Match Identity

Welcome to Match Studio! This is an interactive tool for evaluating and configuring Name Match (Match) for record matching. Match Studio uses Match for fuzzy name retrieval and name matching. It includes a multitude of configurable parameters that can be used to optimize the tool for a given use case.

Matching refers to the process of comparing identifying information about an individual, such as their name, company, address, and/or age, between two records. With Match Studio, you can enter one or more pieces of identifying information and Match Studio will return a list of potential matches from your loaded index. Each match will have a score, between 0% and 100%, indicating the match strength.

Name matching is the core of multi-field entity matching. Names are complex to match because of the large number of variations that occur within a language and across languages. These include, but are not limited to, typographical errors, phonetic spelling variations, transliteration differences, initials, and nicknames.

Over the course of this tutorial, you will learn how to import an index, search for names within it, edit search and match configurations, perform evaluations, and more. This quick-start guide will focus only on name matching. However, many of the concepts we will discuss can be applied to other data types as well. For more information on matching other fields, see the Match Studio Help.

Documentation

At any point while using Match Studio, you can hover over the question mark icon on the navigation bar and select Documentation to open the user guide.

Before we start matching, we need to create an index to match against. An index is a list of records, each entry containing one or more fields. In this guide, we will be creating an index of names and related identifiers. Later on, we will search this index for names and examine how Match Studio scores matches between index entries and queries.

Creating an index involves uploading structured data and mapping record fields to it to tell Match Studio what kind of data the index contains. While this tutorial will only cover one index, Match Studio may contain multiple indices. You can switch between indices to search different bodies of data.

The data we are about to import is a list of names in a .csv file with a header. More complete indices may include additional columns for data points such as address or age. For the purposes of this quick-start guide, we will focus only on names.

Now, it's time to start mapping.

Mapping is the process of assigning data types, or fields, to the columns in your dataset. Each column must have a field type assigned to it.

While the included file contains both person names and organization names, we will only be focusing on person names in this tutorial. Other possible field types include location name, date, address, and any other data type that is valid in Match.

Now that we have created the index, we can perform a search. Search returns all matching records from the index for a single query. This is useful for situations in which you want to search a single name against an existing index.

After completing the above steps, scroll down to view a list of potential name matches from the index in descending order by match score. This match score is an indication of how similar two names are and ranges from 0% to 100%. The top result is Phillip Ward with a match score of 100%. (A match score is a percentage indicating how similar two names are, with a higher percentage indicating a closer match. For more information, see Understanding name match scores.) The match score is highlighted in green to indicate that it is a good match. Any match with a score above the match threshold, a configurable value which is set to 70% by default, is highlighted in this manner.

The next highest match score is 44% for Felipe Viana. This can be attributed to the fact that while Felipe and Phillip are similar, Viana and Ward are not. The score is not high enough to be considered a match in our current configuration.

Let's explore what happens when we search for a company name as well.

Now, the match score for the Phillip Ward result should drop to 87%. Let's find out why.

A match score is displayed beneath each field. This is because each field is scored separately. While the person names in this search have a match score of 100%, the business score is only 75%. Each field also has a configurable weight, but the default is to divide the weight evenly among all searched fields. To determine the final match score, Match Studio calculates the sum of all field scores multiplied by their corresponding field weights. In this case: (75% * 50%) + (100% * 50%) = 87%.

The Compare function allows you to see more details about the match score of two values with the same field type. In this case, we will compare a search value with a match value.

The first step of calculating a name match score is tokenization, in which each name is broken up into smaller pieces called tokens. The score computation table separates each name into tokens and calculates how much each token from one name matches each token from the other name.

The tokens from the left name are listed down the first column, while the tokens from the right name are along the top row.

The shaded boxes highlight the token pairs selected during matching that produce the best score. A token pair is a token from the left name and its matching token from the right name.

If you scroll further down, you will see each token's weight. The weightings determine how important the token pair match is in calculating the final score.  Unusual tokens get a higher weighting than common names because their match is more significant, and initials are weighted less than full names.  

Let’s see what happens when we misspell one of the tokens.

Now let’s examine how the following have changed:

The final score for these names is 87.4%. While not an exact match, this is still considered a match in the current configuration, as indicated by the green color.

What happens if one of the names is out of order?

Now let’s examine how the following have changed:

The final score for these names is 91.5%.

Next, let’s give Match Studio a bigger challenge by misspelling the first and last name, placing them out of order, and adding an initial.

Now let’s examine how the following have changed:

The final score for these names is 66.9%. This is not high enough to be considered a match in our current configuration (note how the circle surrounding the score is no longer green), but it is close. What if we want Match Studio to consider these two names a match? We will look at a few ways to accomplish this next.

