Dirty ER¶
In this notebook we present the pyJedAI approach in the well-known ABT-BUY dataset. Dirty ER, is the process of dedeplication of one set.
Instalation¶
pyJedAI is an open-source library that can be installed from PyPI.
For more: pypi.org/project/pyjedai/
In [ ]:
!python --version
In [ ]:
!pip install pyjedai -U
In [2]:
!pip show pyjedai
Name: pyjedai Version: 0.0.5 Summary: An open-source library that builds powerful end-to-end Entity Resolution workflows. Home-page: Author: Author-email: Konstantinos Nikoletos <nikoletos.kon@gmail.com>, George Papadakis <gpapadis84@gmail.com> License: Apache Software License 2.0 Location: c:\users\nikol\anaconda3\lib\site-packages Requires: strsim, seaborn, matplotlib, optuna, networkx, faiss-cpu, scipy, rdflib, strsimpy, rdfpandas, regex, sentence-transformers, tqdm, nltk, pandas, pandocfilters, matplotlib-inline, numpy, pandas-profiling, PyYAML, transformers, tomli, gensim Required-by:
Imports
In [5]:
import os
import sys
import pandas as pd
import networkx
from networkx import draw, Graph
from pyjedai.utils import print_clusters, print_blocks, print_candidate_pairs
from pyjedai.evaluation import Evaluation
Reading the dataset - Dirty ER example¶
pyJedAI in order to perfrom needs only the tranformation of the initial data into a pandas DataFrame. Hence, pyJedAI can function in every structured or semi-structured data. In this case Abt-Buy dataset is provided as .csv files.
pyjedai <Data> module¶
Data module offers a numpber of options
- Selecting the parameters (columns) of the dataframe, in D1 (and in D2)
- Prints a detailed text analysis
- Stores a hidden mapping of the ids, and creates it if not exists.
In [7]:
from pyjedai.datamodel import Data
d1 = pd.read_csv("./../data/der/cora/cora.csv", sep='|')
gt = pd.read_csv("./../data/der/cora/cora_gt.csv", sep='|', header=None)
attr = ['Entity Id','author', 'title']
Data is the connecting module of all steps of the workflow
In [9]:
data = Data(
dataset_1=d1,
id_column_name_1='Entity Id',
ground_truth=gt,
attributes_1=attr
)
Block Building¶
It clusters entities into overlapping blocks in a lazy manner that relies on unsupervised blocking keys: every token in an attribute value forms a key. Blocks are then extracted, possibly using a transformation, based on its equality or on its similarity with other keys.
The following methods are currently supported:
- Standard/Token Blocking
- Sorted Neighborhood
- Extended Sorted Neighborhood
- Q-Grams Blocking
- Extended Q-Grams Blocking
- Suffix Arrays Blocking
- Extended Suffix Arrays Blocking
In [10]:
from pyjedai.block_building import (
StandardBlocking,
QGramsBlocking,
SuffixArraysBlocking,
ExtendedSuffixArraysBlocking,
ExtendedQGramsBlocking
)
In [11]:
bb = SuffixArraysBlocking(suffix_length=2)
blocks = bb.build_blocks(data)
Suffix Arrays Blocking: 0%| | 0/1295 [00:00<?, ?it/s]
In [13]:
_ = bb.evaluate(blocks)
***************************************************************************************************************************
Μethod: Suffix Arrays Blocking
***************************************************************************************************************************
Method name: Suffix Arrays Blocking
Parameters:
Suffix length: 2
Maximum Block Size: 53
Runtime: 0.1680 seconds
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Performance:
Precision: 4.28%
Recall: 75.77%
F1-score: 8.10%
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Block Cleaning¶
___Optional step___
Its goal is to clean a set of overlapping blocks from unnecessary comparisons, which can be either redundant (i.e., repeated comparisons that have already been executed in a previously examined block) or superfluous (i.e., comparisons that involve non-matching entities). Its methods operate on the coarse level of individual blocks or entities.
