Improving Quality of Entity Resolution Using a Cascade Approach

Khizer Syed, Onais Khan Mohammed, John Talburt, Adeeba Tarannum,

Altaf Mohammed, Mudasar Ali Mir and Mahboob Khan Mohammed

Department of Information Quality, University of Arkansas at Little Rock, Little Rock, U.S.A.

Keywords: Entity Resolution, Record Linkage, Cascade, Group Membership Graph, Household Discovery,

Entity Detection, Entity Relationship, Entity Recognition, Data Standardization, Master Data Management.

Abstract: Entity Resolution (ER) is a critical technique in data management, designed to determine whether two or more

data references correspond to the same real-world entity. This process is essential for cleansing datasets and

linking information across diverse records. A variant of this technique, Binary Entity Resolution, focuses on

the direct comparison of data pairs without incorporating the transitive closure typically found in cluster-

based approaches. Unlike cluster-based ER, where indirect linkages imply broader associations among

multiple records (e.g., A is linked with B, and B is linked with C, thereby linking A with C indirectly), Binary

ER performs pairwise matching, resulting in a straightforward outcome—a series of pairs from two distinct

sources. In this paper, we present a novel improvement to the cascade process used in entity resolution.

Specifically, our data-centric, descending confidence cascade approach systematically orders linking methods

based on their confidence levels in descending order. This method ensures that higher confidence methods,

which are more accurate, are applied first, potentially enhancing the accuracy of subsequent, lower-confidence

methods. As a result, our approach produces better quality matches than traditional methods that do not utilize

a cascading approach, leading to more accurate entity resolution while maintaining high-quality links. This

improvement is particularly significant in Binary ER, where the focus is on pairwise matches, and the quality

of each link is crucial.

1 INTRODUCTION

Entity Resolution techniques have evolved to help

solve the problem of linking entities across large

datasets and is also referred to as record linkage or

deduplication. Entity Resolution is the process of

identifying and matching records that pair to the same

real-world entity across different data sets (Cohen,

2003). Binary Entity Resolution is an approach to

evolve cluster comparisons to more effectively link

pairs of records between two data sets. This technique

has several advantages over traditional methods used

to handle complex data structures.

The cascading process involves multiple stages

where records not linked in initial attempts are passed

through subsequent cascades for further analysis.

This iterative approach allows the system to refine

and attempt to link more records at each stage.

Cascading indices can be defined as the stricter ones

to match prior to the lesser strict. For example, the

first cascade can be an exact match as variation i.e,

'Johnny Doe' versus 'John Doe' are considered. By the

third and fourth cascades, even more lenient matching

criteria might be used to link records that were

previously unlinked. After completing the cascade

stages, any remaining unlinked records are put

through the household discovery process. Here,

records are analyzed to identify households, linking

records based on shared addresses or other household

indicators. Links established through earlier cascades

are considered strong links, while those made during

household discovery are categorized as weak.

(Mohammed, O.K. et al, 2024). The use of cascading

stages allows the system to handle large datasets

efficiently by systematically narrowing the focus of

computational resources to the most promising

matches first, before exploring more complex and

nuanced potential links in later stages. Record linkage

is the task of quickly and accurately identifying

records corresponding to the same entity from one or

more data sources. Entities of interest include

individuals, companies, geographic regions, families,

or households (Fellegi, I. P., & Sunter, 1969).

Clustering divides data patterns into subsets in such a

Syed, K., Mohammed, O. K., Talburt, J., Tarannum, A., Mohammed, A., Mir, M. A. and Mohammed, M. K.

Improving Quality of Entity Resolution Using a Cascade Approach.

DOI: 10.5220/0013077400003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 60-69

ISBN: 978-989-758-737-5; ISSN: 2184-433X

way that similar patterns are clustered together. The

patterns are thereby managed into a well-formed

evaluation that designates the population being

sampled (Winkler, W. E., 1990). Binary entity

resolution (ER) and cluster-based ER represent

different techniques to resolving entities in datasets.

In binary ER, the focus is on comparing individual

pairs of references from separate files to determine

equivalence. Unlike cluster-based ER, binary ER

does not use transitive closure, in that it doesn't

automatically link references that are indirectly

related. The output of binary ER consists of linked

pairs, with each pair representing a match between a

reference from one file to another. Whereas cluster

ER operates by initially identifying pairs of

references that match.