You should always configure Match Studio to suit the goals of any given search. This might mean minimizing false negatives at the expense of increased false positives, or vice versa. A false negative does not match with the searched name when it should. A false positive matches the searched name when it should not.

Whether two names match or not depends on the match threshold, a configurable percentage at or above which two names are considered a match. The default match threshold is 70%. With this threshold, two names with a match score of 60% would not be considered a match, but two names with a match score of 80% would be. A match threshold that is too high increases the risk of false negatives, while a match threshold that is too low increases the risk of false positives.

The score of 62% is below the default threshold of 70%, but it isn't hard to see why someone might want these names to match. The names are spelled slightly differently and out of order, but phonetically they are very similar. We could easily make them match by changing the match threshold.

The problem with decreasing the match threshold to force a match between two names is that it increases your risk of false positives without actually tuning Match Studio to better suit your search goals. Later in the tutorial, we will look at how evaluations can help us determine the ideal match threshold. For now, let's reset the match threshold to the default value.

The Search page also has options to edit the display threshold and choose a match configuration. The display threshold's function is very similar to that of the match threshold, but instead of determining whether a name matches for a searched name, it determines whether a name appears in the list of results for a searched name. We will learn about match configurations later in the tutorial.

Evaluations allow you to see how Match Studio's results for a given index change with different match configurations and thresholds. It does this by comparing its own search results to user-provided gold data. You can think of gold data as the correct answers to a test that Match Studio is administering to itself. The gold data file consists of a list of name pairs and whether or not they should match. For the purposes of this tutorial, we have provided you with a pre-made gold data file for use in your first evaluation.

When the Evaluation is complete, you will see a table showing the match configuration used to perform the evaluation, date of the evaluation, and the following additional information:

From this evaluation, we can see that our current threshold of 70% gives us very good results for precision, recall, and F1 measure.

Select Threshold Report in the Options column of the evaluation to open the threshold report. This is a graph displaying how precision, recall, and F1 measure change as the match threshold increases or decreases.

Earlier in the tutorial, we decreased the match threshold to 0.62 to force a match between the names "Kristy Hart" and "Heart Kirsten". Mouse over the graph at a match threshold of 0.62 to see the performance at that threshold. You will see that while we have perfect recall (no false negatives) at this threshold, precision has significantly decreased (more false positives), resulting in an overall lower F1 measure.

When false negatives (missed matches) are troublesome or even dangerous, select a threshold that favors recall. But imagine we are using Match Studio to match medical records; in this case, we don't want to have so many extra matches (false positives) that it is impossible to find the right one (true positive). This situation could result in someone receiving incorrect medical care. With that in mind, let's take another approach to tune our model for this scenario.

Name order, initialisms, and misspellings are all factors that can affect the match score between two versions of a name. You can edit various match parameters to control how much factors like these affect the match score of a given pair of names. The Compare page has additional features that will allow us to test out different match parameters.

Now that we have identified which parameter to edit in order to get our test names to match, we can use that information to create a new match configuration. Match configurations allow you to save specific combinations of parameter values, allowing you to easily swap or compare them on the fly.

Now, let's combine what we've learned about evaluations and match configurations to evaluate the performance of our new match configuration.

Looking at the evaluation table, you will see that precision, recall, and F1 are unchanged. However, both precision and F1 are higher at our 0.7 match threshold than they were for the default configuration at a 0.62 match threshold. By creating a new match configuration, we have managed to make Match Studio match our two test names without becoming overly vulnerable to false positives.

Figure 2. Old Match Configuration at 62% Threshold

Figure 3. New Match Configuration at 70% Threshold

Now that we have created a new match configuration and evaluated it to determine the best threshold, we can apply it to the search configuration to see its effects in practice.

The search now has a 70.1% match score with Kristy Hart, making them a match with our original match threshold.

You can also export any match configurations you create for use with any version of Name Match or other instances of Match Studio.

To import the match configuration file into the Match plugin, first rename the file to parameter_profiles.yaml. Then replace the existing parameter_profiles.yaml file in the plugin with the new version. After restarting, the parameters in the new file will take effect.

Congratulations! You have just learned how to search with Match Studio, run an evaluation on your data, adjust parameters, and save a modified match configuration.

This was a simplified test project for the purpose of familiarizing you with the uses and general workflow of Match Studio. When using Match Studio for real applications, be prepared to work with larger indices, more fields, and even multiple languages.

For more information on Match Studio, see our online Match Studio documentation.