In [14]:
from pyjedai.block_cleaning import BlockFiltering
In [16]:
bc = BlockFiltering(ratio=0.9)
blocks = bc.process(blocks, data)
Block Filtering: 0%| | 0/3 [00:00<?, ?it/s]
In [18]:
_ = bc.evaluate(blocks)
***************************************************************************************************************************
Μethod: Block Filtering
***************************************************************************************************************************
Method name: Block Filtering
Parameters:
Ratio: 0.9
Runtime: 0.0930 seconds
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Performance:
Precision: 4.90%
Recall: 75.56%
F1-score: 9.21%
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Comparison Cleaning¶
___Optional step___
Similar to Block Cleaning, this step aims to clean a set of blocks from both redundant and superfluous comparisons. Unlike Block Cleaning, its methods operate on the finer granularity of individual comparisons.
The following methods are currently supported:
- Comparison Propagation
- Cardinality Edge Pruning (CEP)
- Cardinality Node Pruning (CNP)
- Weighed Edge Pruning (WEP)
- Weighed Node Pruning (WNP)
- Reciprocal Cardinality Node Pruning (ReCNP)
- Reciprocal Weighed Node Pruning (ReWNP)
- BLAST
Most of these methods are Meta-blocking techniques. All methods are optional, but competive, in the sense that only one of them can part of an ER workflow. For more details on the functionality of these methods, see here. They can be combined with one of the following weighting schemes:
- Aggregate Reciprocal Comparisons Scheme (ARCS)
- Common Blocks Scheme (CBS)
- Enhanced Common Blocks Scheme (ECBS)
- Jaccard Scheme (JS)
- Enhanced Jaccard Scheme (EJS)
In [20]:
from pyjedai.block_cleaning import BlockPurging
In [21]:
bp = BlockPurging(smoothing_factor=0.008)
blocks = bp.process(blocks, data)
Block Purging: 0%| | 0/3192 [00:00<?, ?it/s]
In [23]:
_ = bp.evaluate(blocks)
***************************************************************************************************************************
Μethod: Block Purging
***************************************************************************************************************************
Method name: Block Purging
Parameters:
Smoothing factor: 0.008
Max Comparisons per Block: 6.0
Runtime: 0.0480 seconds
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Performance:
Precision: 9.87%
Recall: 0.72%
F1-score: 1.33%
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Meta Blocking¶
In [24]:
from pyjedai.comparison_cleaning import (
WeightedEdgePruning,
WeightedNodePruning,
CardinalityEdgePruning,
CardinalityNodePruning,
BLAST,
ReciprocalCardinalityNodePruning,
ComparisonPropagation
)
In [25]:
mb = WeightedEdgePruning(weighting_scheme='CBS')
blocks = mb.process(blocks, data)
Weighted Edge Pruning: 0%| | 0/1295 [00:00<?, ?it/s]
In [28]:
_ = mb.evaluate(blocks)
***************************************************************************************************************************
Μethod: Weighted Edge Pruning
***************************************************************************************************************************
Method name: Weighted Edge Pruning
Parameters:
Node centric: False
Weighting scheme: CBS
Runtime: 0.0457 seconds
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Performance:
Precision: 75.49%
Recall: 0.45%
F1-score: 0.89%
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Entity Matching¶
It compares pairs of entity profiles, associating every pair with a similarity in [0,1]. Its output comprises the similarity graph, i.e., an undirected, weighted graph where the nodes correspond to entities and the edges connect pairs of compared entities.
In [29]:
from pyjedai.matching import EntityMatching
In [31]:
attr = {
'author' : 0.6,
'title' : 0.4
}
EM = EntityMatching(
metric='jaccard',
similarity_threshold=0.5
)
pairs_graph = EM.predict(blocks, data)
Entity Matching (jaccard): 0%| | 0/59 [00:00<?, ?it/s]
In [32]:
draw(pairs_graph)
In [34]:
_ = EM.evaluate(pairs_graph)
***************************************************************************************************************************
Μethod: Entity Matching
***************************************************************************************************************************
Method name: Entity Matching
Parameters:
Tokenizer: white_space_tokenizer
Metric: jaccard
Similarity Threshold: 0.5
Runtime: 0.0726 seconds
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Performance:
Precision: 97.67%
Recall: 0.24%
F1-score: 0.49%
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Entity Clustering¶
It takes as input the similarity graph produced by Entity Matching and partitions it into a set of equivalence clusters, with every cluster corresponding to a distinct real-world object.
