Matching evaluation documentation

MATCHFORGE PLAYGROUND

Matching evaluation

Generate matching predictions, compare them with verified ground-truth pairs, and evaluate performance using classification metrics.

Understanding this demo

Imagine merging customer records from two different systems. The same customer may have slightly different names, while different customers may even share identical names.

System A

Original customer records

A-1002 · Amit Patel

System B

Records from another database

B-2002 · Amit Patel

B-2003 · Amit R Patel

Matching uses names only. Customer IDs establish the actual identity of records and are used only for ground-truth labels.

01

Matching predictions

Generate predicted pairs using matchRecords().

02

Ground truth

Provide verified match and non-match labels.

03

Evaluation

Calculate metrics using evaluateMatches().

01

Matching configuration

Configure how MatchForge generates its predictions.

80%
0 · More candidates1 · Stricter matching
02

Input datasets

Edit the customer names to experiment with different matching scenarios.

Source dataset

5 records · name

0

A-1001

1

A-1002

2

A-1003

3

A-1004

4

A-1005

Target dataset

7 records · fullName

0

B-2001

1

B-2002

2

B-2003

3

B-2004

4

B-2005

5

B-2006

6

B-2007

Customer IDs are reference information only. They are not passed to the matching algorithm.

03

Ground-truth labels

A small set of manually verified pairs used to evaluate matching predictions.

8

Labelled pairs

4

Actual matches

4

Non-matches

Ready to evaluate?

Generate predictions and compare them with your labels.