In this article, we will cover various methods to filter pandas dataframe in Python. Data Filtering is one of the most frequent data manipulation operation. It is similar to WHERE clause in SQL or you must have used filter in MS Excel for selecting specific rows based on some conditions. In terms of speed, python has an efficient way to perform filtering and aggregation. It has an excellent package called pandas for data wrangling tasks. Pandas has been built on top of numpy package which was written in C language which is a low level language. Hence data manipulation using pandas package is fast and smart way to handle big sized datasets.
Examples of Data Filtering
It is one of the most initial step of data preparation for predictive modeling or any reporting project. It is also called 'Subsetting Data'. See some of the examples of data filtering below.
Import Data
Make sure pandas package is already installed before submitting the following code. You can check it by running !pip show pandas
statement in Ipython console. If it is not installed, you can install it by using the command !pip install pandas
.
We are going to use dataset containing details of flights departing from NYC in 2013. This dataset has 32735 rows and 16 columns. See column names below. To import dataset, we are using read_csv( )
function from pandas package.
['year', 'month', 'day', 'dep_time', 'dep_delay', 'arr_time', 'arr_delay', 'carrier', 'tailnum', 'flight', 'origin', 'dest', 'air_time', 'distance', 'hour', 'minute']
import pandas as pd df = pd.read_csv("https://dyurovsky.github.io/psyc201/data/lab2/nycflights.csv")
Select flights details of JetBlue Airways that has 2 letters carrier code B6
with origin from JFK
airport
Method 1 : DataFrame Way
newdf = df[(df.origin == "JFK") & (df.carrier == "B6")]
newdf.head() Out[23]: year month day dep_time ... air_time distance hour minute 7 2013 8 13 1920 ... 48.0 228.0 19.0 20.0 10 2013 6 17 940 ... 50.0 264.0 9.0 40.0 14 2013 10 21 1217 ... 46.0 266.0 12.0 17.0 23 2013 7 7 2310 ... 223.0 1626.0 23.0 10.0 35 2013 4 12 840 ... 186.0 1598.0 8.0 40.0 [5 rows x 16 columns]
newdf
.&
refers to AND
condition which means meeting both the criteria.(df.origin == "JFK") & (df.carrier == "B6")
returns True / False. True where condition matches and False where the condition does not hold. Later it is passed within df and returns all the rows corresponding to True. It returns 4166 rows.Method 2 : Query Function
In pandas package, there are multiple ways to perform filtering. The above code can also be written like the code shown below. This method is elegant and more readable and you don't need to mention dataframe name everytime when you specify columns (variables).
newdf = df.query('origin == "JFK" & carrier == "B6"')
Method 3 : loc function
loc is an abbreviation of location term. All these 3 methods return same output. It's just a different ways of doing filtering rows.
newdf = df.loc[(df.origin == "JFK") & (df.carrier == "B6")]
Suppose you want to select specific rows by their position (let's say from second through fifth row). We can use df.iloc[ ]
function for the same.
Indexing in python starts from zero. df.iloc[0:5,] refers to first to fifth row (excluding end point 6th row here). df.iloc[0:5,] is equivalent to df.iloc[:5,]
df.iloc[:5,] #First 5 rows df.iloc[1:5,] #Second to Fifth row df.iloc[5,0] #Sixth row and 1st column df.iloc[1:5,0] #Second to Fifth row, first column df.iloc[1:5,:5] #Second to Fifth row, first 5 columns df.iloc[2:7,1:3] #Third to Seventh row, 2nd and 3rd column
Difference between loc and iloc function
loc considers rows based on index labels. Whereas iloc considers rows based on position in the index so it only takes integers. Let's create a sample data for illustration
import numpy as np x = pd.DataFrame({"col1" : np.arange(1,20,2)}, index=[9,8,7,6,0, 1, 2, 3, 4, 5])
col1 9 1 8 3 7 5 6 7 0 9 1 11 2 13 3 15 4 17 5 19
iloc - Index Position
x.iloc[0:5] Output col1 9 1 8 3 7 5 6 7 0 9
Selecting rows based on index or row positionloc - Index Label
x.loc[0:5] Output col1 0 9 1 11 2 13 3 15 4 17 5 19
Selecting rows based on labels of index
How x.loc[0:5]
returns 6 rows (inclusive of 5 which is 6th element)?
