I have two DataFrames which I want to merge based on a column. However, due to alternate spellings, different number of spaces, absence/presence of diacritical marks, I would like to be able to merge as long as they are similar to one another.
Any similarity algorithm will do (soundex, Levenshtein, difflib's).
Say one DataFrame has the following data:
df1 = DataFrame([[1],[2],[3],[4],[5]], index=['one','two','three','four','five'], columns=['number'])
number
one 1
two 2
three 3
four 4
five 5
df2 = DataFrame([['a'],['b'],['c'],['d'],['e']], index=['one','too','three','fours','five'], columns=['letter'])
letter
one a
too b
three c
fours d
five e
Then I want to get the resulting DataFrame
number letter
one 1 a
two 2 b
three 3 c
four 4 d
five 5 e
Similar to #locojay suggestion, you can apply difflib's get_close_matches to df2's index and then apply a join:
In [23]: import difflib
In [24]: difflib.get_close_matches
Out[24]: <function difflib.get_close_matches>
In [25]: df2.index = df2.index.map(lambda x: difflib.get_close_matches(x, df1.index)[0])
In [26]: df2
Out[26]:
letter
one a
two b
three c
four d
five e
In [31]: df1.join(df2)
Out[31]:
number letter
one 1 a
two 2 b
three 3 c
four 4 d
five 5 e
.
If these were columns, in the same vein you could apply to the column then merge:
df1 = DataFrame([[1,'one'],[2,'two'],[3,'three'],[4,'four'],[5,'five']], columns=['number', 'name'])
df2 = DataFrame([['a','one'],['b','too'],['c','three'],['d','fours'],['e','five']], columns=['letter', 'name'])
df2['name'] = df2['name'].apply(lambda x: difflib.get_close_matches(x, df1['name'])[0])
df1.merge(df2)
Using fuzzywuzzy
Since there are no examples with the fuzzywuzzy package, here's a function I wrote which will return all matches based on a threshold you can set as a user:
Example datframe
df1 = pd.DataFrame({'Key':['Apple', 'Banana', 'Orange', 'Strawberry']})
df2 = pd.DataFrame({'Key':['Aple', 'Mango', 'Orag', 'Straw', 'Bannanna', 'Berry']})
# df1
Key
0 Apple
1 Banana
2 Orange
3 Strawberry
# df2
Key
0 Aple
1 Mango
2 Orag
3 Straw
4 Bannanna
5 Berry
Function for fuzzy matching
def fuzzy_merge(df_1, df_2, key1, key2, threshold=90, limit=2):
"""
:param df_1: the left table to join
:param df_2: the right table to join
:param key1: key column of the left table
:param key2: key column of the right table
:param threshold: how close the matches should be to return a match, based on Levenshtein distance
:param limit: the amount of matches that will get returned, these are sorted high to low
:return: dataframe with boths keys and matches
"""
s = df_2[key2].tolist()
m = df_1[key1].apply(lambda x: process.extract(x, s, limit=limit))
df_1['matches'] = m
m2 = df_1['matches'].apply(lambda x: ', '.join([i[0] for i in x if i[1] >= threshold]))
df_1['matches'] = m2
return df_1
Using our function on the dataframes: #1
from fuzzywuzzy import fuzz
from fuzzywuzzy import process
fuzzy_merge(df1, df2, 'Key', 'Key', threshold=80)
Key matches
0 Apple Aple
1 Banana Bannanna
2 Orange Orag
3 Strawberry Straw, Berry
Using our function on the dataframes: #2
df1 = pd.DataFrame({'Col1':['Microsoft', 'Google', 'Amazon', 'IBM']})
df2 = pd.DataFrame({'Col2':['Mcrsoft', 'gogle', 'Amason', 'BIM']})
fuzzy_merge(df1, df2, 'Col1', 'Col2', 80)
Col1 matches
0 Microsoft Mcrsoft
1 Google gogle
2 Amazon Amason
3 IBM
Installation:
Pip
pip install fuzzywuzzy
Anaconda
conda install -c conda-forge fuzzywuzzy
I have written a Python package which aims to solve this problem:
pip install fuzzymatcher
You can find the repo here and docs here.