In [39]:
from pyjedai.clustering import ConnectedComponentsClustering
In [40]:
ec = ConnectedComponentsClustering()
clusters = ec.process(pairs_graph, data)
In [43]:
_ = ec.evaluate(clusters)
***************************************************************************************************************************
Μethod: Connected Components Clustering
***************************************************************************************************************************
Method name: Connected Components Clustering
Parameters:
Runtime: 0.0000 seconds
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Performance:
Precision: 97.92%
Recall: 0.27%
F1-score: 0.55%
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Data Reading¶
Data is the connecting module of all steps of the workflow
In [44]:
from pyjedai.datamodel import Data
d1 = pd.read_csv("./../data/der/cora/cora.csv", sep='|')
gt = pd.read_csv("./../data/der/cora/cora_gt.csv", sep='|', header=None)
attr = ['Entity Id','author', 'title']
data = Data(
dataset_1=d1,
id_column_name_1='Entity Id',
ground_truth=gt,
attributes_1=attr
)
Similarity Joins¶
In [45]:
from pyjedai.joins import ΕJoin, TopKJoin
In [47]:
join = ΕJoin(similarity_threshold = 0.5,
metric = 'jaccard',
tokenization = 'qgrams_multiset',
qgrams = 2)
g = join.fit(data)
EJoin (jaccard): 0%| | 0/2590 [00:00<?, ?it/s]
In [48]:
_ = join.evaluate(g)
***************************************************************************************************************************
Μethod: EJoin
***************************************************************************************************************************
Method name: EJoin
Parameters:
similarity_threshold: 0.5
metric: jaccard
tokenization: qgrams_multiset
qgrams: 2
Runtime: 61.2715 seconds
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Performance:
Precision: 65.80%
Recall: 93.03%
F1-score: 77.08%
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Out[48]:
{'Precision %': 65.79553872746729,
'Recall %': 93.03421787709497,
'F1 %': 77.07921508124006,
'True Positives': 15987,
'False Positives': 8311,
'True Negatives': 828357.0,
'False Negatives': 1197}
In [49]:
topk_join = TopKJoin(K=20,
metric = 'jaccard',
tokenization = 'qgrams',
qgrams = 3)
g = topk_join.fit(data)
Top-K Join (jaccard): 0%| | 0/2590 [00:00<?, ?it/s]
In [50]:
draw(g)
In [51]:
topk_join.evaluate(g)
***************************************************************************************************************************
Μethod: Top-K Join
***************************************************************************************************************************
Method name: Top-K Join
Parameters:
similarity_threshold: 0.25547445255474455
K: 20
metric: jaccard
tokenization: qgrams
qgrams: 3
Runtime: 41.0031 seconds
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Performance:
Precision: 58.34%
Recall: 63.75%
F1-score: 60.92%
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Out[51]:
{'Precision %': 58.340434597358325,
'Recall %': 63.74534450651769,
'F1 %': 60.923248053392655,
'True Positives': 10954,
'False Positives': 7822,
'True Negatives': 823813.0,
'False Negatives': 6230}
Entity Clustering¶
In [52]:
from pyjedai.clustering import ConnectedComponentsClustering
In [54]:
ccc = ConnectedComponentsClustering()
clusters = ccc.process(g, data)
In [56]:
_ = ccc.evaluate(clusters)
***************************************************************************************************************************
Μethod: Connected Components Clustering
***************************************************************************************************************************
Method name: Connected Components Clustering
Parameters:
Runtime: 0.0630 seconds
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Performance:
Precision: 2.05%
Recall: 100.00%
F1-score: 4.02%
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────