The methodology for computing metrics in binary

Entity Resolution (ER) shares similarities with

cluster- based ER, yet there are notable distinctions.

The first key difference lies in the data sources:

cluster ER operates on a unified dataset or single file,

whereas binary ER is executed across two distinct

datasets. The second distinction pertains to the

application of the transitive closure principle; cluster

ER integrates this principle to identify related entities

across multiple records, while binary ER does not

incorporate transitive closure, focusing instead on

direct comparisons between the two datasets to help

identify pairs that represent the same entity. These

differences highlight the unique challenges and

considerations inherent to each ER approach.

Initial objective of our approach is to identify the

cascading methods that can help uniquely identify

links between people, places and things; i.e, Social

Security Number for a person versus a Part Number

for equipment, in order to establish potential matches

(Fellegi, I. P., & Sunter, 1969), The groundwork

around finding the most apparent connections is

fundamental to any ER process. Following this, the

process employs a series of less stringent filters, such

as comparisons on name, address, date-of-birth, and

other demographics of a person. After direct pair

linking, indirect methods such as household

connections may be employed to further additional

links (Mohammed, O.K. et al, 2024). By

methodically narrowing down from the most accurate

identifiers to broader characteristics, the cascade

approach enhances the integrity and utility of data

linkage, making it an essential tool in efficient data

management and integration tasks. Employing a

tiered approach to implement cascading enables

precise linking and helps to ensure that only

equivalent pairs of references with a high confidence

are brought together.

Binary ER can support pairwise linking These

different types of pairwise linking are crucial for

accurately capturing the complexity of real-world

relationships between data records (Mohammed,

O.K. et al, 2024). By supporting these varying levels

of linkage, our Binary ER approach ensures

flexibility and precision in matching records, which

is essential for improving the overall accuracy and

effectiveness of the entity resolution process in this

work.

One to one: One reference in file A matches to one

reference in file B.

One to many: One reference in file A could match

to more than one reference in file B, but each

reference in file B has at most one matching reference

in file A. These different types of pairwise linking are

crucial for accurately capturing the complexity of

real-world relationships between data records. This is

particularly important in scenarios where duplicate

records exist in the database, necessitating the use of

the one-to-many scenario to ensure all possible

matches are identified. Additionally, in certain cases,

such as when generating credit files at a credit score

company, there is a requirement that only one

matching entity is sent to the user, meaning that the

system must handle one-to-one matching with

precision. By supporting these varying levels of

linkage, our Binary ER approach ensures flexibility

and precision in matching records, which is essential

for improving the overall accuracy and effectiveness

of the entity resolution process in this work.

2 LITERATURE REVIEW

Entity Resolution (ER), also known as record linkage

or deduplication, is a critical process in data

management, aimed at identifying and linking records

that refer to the same real-world entity across multiple

datasets. Over the years, various approaches have

been developed to address the challenges of ER,

ranging from traditional rule-based systems to more

advanced machine learning methods. One of the

foundational methods in ER is the Fellegi-Sunter

model, which introduced probabilistic techniques to

resolve records based on a set of matching criteria and

decision rules (Fellegi, I. P., & Sunter, A. B., 1969).

This model laid the groundwork for many subsequent

ER systems by formalizing the process of comparing

and linking records. However, traditional

probabilistic models often struggle with complex and

large-scale datasets, where the sheer volume of

records and variations in data can lead to inaccurate

matches and inefficiencies.

Improving Quality of Entity Resolution Using a Cascade Approach

To address these challenges, more recent research

has explored the use of similarity metrics such as the

Jaro- Winkler distance and the Levenshtein edit

distance, which are effective in handling minor

variations and typographical errors in textual data

(Winkler, W. E., 1990) (Christen, P., 2012). These

metrics have been widely adopted in ER systems for

their ability to enhance the accuracy of matching by

comparing strings based on their similarity. Jaro-

Winkler has been noted for its effectiveness in

resolving names where common prefixes are shared,

while Levenshtein is valuable for its general

applicability across various types of textual data

(Cohen, W. W., Ravikumar, P., & Fienberg, S. E.,

2003). Despite these advancements, traditional ER

methods still face significant limitations when

dealing with large datasets that contain partial or

missing data, duplicates, and diverse data formats. To

mitigate these issues, clustering-based approaches

have been developed, where records are grouped into

clusters based on shared attributes, and links are

established among all records within a cluster

(Papadakis, G., Palpanas, T., & Koutrika, G., 2020).