Basic usage:
Given two dataframes df_left and df_right, which you want to fuzzy join, you can write the following:
from fuzzymatcher import link_table, fuzzy_left_join
# Columns to match on from df_left
left_on = ["fname", "mname", "lname", "dob"]
# Columns to match on from df_right
right_on = ["name", "middlename", "surname", "date"]
# The link table potentially contains several matches for each record
fuzzymatcher.link_table(df_left, df_right, left_on, right_on)
Or if you just want to link on the closest match:
fuzzymatcher.fuzzy_left_join(df_left, df_right, left_on, right_on)
I would use Jaro-Winkler, because it is one of the most performant and accurate approximate string matching algorithms currently available [Cohen, et al.], [Winkler].
This is how I would do it with Jaro-Winkler from the jellyfish package:
def get_closest_match(x, list_strings):
best_match = None
highest_jw = 0
for current_string in list_strings:
current_score = jellyfish.jaro_winkler(x, current_string)
if(current_score > highest_jw):
highest_jw = current_score
best_match = current_string
return best_match
df1 = pandas.DataFrame([[1],[2],[3],[4],[5]], index=['one','two','three','four','five'], columns=['number'])
df2 = pandas.DataFrame([['a'],['b'],['c'],['d'],['e']], index=['one','too','three','fours','five'], columns=['letter'])
df2.index = df2.index.map(lambda x: get_closest_match(x, df1.index))
df1.join(df2)
Output:
number letter
one 1 a
two 2 b
three 3 c
four 4 d
five 5 e
For a general approach: fuzzy_merge
For a more general scenario in which we want to merge columns from two dataframes which contain slightly different strings, the following function uses difflib.get_close_matches along with merge in order to mimic the functionality of pandas' merge but with fuzzy matching:
import difflib
def fuzzy_merge(df1, df2, left_on, right_on, how='inner', cutoff=0.6):
df_other= df2.copy()
df_other[left_on] = [get_closest_match(x, df1[left_on], cutoff)
for x in df_other[right_on]]
return df1.merge(df_other, on=left_on, how=how)
def get_closest_match(x, other, cutoff):
matches = difflib.get_close_matches(x, other, cutoff=cutoff)
return matches[0] if matches else None
Here are some use cases with two sample dataframes:
print(df1)
key number
0 one 1
1 two 2
2 three 3
3 four 4
4 five 5
print(df2)
key_close letter
0 three c
1 one a
2 too b
3 fours d
4 a very different string e
With the above example, we'd get:
fuzzy_merge(df1, df2, left_on='key', right_on='key_close')
key number key_close letter
0 one 1 one a
1 two 2 too b
2 three 3 three c
3 four 4 fours d
And we could do a left join with:
fuzzy_merge(df1, df2, left_on='key', right_on='key_close', how='left')
key number key_close letter
0 one 1 one a
1 two 2 too b
2 three 3 three c
3 four 4 fours d
4 five 5 NaN NaN
For a right join, we'd have all non-matching keys in the left dataframe to None:
fuzzy_merge(df1, df2, left_on='key', right_on='key_close', how='right')
key number key_close letter
0 one 1.0 one a
1 two 2.0 too b
2 three 3.0 three c
3 four 4.0 fours d
4 None NaN a very different string e
Also note that difflib.get_close_matches will return an empty list if no item is matched within the cutoff. In the shared example, if we change the last index in df2 to say:
print(df2)
letter
one a
too b
three c
fours d
a very different string e
We'd get an index out of range error:
df2.index.map(lambda x: difflib.get_close_matches(x, df1.index)[0])
IndexError: list index out of range
In order to solve this the above function get_closest_match will return the closest match by indexing the list returned by difflib.get_close_matches only if it actually contains any matches.
http://pandas.pydata.org/pandas-docs/dev/merging.html does not have a hook function to do this on the fly. Would be nice though...
