'numerical', 'distribution': distribution } features_list.append(feature_dict) return features_list def compute_split_cat_count(self, col, exclude_outlier=False): if exclude_outlier: cc = col.value_counts() mask = col.isin(cc.index[cc < len(col) * .005]) col.loc[mask] = 'OTHER_CATEGORY' cc0 = col[self.target == 0].value_counts(normalize=True, dropna=False) cc1 = col[self.target == 1].value_counts(normalize=True, dropna=False) count_df = pd.concat([cc0, cc1], axis=1).fillna(0) count_df.columns = [0, 1] count_df.index = count_df.index.fillna('NO_CATEGORY') buffer = 1. / float(len(self.target)) count_df['ratio'] = count_df.apply(lambda row: (row[0] + buffer) / (row[1] + buffer), axis=1) count_df.sort_values('ratio', inplace=True) return np.stack((count_df.index.values, 10000*count_df[0].values, 10000*count_df[1].values)) def compute_split_kde(self, col, exclude_outlier=True): try: col_value_range = RANGE_FILTER[col.name] except: if exclude_outlier: col_value_range = [np.percentile(col, 1), np.percentile(col, 99)] else: col_value_range = [np.min(col), np.max(col)] x = np.linspace(col_value_range[0], col_value_range[1], num=NUMERICAL_DOMAIN_INTERVAL) # could also use pd.Series.value_counts, (use bins) # if unique values in array <=1, kde will have exception try: kde0 = gaussian_kde(col[self.target == 0], bw_method=0.1) except LinAlgError: kde0 = lambda x: get_single_value_distribution(x, col[self.target == 0]) try: kde1 = gaussian_kde(col[self.target == 1], bw_method=0.1) except LinAlgError: kde1 = lambda x: get_single_value_distribution(x, col[self.target == 1]) return np.stack((x, 10000*kde0(x), 10000*kde1(x))) def get_features_meta_data(self): return self.features_meta_data def get_feature_distribution_by_segment_group(self, feature_name): distribution_dict = {} main_name = feature_name.split(DUMMY_PREFIX_SEP)[0] distribution_dict['name'] = main_name if main_name in self.cat_dict: distribution_dict['distribution'] = self.compute_split_cat_count(self.feature_df[main_name]).tolist() distribution_dict['type'] = 'categorical' else: distribution_dict['distribution'] = self.compute_split_kde(self.feature_df[main_name]).tolist() distribution_dict['type'] = 'numerical' return distribution_dict def get_features_distribution_by_segment_group(self): distribution_list = [] buffer = 1. / float(len(self.target)) for feature_name in self.feature_df.columns: distribution_dict = self.get_feature_distribution_by_segment_group(feature_name) # ignore features with too may categories, like uuid; ignore features with only one category if len(distribution_dict['distribution'][0]) <= 1 or \ len(distribution_dict['distribution'][0]) > max(len(self.target) / 10., 100): continue distribution_dict['divergence'] = entropy( np.array(distribution_dict['distribution'][1]) + buffer, np.array(distribution_dict['distribution'][2]) + buffer ) distribution_list.append(distribution_dict) return sorted(distribution_list, key=lambda x: x['divergence'], reverse=True)

bindings/python-compute/feature_differentiation.py (122 lines of code) (raw):