While clustering can improve the handling of indirect

links, it often introduces complexity and can lead to

over- linking, where unrelated records are incorrectly

grouped together.

In the industry, ER plays a crucial role in sectors

such as finance, healthcare, and retail, where accurate

data linkage is essential for operational efficiency and

regulatory compliance. For example, credit reporting

agencies use ER to aggregate and maintain accurate

credit histories by linking records from different

financial institutions. In healthcare, ER is used to

integrate patient records across different providers,

ensuring that healthcare professionals have a

comprehensive view of a patient’s medical history.

Retail companies leverage ER to create unified

customer profiles by linking purchase data across

multiple channels, enabling personalized marketing

strategies and improved customer service (Talburt, J.

R., 2011). The approach presented in this paper builds

on these established methods but introduces a novel

improvement through the use of a data-centric,

descending confidence cascade framework for Binary

ER. Unlike clustering-based methods, our approach

focuses on pairwise record matching, ensuring that

each link is directly evaluated and validated. This

cascade approach begins with high-confidence

attributes, such as Social Security Number (SSN), and

progressively incorporates less distinct attributes,

such as name and address, in subsequent stages. By

employing this descending confidence strategy, our

method not only improves the precision of matches

but also efficiently handles the challenges of partial

data and duplicates (Papadakis, G., Kirielle, N., &

Palpanas, T., 2024).

Furthermore, while traditional ER methods often

rely on a single-shot matching process, our cascading

approach introduces iterative refinement, where

records that are not matched in early stages are

reconsidered with more lenient criteria in later stages.

This iterative process, coupled with the use of hashing

techniques to generate unique IDs for records,

ensures that our system is both space-efficient and

highly accurate in linking records across large and

complex datasets (Christen, P., Vatsalan, D., &

Verykios, V. S., 2014)

3 METHODOLOGY

The proposed binary entity resolution process is

designed to effectively link records across two

datasets, named FileA and FileB. Each step in this

process, along with iterative cascading and household

discovery, is structured to maximize accuracy and

efficiency in identifying records that refer to the same

entity.

Figure 1: Flow of binary entity resolution (in detail).

3.1 Data Preparation

Initially, records from FileA and FileB are

distinguished by prefixing their IDs with 'A' and 'B'

respectively. This step ensures that each record can be

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

uniquely identified even after the datasets are merged

into a single data frame, for example, If FileA

contains a record ID '123', it becomes 'A123';

similarly, '123' from FileB becomes 'B123'.

In addition to using similarity metrics like Jaro-

Winkler and Levenshtein edit distance, another

effective technique in Binary Entity Resolution (ER) is

the use of hashing to generate unique identifiers for

records. By creating a hash of the record, or a

combination of key attributes within the record, we can

produce a unique ID that serves as a concise and space-

efficient representation of the record. This method

significantly reduces the storage requirements

compared to storing full records or complex identifiers,

while still enabling precise matching. Hashing is

particularly useful when dealing with large datasets, as

it simplifies the process of record linkage by

converting potentially large and complex strings into

fixed-size hash values. This approach can complement

traditional similarity metrics, providing a fast and

efficient way to identify and link records with minimal

computational overhead.

3.2 Tokenization and Frequency

Calculation with Blocking

Attributes or columns from the combined data frame

are tokenized, breaking down data into manageable

parts or tokens, and frequencies of these tokens are

calculated. This aids in identifying and comparing

elements across the datasets. For an instance the name

'John Doe' might be tokenized into 'John' and 'Doe'. If

these tokens appear frequently, they help in the

blocking and matching processes. A blocking strategy

is applied to group records by shared tokens, reducing

computational complexity by ensuring that only

likely matching records are compared. All records

with the token 'Doe' are grouped together, reducing

the number of comparisons necessary by focusing

only on records within the same block.