It is because loc
does not produce output based on index position. It considers labels of index only which can be alphabet as well and includes both starting and end point. Refer the example below.
x = pd.DataFrame({"col1" : range(1,5)}, index=['a','b','c','d']) x.loc['a':'c'] # equivalent to x.iloc[0:3] col1 a 1 b 2 c 3
Here we are selecting first five rows of two columns named origin and dest.
df.loc[df.index[0:5],["origin","dest"]]
df.index
returns index labels. df.index[0:5] is required instead of 0:5 (without df.index) because index labels do not always in sequence and start from 0. It can start from any number or even can have alphabet letters. Refer the example where we showed comparison of iloc and loc.
Suppose you want to include all the flight details where origin is either JFK or LGA.
# Long Way newdf = df.loc[(df.origin == "JFK") | (df.origin == "LGA")] # Smart Way newdf = df[df.origin.isin(["JFK", "LGA"])]
|
implies OR condition which means any of the conditions holds True. isin( )
is similar to IN operator in SAS and R which can take many values and apply OR condition. Make sure you specify values in list [ ].
In this example, we are deleting all the flight details where origin is from JFK. !=
implies NOT EQUAL TO.
newdf = df.loc[(df.origin != "JFK") & (df.carrier == "B6")]
Let's check whether the above line of code works fine or not by looking at unique values of column origin in newdf.
pd.unique(newdf.origin) ['LGA', 'EWR']
Tilde ~
is used to negate the condition. It is equivalent to NOT operator in SAS and R.
newdf = df[~((df.origin == "JFK") & (df.carrier == "B6"))]
With the use of notnull()
function, you can exclude or remove NA and NAN values. In the example below, we are removing missing values from origin column. Since this dataframe does not contain any blank values, you would find same number of rows in newdf.
newdf = df[df.origin.notnull()]
It is generally considered tricky to handle text data. But python makes it easier when it comes to dealing character or string columns. Let's prepare a fake data for example.
import pandas as pd df = pd.DataFrame({"var1": ["AA_2", "B_1", "C_2", "A_2"]}) var1 0 AA_2 1 B_1 2 C_2 3 A_2
Select rows having values starting from letter 'A'
By using .str
, you can enable string functions and can apply on pandas dataframe. str[0] means first letter.
df[df['var1'].str[0] == 'A']
Filter rows having string length greater than 3
len( )
function calculates length of iterable.
df[df['var1'].str.len()>3]
Select string containing letters A or B
contains( )
function is similar to LIKE statement in SQL and SAS. You can subset data by mentioning pattern in contains( ) function.
df[df['var1'].str.contains('A|B')] Output var1 0 AA_2 1 B_1 3 A_2
You can perform filtering using pure python methods without dependency on pandas package.
Warning : Methods shown below for filtering are not efficient ones. The main objective of showing the following methods is to show how to do subsetting without using pandas package. In your live project, you should use pandas' builtin functions (query( ), loc[ ], iloc[ ]) which are explained above.
We don't need to create a dataframe to store data. We can stock it in list data structure. lst_df
contains flights data which were imported from CSV file.
import csv import requests response = requests.get('https://dyurovsky.github.io/psyc201/data/lab2/nycflights.csv').text lines = response.splitlines() d = csv.DictReader(lines) lst_df = list(d)
Lambda Method for Filtering
Lambda is an alternative way of defining user defined function. With the use of lambda, you can define function in a single line of code. You can check out this link to learn more about it.
l1 = list(filter(lambda x: x["origin"] == 'JFK' and x["carrier"] == 'B6', lst_df))
If you are wondering how to use this lambda function on a dataframe, you can submit the code below.
newdf = df[df.apply(lambda x: x["origin"] == 'JFK' and x["carrier"] == 'B6', axis=1)]
List Comprehension Method for Filtering
List comprehension is an alternative to lambda function and makes code more readable. Detailed Tutorial : List Comprehension
l2 = list(x for x in lst_df if x["origin"] == 'JFK' and x["carrier"] == 'B6')
You can use list comprehension on dataframe like the way shown below.
newdf = df.iloc[[index for index,row in df.iterrows() if row['origin'] == 'JFK' and row['carrier'] == 'B6']]
Create Class for Filtering
Python is an object-oriented programming language in which code is implemented using class
.
class filter: def __init__(self, l, query): self.output = [] for data in l: if eval(query): self.output.append(data) l3 = filter(lst_df, 'data["origin"] == "JFK" and data["carrier"] == "B6"').output