I would just do a separate step and use difflib getclosest_matches to create a new column in one of the 2 dataframes and the merge/join on the fuzzy matched column
I used Fuzzymatcher package and this worked well for me. Visit this link for more details on this.
use the below command to install
pip install fuzzymatcher
Below is the sample Code (already submitted by RobinL above)
from fuzzymatcher import link_table, fuzzy_left_join
# Columns to match on from df_left
left_on = ["fname", "mname", "lname", "dob"]
# Columns to match on from df_right
right_on = ["name", "middlename", "surname", "date"]
# The link table potentially contains several matches for each record
fuzzymatcher.link_table(df_left, df_right, left_on, right_on)
Errors you may get
ZeroDivisionError: float division by zero---> Refer to this
link to resolve it
OperationalError: No Such Module:fts4 --> downlaod the sqlite3.dll
from here and replace the DLL file in your python or anaconda
DLLs folder.
Pros :
Works faster. In my case, I compared one dataframe with 3000 rows with anohter dataframe with 170,000 records . This also uses SQLite3 search across text. So faster than many
Can check across multiple columns and 2 dataframes. In my case, I was looking for closest match based on address and company name. Sometimes, company name might be same but address is the good thing to check too.
Gives you score for all the closest matches for the same record. you choose whats the cutoff score.
cons:
Original package installation is buggy
Required C++ and visual studios installed too
Wont work for 64 bit anaconda/Python
There is a package called fuzzy_pandas that can use levenshtein, jaro, metaphone and bilenco methods. With some great examples here
import pandas as pd
import fuzzy_pandas as fpd
df1 = pd.DataFrame({'Key':['Apple', 'Banana', 'Orange', 'Strawberry']})
df2 = pd.DataFrame({'Key':['Aple', 'Mango', 'Orag', 'Straw', 'Bannanna', 'Berry']})
results = fpd.fuzzy_merge(df1, df2,
left_on='Key',
right_on='Key',
method='levenshtein',
threshold=0.6)
results.head()
Key Key
0 Apple Aple
1 Banana Bannanna
2 Orange Orag
As a heads up, this basically works, except if no match is found, or if you have NaNs in either column. Instead of directly applying get_close_matches, I found it easier to apply the following function. The choice of NaN replacements will depend a lot on your dataset.
def fuzzy_match(a, b):
left = '1' if pd.isnull(a) else a
right = b.fillna('2')
out = difflib.get_close_matches(left, right)
return out[0] if out else np.NaN
You can use d6tjoin for that
import d6tjoin.top1
d6tjoin.top1.MergeTop1(df1.reset_index(),df2.reset_index(),
fuzzy_left_on=['index'],fuzzy_right_on=['index']).merge()['merged']
index number index_right letter
0 one 1 one a
1 two 2 too b
2 three 3 three c
3 four 4 fours d
4 five 5 five e
It has a variety of additional features such as:
check join quality, pre and post join
customize similarity function, eg edit distance vs hamming distance
specify max distance
multi-core compute
For details see
MergeTop1 examples - Best match join examples notebook
PreJoin examples - Examples for diagnosing join problems
I have used fuzzywuzz in a very minimal way whilst matching the existing behaviour and keywords of merge in pandas.
Just specify your accepted threshold for matching (between 0 and 100):
from fuzzywuzzy import process
def fuzzy_merge(df, df2, on=None, left_on=None, right_on=None, how='inner', threshold=80):
def fuzzy_apply(x, df, column, threshold=threshold):
if type(x)!=str:
return None
match, score, *_ = process.extract(x, df[column], limit=1)[0]
if score >= threshold:
return match
else:
return None
if on is not None:
left_on = on
right_on = on
# create temp column as the best fuzzy match (or None!)
df2['tmp'] = df2[right_on].apply(
fuzzy_apply,
df=df,
column=left_on,
threshold=threshold
)
merged_df = df.merge(df2, how=how, left_on=left_on, right_on='tmp')
del merged_df['tmp']
return merged_df
Try it out using the example data:
df1 = pd.DataFrame({'Key':['Apple', 'Banana', 'Orange', 'Strawberry']})
df2 = pd.DataFrame({'Key':['Aple', 'Mango', 'Orag', 'Straw', 'Bannanna', 'Berry']})
fuzzy_merge(df, df2, on='Key', threshold=80)
Using thefuzz
Using SeatGeek's great package thefuzz, which makes use of Levenshtein distance. This works with data held in columns. It adds matches as rows rather than columns, to preserve a tidy dataset, and allows additional columns to be easily pulled through to the output dataframe.