3.3 Pair Generation and Sorting

Within each block, pairs of records (one from FileA

and one from FileB) are generated and sorted based

on the tokens they share. This is critical for efficiently

finding potential matches. A record 'A123: John Doe'

might be paired with 'B456: Johnny Doe' because

they share the token 'Doe'.

3.4 Similarity Calculation and

Selection

Similarity metrics are calculated for each pair to

determine how closely they match based on

predefined criteria. Pairs with similarity scores above

a certain threshold are selected as potential matches.

Example: The pair 'A123: John Doe' and 'B456:

Johnny Doe' might have a high similarity score if

additional attributes like address or date of birth also

match.

In the context of Binary Entity Resolution (ER),

selecting appropriate similarity metrics is crucial for

accurately matching records. Two widely used

metrics in this domain are the Jaro-Winkler similarity

and the Levenshtein edit distance. Jaro-Winkler is

particularly effective for identifying similar strings

that share common prefixes, making it well-suited for

matching names and other textual data where minor

spelling variations or typographical errors are

common. This metric gives higher scores to strings

that match from the beginning, which can be

particularly useful in resolving names or addresses

where such patterns often occur.

On the other hand, the Levenshtein edit distance

measures the minimum number of single-character

edits (insertions, deletions, or substitutions) required

to transform one string into another. This metric is

highly effective for matching records where small

variations in text are expected, such as in addresses,

names, or other attributes. By applying these metrics

within the Binary ER framework, we can enhance the

accuracy of pairwise record comparisons, ensuring

that even records with slight discrepancies are

correctly identified as matches. (Winkler, W. E.,

1990). The combination of Jaro-Winkler for prefix-

sensitive comparisons and Levenshtein for more

general string similarity provides a robust approach to

handling the diverse challenges inherent in entity

resolution tasks. Consider two datasets, Dataset A

(denoted as A) and Dataset B (denoted as B), each

containing records that represent individuals. These

records include various attributes such as Social

Security Number (SSN), Name, Date of Birth (DOB),

and Street Number. In our cascade approach, we

begin by blocking and matching entities using the

most distinct attributes in the first cascade. For

example, in the first cascade (i = 1), we might use

SSN as the blocking attribute (B1). This means that

records ea from A and eb from B are grouped together

in the same block only if they share the same SSN.

Within each block, the similarity score S(ea, eb) is

calculated, and only those pairs with a similarity score

above a high matching threshold θ(1) are considered

as matches.

As we move to the second cascade (i = 2), we

lower the distinctiveness of the blocking attribute.

Here, we might use a combination of Name and Date

Improving Quality of Entity Resolution Using a Cascade Approach

of Birth (B2) for blocking. This allows for a broader

comparison, grouping entities that might not have

matched on SSN but could potentially match on

Name and DOB. The matching threshold θ(2) is

slightly lowered to account for the increased

possibility of variation in these attributes.

In the third cascade (i = 3), we might further relax

the blocking criteria by including Street Number (B3)

in the blocking attributes. This allows the system to

capture potential matches that share a name and

birthdate but may have slight differences in their SSN

or other identifying details. The matching threshold

θ(3) is adjusted accordingly to further refine the

matching process.

After the cascades, we might encounter situations

where we only have partial data, such as names and

addresses. For instance, if we find that "John Doe"

and "Mary Doe" are living at "123 Oak Street" in one

dataset, and we also find "John Doe" and "Mary Doe"

living at "456 Pine Street" in another dataset, the

system can attempt to link these entities through

household discovery. If one of these pairs was already

linked in an earlier cascade (e.g., Cascade 1), we

might consider this a "strong household link."

However, if neither pair was linked previously, and

the connection is made based solely on the shared

household members and address patterns, it could be

classified as a "weak household link."

