A tool facilitating matching for any dataset discovery method. Also, an extensible experiment suite for state-of-the-art schema matching methods.
A python package for capturing potential relationships among columns of different tabular datasets, which are given in the form of pandas DataFrames. Valentine is based on Valentine: Evaluating Matching Techniques for Dataset Discovery
You can find more information about the research supporting Valentine here.
The original experimental suite version of Valentine, as first published for the needs of the research paper, can be still found here.
To install Valentine simply run:
pip install valentine
Valentine can be used to find matches among columns of a given pair of pandas DataFrames.
In order to do so, the user can choose one of the following 5 matching methods:
Coma(int: max_n, bool: use_instances, str: java_xmx)
is a python wrapper around COMA 3.0 Comunity edition
Cupid(float: w_struct, float: leaf_w_struct, float: th_accept)
is the python implementation of the paper Generic Schema Matching with Cupid
DistributionBased(float: threshold1, float: threshold2)
is the python implementation of the paper Automatic Discovery of Attributes in Relational Databases
JaccardDistanceMatcher(float: threshold_dist)
is a baseline method that uses Jaccard Similarity between columns to assess their correspondence score, optionally enhanced by a string similarity measure of choice.
threshold_dist(float) - Acceptance threshold for assessing two strings as equal, default is 0.8.
distance_fun(StringDistanceFunction) - String similarity function used to assess whether two strings are equal. The enumeration class type StringDistanceFunction
can be imported from valentine.algorithms.jaccard_distance
. Functions currently supported are:
StringDistanceFunction.Levenshtein
: Levenshtein distance
StringDistanceFunction.DamerauLevenshtein
: Damerau-Levenshtein distance
StringDistanceFunction.Hamming
: Hamming distance
StringDistanceFunction.Jaro
: Jaro distance
StringDistanceFunction.JaroWinkler
: Jaro-Winkler distance
* StringDistanceFunction.Exact
: String equality ==
SimilarityFlooding(str: coeff_policy, str: formula)
is the python implementation of the paper Similarity Flooding: A Versatile Graph Matching Algorithmand its Application to Schema Matching
After selecting one of the 5 matching methods, the user can initiate the pairwise matching process in the following way:
matches = valentine_match(df1, df2, matcher, df1_name, df2_name)
where df1 and df2 are the two pandas DataFrames for which we want to find matches and matcher is one of Coma, Cupid, DistributionBased, JaccardLevenMatcher or SimilarityFlooding. The user can also input a name for each DataFrame (defaults are "table_1" and "table_2"). Function valentine_match
returns a MatcherResults object, which is a dictionary with additional convenience methods, such as one_to_one
, take_top_percent
, get_metrics
and more. It stores as keys column pairs from the two DataFrames and as values the corresponding similarity scores.
After selecting one of the 5 matching methods, the user can initiate the batch matching process in the following way:
matches = valentine_match_batch(df_iter_1, df_iter_2, matcher, df_iter_1_names, df_iter_2_names)
where df_iter_1 and df_iter_2 are the two iterable structures containing pandas DataFrames for which we want to find matches and matcher is one of Coma, Cupid, DistributionBased, JaccardLevenMatcher or SimilarityFlooding. The user can also input an iterable with names for each DataFrame. Function valentine_match_batch
returns a MatcherResults object, which is a dictionary with additional convenience methods, such as one_to_one
, take_top_percent
, get_metrics
and more. It stores as keys column pairs from the two DataFrames and as values the corresponding similarity scores.
The MatcherResults
instance has some convenience methods that the user can use to either obtain a subset of the data or to transform the data. This instance is a dictionary and is sorted upon instantiation, from high similarity to low similarity.
top_n_matches = matches.take_top_n(5)
top_n_percent_matches = matches.take_top_percent(25)
one_to_one_matches = matches.one_to_one()
The MatcherResults instance that is returned by valentine_match
or valentine_match_batch
also has a get_metrics
method that the user can use
metrics = matches.get_metrics(ground_truth)
in order to get all effectiveness metrics, such as Precision, Recall, F1-score and others as described in the original Valentine paper. In order to do so, the user needs to also input the ground truth of matches based on which the metrics will be calculated. The ground truth can be given as a list of tuples representing column matches that should hold (see example below).
By default, all the core metrics will be used for this with default parameters, but the user can also customize which metrics to run with what parameters, and implement own custom metrics by extending from the Metric
base class. Some sets of metrics are available as well.
from valentine.metrics import F1Score, PrecisionTopNPercent, METRICS_PRECISION_INCREASING_N
metrics_custom = matches.get_metrics(ground_truth, metrics={F1Score(one_to_one=False), PrecisionTopNPercent(n=70)})
metrics_prefefined_set = matches.get_metrics(ground_truth, metrics=METRICS_PRECISION_INCREASING_N)
The following block of code shows: 1) how to run a matcher from Valentine on two DataFrames storing information about authors and their publications, and then 2) how to assess its effectiveness based on a given ground truth (a more extensive example is shown in valentine_example.py
):
import os
import pandas as pd
from valentine import valentine_match
from valentine.algorithms import Coma
# Load data using pandas
d1_path = os.path.join('data', 'authors1.csv')
d2_path = os.path.join('data', 'authors2.csv')
df1 = pd.read_csv(d1_path)
df2 = pd.read_csv(d2_path)
# Instantiate matcher and run
matcher = Coma(use_instances=True)
matches = valentine_match(df1, df2, matcher)
print(matches)
# If ground truth available valentine could calculate the metrics
ground_truth = [('Cited by', 'Cited by'),
('Authors', 'Authors'),
('EID', 'EID')]
metrics = matches.get_metrics(ground_truth)
print(metrics)
The output of the above code block is:
{
(('table_1', 'Cited by'), ('table_2', 'Cited by')): 0.86994505,
(('table_1', 'Authors'), ('table_2', 'Authors')): 0.8679843,
(('table_1', 'EID'), ('table_2', 'EID')): 0.8571245
}
{
'Recall': 1.0,
'F1Score': 1.0,
'RecallAtSizeofGroundTruth': 1.0,
'Precision': 1.0,
'PrecisionTop10Percent': 1.0
}
Original Valentine paper:
@inproceedings{koutras2021valentine,
title={Valentine: Evaluating Matching Techniques for Dataset Discovery},
author={Koutras, Christos and Siachamis, George and Ionescu, Andra and Psarakis, Kyriakos and Brons, Jerry and Fragkoulis, Marios and Lofi, Christoph and Bonifati, Angela and Katsifodimos, Asterios},
booktitle={2021 IEEE 37th International Conference on Data Engineering (ICDE)},
pages={468--479},
year={2021},
organization={IEEE}
}
Demo Paper:
@article{koutras2021demo,
title={Valentine in Action: Matching Tabular Data at Scale},
author={Koutras, Christos and Psarakis, Kyriakos and Siachamis, George and Ionescu, Andra and Fragkoulis, Marios and Bonifati, Angela and Katsifodimos, Asterios},
journal={VLDB},
volume={14},
number={12},
pages={2871--2874},
year={2021},
publisher={VLDB Endowment}
}