cra-window-rules / modules /search_columns.py
Mark Febrizio
Documentation (#24)
fe4f734 unverified
raw
history blame
3.79 kB
import itertools
import re
from numpy import array
from pandas import DataFrame
class SearchError(Exception):
"""Search returned misaligned results."""
pass
# Defining a function to search for string patterns within dataframe columns
def search_columns(
df: DataFrame,
patterns: list,
columns: list,
return_as: str = "indicator_column",
return_column: str = "indicator",
re_flags = re.I | re.X
):
"""Search columns for string patterns within dataframe columns.
Args:
df (DataFrame): Input data in format of pandas dataframe.
patterns (list): List of string patterns to input, compatible with regex.
columns (list): List of column names to search for input patterns.
return_as (str, optional): Return a DataFrame with indicator column ("indicator_column") or filtered by the search terms ("filtered_df"). Defaults to "indicator_column".
re_flags (optional): Regex flags to use. Defaults to re.I | re.X.
Raises:
TypeError: Raises exception when `patterns` or `columns` parameters are not lists.
ValueError: Raises exception when `patterns` or `columns` parameters have incorrect length.
ValueError: Raises exception when `return_as` parameter receives an incorrect value.
Returns:
DataFrame: DataFrame with "indicator" column or filtered by search terms.
"""
# create list object for appending boolean arrays
bool_list = []
# ensure that input patterns and columns are formatted as lists
if not (isinstance(patterns, list) and isinstance(columns, list)):
raise TypeError('Inputs for "patterns" and "columns" keywords must be lists.')
if len(patterns) == len(columns):
# create list of inputs in format [(pattern1, column1),(pattern2, column2), ...]
inputs = list(zip(patterns,columns))
# loop over list of inputs
for i in inputs:
searchre = df[i[1]].str.contains(i[0], regex=True, case=False, flags=re_flags)
searchbool = array([True if n is True else False for n in searchre])
bool_list.append(searchbool)
elif (len(patterns) == 1) and (len(patterns) != len(columns)):
# create list of inputs in format [(pattern, column1),(pattern, column2), ...]
inputs = list(itertools.product(patterns, columns))
# loop over list of inputs
for i in inputs:
searchre = df[i[1]].str.contains(i[0], regex=True, case=False, flags=re_flags)
searchbool = array([True if n is True else False for n in searchre])
bool_list.append(searchbool)
else: # eg, patterns formatted as a list of len(n>1) but does not match len(columns)
raise ValueError("Length of inputs are incorrect. Lengths of 'patterns' and 'columns' can either match or a single pattern can map to multiple columns.")
# combine each "searchbool" array elementwise
# we want a positive match for any column to evaluate as True
# equivalent to (bool_list[0] | bool_list[1] | bool_list[2] | ... | bool_list[n-1])
filter_bool = array(bool_list).any(axis=0)
if return_as == "indicator_column":
dfResults = df.copy(deep=True)
dfResults.loc[:, return_column] = 0
dfResults.loc[filter_bool, return_column] = 1
#print(f"Count {return_column}: {sum(dfResults[return_column].values)}")
return dfResults
elif return_as == "filtered_df":
# filter results
dfResults = df.loc[filter_bool, :].copy(deep=True)
#print(f"Count {return_column}: {len(dfResults)}")
return dfResults
else:
raise ValueError("Incorrect input for 'return_as' parameter.")