3.5 Pseudocode

# Load data from two sources datasetA =

load_data("Dataset_A") datasetB =

load_data("Dataset_B")

# Initialize variables for unmatched

records and final matches unmatchedA,

unmatchedB = datasetA, datasetB

all_matches = []

# Define the number of cascades

cascades = 4

# Perform cascading entity

resolution

for i in range(1, cascades + 1):

threshold = adjust_threshold(i)

matches =

entity_resolution(unmatchedA,

unmatchedB, threshold)

all_matches.extend(matches)

unmatchedA,unmatchedB =

update_unmatched(unmatchedA,

unmatchedB, matches)

# Conduct household discovery for

remaining unmatched records

household_links=

household_discovery(unmatchedA +

unmatchedB)

# Append household discovery links to

all matches

all_matches.extend(household_links)

# Output or process all matches from

cascades and household discovery

output_matches(all_matches)

3.5.1 Annotated Dataset

In this study, we focus on the task of Binary Entity

Resolution (ER) across three distinct datasets, each

presenting unique characteristics and challenges. The

datasets include TruthFile, EasyFile, and

MediumFile. TruthFile serves as the input or request

file in the ER process, containing the ground truth for

entities with comprehensive attribute coverage. It

comprises 1 million records, providing a baseline for

matching against other datasets.

EasyFile is characterized by its completeness,

including all relevant attributes such as SSN,

FirstName, LastName, House Number, Street

Address, City, State, Zip, DOB-Day, DOB-Month,

DOB-Year, Phone, Occupation, and Salary. While

EasyFile includes the full SSN and has no missing

data, it does contain some errors, such as nicknames,

switched dates of birth (DOB), and slightly rounded

salaries. Additionally, a certain percentage of

individuals in EasyFile belong to the same household

and share the same phone number, which adds an

element of complexity to the matching process.

Despite these errors, EasyFile remains a rich source

of data for accurate matching due to its overall

completeness.

MediumFile differs from EasyFile by lacking full

SSNs, instead providing only the last four digits. It

also omits telephone numbers, and the salary data is

further rounded compared to EasyFile. Additionally,

MediumFile contains approximately 900,000 records

with duplicates, increasing the complexity of the ER

process. The presence of partial identifiers and

duplicates requires more sophisticated matching

techniques. When matching TruthFile to EasyFile,

the process is relatively straightforward, following a

one- to-one mapping facilitated by the completeness

and uniqueness of the EasyFile records. However, the

complexity increases significantly when dealing with

matches from TruthFile to MediumFile. In these

cases, one-to-many relationships arise due to the

presence of duplicate records in MediumFile.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

4 BASELINE SYSTEM

In our study, we incorporate a baseline system that

employs a traditional approach to entity resolution,

implemented using the record-linkage library in

Python. This system performs record linkage without

cascades, executing a single round of blocking and

matching to identify all possible record pairs at once.

The one-shot method applies a fixed set of blocking

and matching rules, using attributes such as the last

four digits of the SSN, the last three characters of the

name combined with the address, and the street name

combined with the last name. For matching, the system

utilizes the Levenshtein edit distance with a threshold

of 75 for name, address, and date of birth (DOB),

allowing for a DOB tolerance of 1 year. This approach

is applied uniformly across the MediumFile dataset.

For the EasyFile, the baseline system uses the

same blocking criteria but applies a higher matching

threshold of 85, which reflects the cleaner and more

complete data available in this file. In contrast, our

binary ER framework utilizes a cascading approach,

which involves multiple stages with progressively

relaxed matching criteria. The first cascade uses strict

rules, like the high cutoffs in the baseline system, to

identify high-confidence matches. Records

successfully linked in this stage are removed from

subsequent cascades. An exception is made for one-

to- many relationships, where only the response

records that have been linked are removed, allowing

the request records to be considered for further

linkages. This iterative approach helps manage

complexities such as duplicates, missing data, and

closely related records, providing a more nuanced and

accurate resolution.

Table 1: Configurations of the record-linkage library,

blocking and matching criteria. (baseline system).

Block

Criteria

Matching

Attributes

Matching

Technique

Threshold

Last4SSN,

Name with

Address

Last three of

name with

address, street

with last name

Levenshtein

edit distance

Last4SSN,

Name with

Address

Last three of

name with

address, street

with last name

Levenshtein

edit distance

The baseline system serves as a useful

comparative benchmark, highlighting the advantages

of the cascading method in handling the intricacies of

our datasets, including the TruthFile, EasyFile, and

MediumFile, and achieving more precise entity

resolution outcomes.

5 RESULTS

Table 2: Configurations of blocking and matching criteria

for easy file.