Sample data
df1 = pd.DataFrame({'col_a':['one','two','three','four','five'], 'col_b':[1, 2, 3, 4, 5]})
col_a col_b
0 one 1
1 two 2
2 three 3
3 four 4
4 five 5
df2 = pd.DataFrame({'col_a':['one','too','three','fours','five'], 'col_b':['a','b','c','d','e']})
col_a col_b
0 one a
1 too b
2 three c
3 fours d
4 five e
Function used to do the matching
def fuzzy_match(
df_left, df_right, column_left, column_right, threshold=90, limit=1
):
# Create a series
series_matches = df_left[column_left].apply(
lambda x: process.extract(x, df_right[column_right], limit=limit) # Creates a series with id from df_left and column name _column_left_, with _limit_ matches per item
)
# Convert matches to a tidy dataframe
df_matches = series_matches.to_frame()
df_matches = df_matches.explode(column_left) # Convert list of matches to rows
df_matches[
['match_string', 'match_score', 'df_right_id']
] = pd.DataFrame(df_matches[column_left].tolist(), index=df_matches.index) # Convert match tuple to columns
df_matches.drop(column_left, axis=1, inplace=True) # Drop column of match tuples
# Reset index, as in creating a tidy dataframe we've introduced multiple rows per id, so that no longer functions well as the index
if df_matches.index.name:
index_name = df_matches.index.name # Stash index name
else:
index_name = 'index' # Default used by pandas
df_matches.reset_index(inplace=True)
df_matches.rename(columns={index_name: 'df_left_id'}, inplace=True) # The previous index has now become a column: rename for ease of reference
# Drop matches below threshold
df_matches.drop(
df_matches.loc[df_matches['match_score'] < threshold].index,
inplace=True
)
return df_matches
Use function and merge data
import pandas as pd
from thefuzz import process
df_matches = fuzzy_match(
df1,
df2,
'col_a',
'col_a',
threshold=60,
limit=1
)
df_output = df1.merge(
df_matches,
how='left',
left_index=True,
right_on='df_left_id'
).merge(
df2,
how='left',
left_on='df_right_id',
right_index=True,
suffixes=['_df1', '_df2']
)
df_output.set_index('df_left_id', inplace=True) # For some reason the first merge operation wrecks the dataframe's index. Recreated from the value we have in the matches lookup table
df_output = df_output[['col_a_df1', 'col_b_df1', 'col_b_df2']] # Drop columns used in the matching
df_output.index.name = 'id'
id col_a_df1 col_b_df1 col_b_df2
0 one 1 a
1 two 2 b
2 three 3 c
3 four 4 d
4 five 5 e
Tip: Fuzzy matching using thefuzz is much quicker if you optionally install the python-Levenshtein package too.
For more complex use cases to match rows with many columns you can use recordlinkage package. recordlinkage provides all the tools to fuzzy match rows between pandas data frames which helps to deduplicate your data when merging. I have written a detailed article about the package here
if the join axis is numeric this could also be used to match indexes with a specified tolerance:
def fuzzy_left_join(df1, df2, tol=None):
index1 = df1.index.values
index2 = df2.index.values
diff = np.abs(index1.reshape((-1, 1)) - index2)
mask_j = np.argmin(diff, axis=1) # min. of each column
mask_i = np.arange(mask_j.shape[0])
df1_ = df1.iloc[mask_i]
df2_ = df2.iloc[mask_j]
if tol is not None:
mask = np.abs(df2_.index.values - df1_.index.values) <= tol
df1_ = df1_.loc[mask]
df2_ = df2_.loc[mask]
df2_.index = df1_.index
out = pd.concat([df1_, df2_], axis=1)
return out
TheFuzz is the new version of a fuzzywuzzy
In order to fuzzy-join string-elements in two big tables you can do this:
Use apply to go row by row
Use swifter to parallel, speed up and visualize default apply function (with colored progress bar)
Use OrderedDict from collections to get rid of duplicates in the output of merge and keep the initial order
Increase limit in thefuzz.process.extract to see more options for merge (stored in a list of tuples with % of similarity)
'*' You can use thefuzz.process.extractOne instead of thefuzz.process.extract to return just one best-matched item (without specifying any limit). However, be aware that several results could have same % of similarity and you will get only one of them.
