Dedupe - How it works

Using advanced machine learning and statistics, learns the best way to identify similar records in any dataset. Learn the specifics of our research-driven approach to record matching and entity resolution.

Matching records

If you look at the following two records, you might think it’s pretty clear that they are about the same person.

first name last name address phone
bob roberts 1600 pennsylvania ave. 555-0123
Robert Roberts 1600 Pensylvannia Avenue

However, I bet it would be pretty hard for you to explicitly write down all the reasons why you think these records are about the same Mr. Roberts.

Record similarity

One way that people have approached this problem is by saying that records that are more similar are more likely to be duplicates. That’s a good first step, but then we have to precisely define what we mean for two records to be similar.

The default way that we do this in is to use what’s called a string metric. A string metric is an way of taking two strings and returning a number that is low if the strings are similar and high if they are dissimilar.

One famous string metric is called the Hamming distance. It counts the number of substitutions that must be made to turn one string into another. For example, roberts and Roberts would have Hamming distance of 1 because we have to substitute r for R in order to turn roberts into Roberts.

There are lots of different string metrics, and we actually use a metric called the Affine Gap Distance, which is a variation on the Hamming distance.

Record by record or field by field

When we are calculating whether two records are similar we could treat each record as if it was a long block of text.

Record A Record B
bob roberts 1600 pennsylvania ave. 555-0123 Robert Roberts 1600 Pensylvannia Avenue

Alternately, we could compare field by field:

Record A Record B
first name bob Robert
last name roberts Roberts
address 1600 pennsylvania ave. 1600 Pensylvannia Avenue
phone 555-0123

The major advantage of comparing field by field is that we don’t have to treat each field string distance equally. Maybe we think that its more important that the first name and phone numbers are close over the other fields. We can express that importance with numeric weights, i.e.

Record A Record B Weight
first name bob Robert x 0.5
last name roberts Roberts x 0.2
address 1600 pennsylvania ave. 1600 Pensylvannia Avenue x 0.2
phone 555-0123 x 0.5

Setting weights and making decisions

Let’s say we calculated the record distance and we found that it had a value of 8. That number, by itself, is not that helpful.

Ultimately, we are trying to decide whether a pair of records are duplicates. Does an 8 mean that the pair of records are really similar or really far apart, likely or unlikely to be duplicates? We’d like to define the record distances so that we can look at the number and know whether to decide whether it’s a duplicate.

Also, we really would rather not have to set the weights manually every time and it can be very tricky to know which fields are going to matter. Even if we know that some fields are more important than others, how do we quantify that? Is it 2 times more important or 1.3 times?

Fortunately, we can solve both problems with a technique called regularized logistic regression. If we supply pairs of records that we label as either being duplicates or distinct, then will learn a set of weights such that the record distance can easily be transformed into our best estimate of the probability that a pair of records are duplicates.

Once we have learned these good weights, we want to use them to find which records are duplicates. However, it turns out that doing this the naive way will usually not work, and we’ll have to do something smarter to make the right comparisons.

Active learning

In order to learn the field weights, needs example pairs with labels. Most of the time, we will need people to supply those labels. To reduce the amount of time that labeling takes, we use an approach called active learning.

With active learning, keeps track of unlabeled pairs and their currently learned weights. At any time, there will be a record pair will believe have a near a 50% chance of being a duplicate or distinct. By always asking you to label the record pair is least certain about, we will learn the most we possibly can about your dataset for each training pair.

Once a record pair is labeled, automatically relearns the field weights. With these new weights, there will now be a different record pair that is most uncertain about, and that’s the next one you’ll be asked to label.

Making smart comparisons

Say we have magic tool that can compare two records and automatically know if they are matches or not. Let’s say that this tool takes took one second to return.

How long would it take to duplicate a thousand records?

Within a dataset of thousand records, there are (1,000 x 999) / 2 = 499,500 unique pairs of records. If we compared all of them using our magic tool it would take six days.

In the computer world, though, one second is a long time. Let’s say we sped it up so that we can make 10,000 comparisons per second. Now we can get through our 1,000 records in less than a minute.

Feeling good about our super-fast comparison tool, let’s take on a dataset of 100,000 records.

Now there are (100,000 x 99,999) / 2 = 4,999,950,000 unique possible pairs. If we compare all of them with our super-fast comparison tool, it will take six days again.

If we want to work with moderately sized data, we have to find a way of making fewer comparisons.

Duplicates are rare

In real world data, nearly all possible pairs of records are not duplicates.

In the example below, two pairs of records are duplicates—(1, 2) and (3, 4), while there are four unique pairs of records that are not duplicates—(1,3), (1,4), (2,3), and (2,4).

Typically, as the size of the dataset grows, the fraction of pairs of records that are duplicates gets very small very quickly.

record id first name last name address phone
1 bob roberts 1600 pennsylvania ave. 555-0123
2 Robert Roberts 1600 Pensylvannia Avenue
3 steve Jones 123 Cowabunga Lane 555-0000
4 Stephen Jones 123 Cawabunga Ln 444-555-0000

If we could only compare records that were true duplicates we would not run into the explosion of comparisons. Of course if we already knew where the true duplicates were, we wouldn’t need to compare any records. Unfortunately we do quite well if we just compare records that are somewhat similar.