Cascade

Blocking

Attribute

Matchi

Attrib

utes

Threshold

Cascade 1

TruthSSN

(full)

TruthLastName

→

EasyLastName

First N=5

Cascade 2

TruthFirstN

ame (first 2),

TruthSSN

(last 4),

TruthState

(full),

TruthZip

(first 3)

TruthLastName

→

EasyLastName

First N=5

Cascade 3

TruthDOBY

ear (last 4),

TruthDOB

Month (full),

TruthHouse

Number (last

4),

TruthZip

(first 3)

TruthStreetAddr

ess

→

EasyStreetAddr

ess

Cascade 4

TruthPho

(last

4),

TruthHou

seNumber

(first 1)

TruthLastName

→

EasyFirstName,

TruthFirstNam

e →

EasyLastName,

TruthState

→ EasyState

Cascade 5

TruthStre

etAddress

(full),

TruthZip

(first 3)

TruthLastName

→

EasyLastName

First N=5

Cascade 6

TruthCity

(full),

TruthState

(full)

TruthLastName

→

EasyLastName

First N=5

Cascade 7

TruthSSN

(last 4)

TruthCity

→ EasyCity

First N=5

Improving Quality of Entity Resolution Using a Cascade Approach

As shown in Table-1 In the Easy dataset, the initial

cascade applied strict matching criteria, focusing on

exact matches for high-confidence identification.

This stage resulted in 853,284 true positives (TP),

with a precision of 1.0, indicating no false positives

(FP) were identified. However, the initial recall was

0.7757, reflecting the system's conservative

approach, which led to 246,716 false negatives (FN).

As the criteria relaxed in subsequent cascades, the

system's recall steadily improved, reaching 0.9526 in

the second cascade and 0.9867 by the third. The

precision remained consistently at 1.0 across all

stages, underscoring the method's reliability in

avoiding incorrect matches. By the fifth cascade, the

system achieved perfect scores in precision, recall,

and F-measure (1.0), indicating that all expected

matches were correctly identified without error.

Table 3: Precision, Recall, and F-Measure Across Cascades

for EasyFile.

Casca

(TP) (FN) P R

F-

meas

ure

1 8,53,284 2,46,716 1 0.7757

0.873

2 10,47,880 52,120 1 0.9526

0.975

3 10,85,356 14,644 1 0.9867

0.993

10,97,211

2,789 1 0.9975

0.998

10,99,953

47 1 1 1

10,99,993

7 1 1 1

11,00,000

0 1 1 1

Table 4: Comparison of blocking and matching criteria for

medium file.

Cascade

Blocking

Attributes

Matching

Attributes

Threshold

Cascade

SSN (last 4 digits),

State (full),

DOBYear (full),

HouseNumber

(full), Salary (full)

StreetAddr

ess, City,

Occupation

0.69

Cascade

Zip (full),

Salary (full),

Occupation (full),

HouseNumber (first

2 digits)

City, State,

StreetAddr

ess

0.69

Cascade

DOBDay (full),

DOBMonth (full),

DOBYear (full),

Salary (full)

City, State 0.69

Cascade

City (full),

State (full),

Salary (full)

StreetAddr

ess,

FirstName

(first 3

letters)

0.69

Cascade

SSN (last 2 digits),

State (full),

Salary (full)

StreetAddr

ess, City,

Occupation

0.69

Cascade

SSN (last 4 digits),

Salary (full),

State (full)

State,

LastName

(first 3

letters)

0.69

Cascade

Salary (full), State

(full), HouseNumber

(full)

StreetAddr

ess, City,

Occupation

0.49

Across both datasets, the cascading approach proved

to be highly effective, consistently delivering high

precision from the outset and progressively

improving recall. The stepwise relaxation of

matching criteria allowed the system to initially focus

on high- confidence matches and subsequently

broaden its scope to include more complex cases. The

final stages of both the Easy and Medium datasets

showed that the system could accurately identify all

true matches without any false positives, resulting in

perfect F- measure scores.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

Table 5: Precision, Recall, and F-Measure Across Cascades

for Medium.

Cascade FP FN P R

F- measure

1 0

1,82,223

1 0.8178 0.8998

2 0

1,08,046

1 0.892 0.9429

3 0 19,177 1 0.9808 0.9903

4 0 1,689 1 0.9983 0.9991

5 0 290 1 0.9997 0.9998

6 0 124 1 0.9999 0.9999

7 0 0 1 1 1

Table 6: Comparison of record-linkage and Binary ER

results.