'**' Somehow the swifter takes a minute or two before starting the actual apply. If you need to process small tables you can skip this step and just use progress_apply instead
from thefuzz import process
from collections import OrderedDict
import swifter
def match(x):
matches = process.extract(x, df1, limit=6)
matches = list(OrderedDict((x, True) for x in matches).keys())
print(f'{x:20} : {matches}')
return str(matches)
df1 = df['name'].values
df2['matches'] = df2['name'].swifter.apply(lambda x: match(x))
I come from an Excel background but I love pandas and it has truly made me more efficient. Unfortunately, I probably carry over some bad habits from Excel. I have three large files (between 2 million and 13 million rows each) which contain data on interactions which could be tied together, unfortunately, there is no unique key connecting the files. I am literally concatenating (Excel formula) 3 fields into one new column on all three files.
Three columns which exist on each file which I combined together (the other fields would be like the reason for interaction on one file, the score on another file, and the some other data on the third file which I would like to tie together back to a certain agentID):
Date | CustomerID | AgentID
I edit my date format to be uniform on each file:
df[Date] = pd.to_datetime(df['Date'], coerce = True)
df[Date] = df[Date].apply(lambda x:x.date().strftime('%Y-%m-%d'))
Then I create a unique column (well, as unique as I can get it.. sometimes the same customer interacts with the same agent on the same date but this should be quite rare):
df[Unique] = df[Date].astype(str) + df[CustomerID].astype(str) + df[AgentID].astype(str)
I do the same steps for df2 and then:
combined = pd.merge(df, df2, how = 'left', on = 'Unique')
I typically send that to a new csv in case something crashes, gzip it, then read it again and do the same process again with the third file.
final = pd.merge(combined, df2, how = 'left', on = 'Unique')
As you can see, this takes time. I have to format the dates on each and then turn them into text, create an object column which adds to the filesize, and (due to the raw data issues themselves) drop duplicates so I don't accidentally inflate numbers. Is there a more efficient workflow for me to follow?
Instead of using on = 'Unique':
combined = pd.merge(df, df2, how = 'left', on = 'Unique')
you can pass a list of columns to the on keyword parameter:
combined = pd.merge(df, df2, how='left', on=['Date', 'CustomerID', 'AgentID'])
Pandas will correctly merge rows based on the triplet of values from the 'Date', 'CustomerID', 'AgentID' columns. This is safer (see below) and easier than building the Unique column.
For example,
import pandas as pd
import numpy as np
np.random.seed(2015)
df = pd.DataFrame({'Date': pd.to_datetime(['2000-1-1','2000-1-1','2000-1-2']),
'CustomerID':[1,1,2],
'AgentID':[10,10,11]})
df2 = df.copy()
df3 = df.copy()
L = len(df)
df['ABC'] = np.random.choice(list('ABC'), L)
df2['DEF'] = np.random.choice(list('DEF'), L)
df3['GHI'] = np.random.choice(list('GHI'), L)
df2 = df2.iloc[[0,2]]
combined = df
for x in [df2, df3]:
combined = pd.merge(combined, x, how='left', on=['Date','CustomerID', 'AgentID'])
yields
In [200]: combined
Out[200]:
AgentID CustomerID Date ABC DEF GHI
0 10 1 2000-1-1 C F H
1 10 1 2000-1-1 C F G
2 10 1 2000-1-1 A F H
3 10 1 2000-1-1 A F G
4 11 2 2000-1-2 A F I
A cautionary note:
Adding the CustomerID to the AgentID to create a Unique ID could be problematic
-- particularly if neither has a fixed-width format.
For example, if CustomerID = '12' and AgentID = '34' Then (ignoring the date which causes no problem since it does have a fixed-width) Unique would be
'1234'. But if CustomerID = '1' and AgentID = '234' then Unique would
again equal '1234'. So the Unique IDs may be mixing entirely different
customer/agent pairs.
PS. It is a good idea to parse the date strings into date-like objects
df['Date'] = pd.to_datetime(df['Date'], coerce=True)
Note that if you use
combined = pd.merge(combined, x, how='left', on=['Date','CustomerID', 'AgentID'])
it is not necessary to convert any of the columns back to strings.