Duplicate records almost always share some thing in common. If we define groups of data that share some thing and only compare the records in that group, or block, then we can dramatically reduce the number of comparisons we will make. If define these blocks well, then we will make very few comparisons and still have confidence that will compare records that truly are duplicates.

This task is called blocking, and we approach it in two ways: predicate blocks and canopies.

Predicate blocks

A predicate block is a bundle of records that all share a feature - a feature produced by a simple function called a predicate.

Predicate functions take in a record field, and output a set of features for that field. These features could be “the first 3 characters of the field,” “every word in the field,” and so on. Records that share the same feature become part of a block.

Let’s take an example. Let’s use a “first 3 character” predicate on the address field below.

record id first name last name address phone
1 bob roberts 1600 pennsylvania ave. 555-0123
2 Robert Roberts 1600 Pensylvannia Avenue
3 steve Jones 123 Cowabunga Lane 555-0000
4 Stephen Jones 123 Cawabunga Ln 444-555-0000

That leaves us with two blocks - The 160 block, which contains records 1 and 2, and the 123 block, which contains records 3 and 4.

Again, we’re applying the “first three words” predicate function to the address field in our data, the function outputs the following features - 160, 160, 123, 123 - and then we group together the records that have identical features into “blocks”.

Others simple predicates uses include:

  • whole field
  • token field
  • common integer
  • same three char start
  • same five char start
  • same seven char start
  • near integers
  • common four gram
  • common six gram

Index Blocks also uses another way of producing blocks from searching and index. First, we create a special data structure, like an inverted index, that lets us quickly find records similar to target records. We populate the index with all the unique values that appear in field.

For each record we search the index for values similar to the record’s field. We block together records that share at least one common search result.

Index predicates require building an index from all the unique values in a field. This can take substantial computation time and memory. Index predicates are also usually slower than predicate blocking.

Combining blocking rules

If it’s good to define blocks of records that share the same city field, it might be even better to block records that share BOTH the city field AND zip code field. tries these cross-field blocks. These combinations blocks are called disjunctive blocks.

Learning good blocking rules for given data uses a long set of predicates, and when these are combined can have hundreds of possible blocking rules to choose from. We will want to find a small set of these rules that covers every labeled duplicated pair but minimizes the total number pairs will have to compare.

While we approach this problem by using greedy algorithms, particularly Chvatal’s Greedy Set-Cover algorithm.

Grouping duplicates

Once we have calculated the probability that pairs of record are duplicates or not, we need to transform pairs of duplicate records into clusters of duplicate records. Three, ten, or even thousands of records could all refer to the same entity (person, organization, ice cream flavor, etc.,) but we only have pairwise measures.

Let’s say we have measured the following pairwise probabilities between records A, B, and C.

A -- 0.6 -- B -- 0.6 -- C

The probability that A and B are duplicates is 60%, the probability that B and C are duplicates is 60%, but what is the probability that A and C are duplicates?

Let’s say that everything is going perfectly and we can say there’s a 36% probability (60% x 60%) that A and C are duplicates. We’d probably want to say that A and C should not be considered duplicates.

Okay, then should we say that A and B are a duplicate pair and C is a distinct record or that A is the distinct record and that B and C are duplicates?

Well… this is a thorny problem, and we tried solving it a few different ways. In the end, we found that hierarchical clustering with centroid linkage gave us the best results. This algorithm treats all points within some distance of centroid as part of the same group. In this example, B would be the centroid - and A, B, and C would all be put in the same group.

Unfortunately, a more principled answer does not exist because the estimated pairwise probabilities are not transitive.

Clustering the groups depends on us setting a threshold for group membership—the distance of the points to the centroid. Depending on how we choose that threshold, we’ll get very different groups, and we will want to choose this threshold wisely.

Choosing a good threshold can predict the probability that a pair of records are duplicates. So, how should we decide that a pair of records really are duplicates? The answer lies in the tradeoff between precision and recall.

As long as we know how much we care about precision vs. recall, we can define an F-score that will let us find a threshold for deciding when records are duplicates that is optimal for our priorities.

Typically, the way that we find that threshold is by looking at the precision and recall of some data where we know the real duplicates. However, we will only get a good threshold if the labeled examples are representative of the data we are trying to classify.

So here’s the problem - the labeled examples that we make with are not at all representative, and that’s by design. In the active learning step, we are not trying to find the most representative data examples. We’re trying to find the ones that will teach us the most.

The approach here is to take a random sample of blocked data, and then calculate the pairwise probability that records will be duplicates within each block. From these probabilities we can calculate the expected number of duplicates and distinct pairs, so we can calculate the expected precision and recall.

Special cases

The process we have been describing is for the most general case—when you have a dataset where an arbitrary number of records can all refer to the same entity.

There are certain special cases where we can make more assumptions about how records can be linked, which if true, make the problem much simpler.

Record linkage

One important case is record linkage: finding matches across two different datasets. By assuming that each dataset, individually, is unique, then this puts a nice, big constraint on how records can match. If this uniqueness assumption holds, then two records can only refer to the same entity if they are from different datasets and no other record can match either of those two records.