Metric

Cascading

Approach

(EasyFile)

Record

Linkage

Library

(EasyFile

)

Cascading

Approach

(MediumFi

le)

Record

Linkag e

Librar y

(Mediu

mFile)

Precision 1 0.99 1 0.94

Recall 1 0.90 1 0.90

F-

measure

1 0.94 1 0.92

Accuracy 1 0.99 1 0.99

Balanced

Accuracy

1 0.99 1 0.99

Figure 2: Comparative Analysis of Entity Resolution

Methods Across EasyFile and MediumFile Datasets.

Figure 3: Comparative Analysis of Entity Resolution

Methods Across EasyFile and MediumFile Datasets.

While comparing the record-linkage library and

the proposed approach of binary entity resolution

table-1 clearly illustrates the comparative

performance of the two approaches across different

metrics. The custom cascading approach consistently

outperformed the record-linkage library in all

measured aspects. Specifically, the cascading method

achieved perfect precision and recall for the EasyFile

dataset, resulting in an F-measure of 1.0, with no false

positives or false negatives. This demonstrates the

method's efficacy in identifying true matches

accurately and eliminating incorrect linkages. For the

MediumFile, the cascadingapproach also achieved

near-perfect metrics with a precision of 0.99, a recall

of 0.99, and an F-measure of 0.99. The slightly lower

performance compared to the EasyFile is attributed to

the increased complexity of the dataset, including

partial data and duplicates. Nonetheless, the approach

maintained a minimal number of false positives

(1,043) and false negatives (368), showcasing its

robustness in handling more challenging scenarios.

In contrast, the record-linkage library exhibited a

lower precision and recall, particularly for the

MediumFile, with an F-measure of 0.924. The higher

false positive and false negative rates in both datasets

highlight the limitations of a one-shot linkage method,

which lacks iterative refinement and handling of

complex data relationships such as one-to-many

linkages. Overall, the comparative analysis

underscores the superiority of the custom cascading

approach in achieving high accuracy and balanced

accuracy, effectively minimizing errors and ensuring a

Improving Quality of Entity Resolution Using a Cascade Approach

comprehensive linkage of records across different

datasets.

While comparing the record-linkage library and

the proposed approach of binary entity resolution

table-1 clearly illustrates the comparative

performance of the two approaches across different

metrics. The custom cascading approach consistently

outperformed the record-linkage library in all

measured aspects. Specifically, the cascading method

achieved perfect precision and recall for the EasyFile

dataset, resulting in an F-measure of 1.0, with no false

positives or false negatives. This demonstrates the

method's efficacy in identifying true matches

accurately and eliminating incorrect linkages. For the

MediumFile, the cascading approach also achieved

near-perfect metrics with a precision of 0.99, a recall

of 0.99, and an F-measure of 0.99. The slightly lower

performance compared to the EasyFile is attributed to

the increased complexity of the dataset, including

partial data and duplicates. Nonetheless, the approach

maintained a minimal number of false positives

(1,043) and false negatives (368), showcasing its

robustness in handling more challenging scenarios.

In contrast, the record-linkage library exhibited a

lower precision and recall, particularly for the

MediumFile, with an F-measure of 0.924. The higher

false positive and false negative rates in both datasets

highlight the limitations of a one-shot linkage

method, which lacks iterative refinement and

handling of complex data relationships such as one-

to-many linkages. Overall, the comparative analysis

underscores the superiority of the custom cascading

approach in achieving high accuracy and balanced

accuracy, effectively minimizing errors and ensuring

a comprehensive linkage of records across different

datasets.

6 CONCLUSION

In this paper, we presented a detailed study on the

effectiveness of a binary entity resolution (ER)

approach using a cascading method. The primary

objective was to accurately link records between

datasets with varying levels of data completeness and

complexity. The datasets, referred to as Easy and

Medium, were subjected to a series of cascades with

progressively relaxed matching criteria, allowing for

a thorough evaluation of the system's performance

across key metrics, including precision, recall, and F-

measure. The results demonstrated that the cascading

method is highly effective in resolving entities, even

in complex data scenarios. For the Easy dataset, the

system achieved perfect precision and progressively

improved recall, culminating in flawless performance

by the fifth cascade. Similarly, the medium dataset,

characterized by incomplete data and duplicates,

showed significant improvements in recall across the

cascades, ultimately reaching perfect precision and

recall in the final stages. This indicates the method's

robustness and adaptability to various data

challenges. The consistent high precision observed

from the first cascade onward highlights the method's

strength in minimizing false positives, ensuring that

identified matches are accurate. The gradual increase

in recall reflects the system's ability to expand its

matching criteria to capture all relevant entities

comprehensively. The cascading approach

effectively balances the need for high precision with

the necessity of thorough recall, making it an ideal

solution for complex ER tasks.

In conclusion, the cascading method proves to be

a robust and efficient approach for entity resolution,

capable of handling datasets with diverse

characteristics and complexities. Its systematic

relaxation of matching criteria ensures high accuracy

and completeness in linking records, making it a

valuable tool in data integration and analysis. Future

work may explore further optimizations in the

cascade design and extend the approach to more

diverse data domains, potentially incorporating

machine learning techniques to enhance matching

precision and recall.

ACKNOWLEDGEMENTS

This research was partially supported by the

Department of Commerce US Census Cooperative

Agreement CB21RMD0160002. We express our

profound gratitude to Dr. John R. Talburt, the Project

Coordinator. His invaluable guidance and support

have been instrumental in the progression and success

of this project. We also extend our thanks to Syed

Yaser Mehdi for his comprehensive research on the

literature review which has significantly enriched our

understanding and findings. We are deeply grateful to

the University of Arkansas at Little Rock for

providing us with the necessary resources and an

inspiring platform to conduct our research. Their

commitment to fostering an environment of academic

excellence and innovation has been crucial to our

work.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

REFERENCES

Fellegi, I. P., & Sunter, A. B. (1969). A Theory for Record

Linkage. Journal of the American Statistical

Association, 64(328), 1183-1210.

Winkler, W. E. (1990). String Comparator Metrics and

Enhanced Decision Rules in the Fellegi-Sunter Model

of Record Linkage. Proceedings of the Section on

Survey Research Methods, American Statistical

Association, 354-359.

Christen, P. (2012). Data Matching: Concepts and

Techniques for Record Linkage, Entity Resolution, and

Duplicate Detection. Springer.

Cohen, W. W., Ravikumar, P., & Fienberg, S. E. (2003). A

Comparison of String Distance Metrics for Name-

Matching Tasks. In IIWeb (Vol. 3, pp. 73-78).

Papadakis, G., Palpanas, T., & Koutrika, G. (2020). Entity

Resolution Methods for Big Data. ACM Computing

Surveys (CSUR), 53(1), 1-42.

Talburt, J. R. (2011). Entity Resolution and Information

Quality. Morgan Kaufmann.

Papadakis, G., Kirielle, N., & Palpanas, T. (2024). A

Critical Re-evaluation of Record Linkage Benchmarks

for Learning-Based Matching Algorithms. Proceedings

of the 40th International Conference on Data

Engineering (ICDE), 3435-3448.

Christen, P., Vatsalan, D., & Verykios, V. S. (2014).

Challenges for Privacy-Preserving Record Linkage.

IEEE Transactions on Knowledge and Data

Engineering, 26(4), 912-925.

Mohammed, O.K. et al. (2024). Household Discovery with

Group Membership Graphs. In: Latifi, S. (eds) ITNG

2024: 21st International Conference on Information

Technology-New Generations. ITNG 2024. Advances

in Intelligent Systems and Computing, vol 1456.

Springer, Cham. https://doi.org/10.1007/978-3-031-

56599-1_31

A. Cleven and F. Wortmann, "Uncovering Four Strategies

to Approach Master Data Management," Methods,

American Statistical Association, 354-359.2010 43rd

Hawaii International Conference on System Sciences,

Honolulu, HI, USA, 2010, pp. 1-10, doi:

10.1109/HICSS.2010.488.

Winkler, W. E. (1990). String Comparator Metrics and

Enhanced Decision Rules in the Fellegi- Sunter Model

of Record Linkage. Proceedings of the Section on

Survey Research.

Improving Quality of Entity Resolution Using a Cascade Approach