The R package fuzzylink implements a probabilistic record linkage procedure proposed in Ornstein (2024). This method allows users to merge datasets with fuzzy matches on a key identifying variable. Suppose, for example, you have the following two datasets:
dfA
#> name age
#> 1 Joe Biden 81
#> 2 Donald Trump 77
#> 3 Barack Obama 62
#> 4 George W. Bush 77
#> 5 Bill Clinton 77
dfB
#> name hobby
#> 1 Joseph Robinette Biden Football
#> 2 Donald John Trump Golf
#> 3 Barack Hussein Obama Basketball
#> 4 George Walker Bush Reading
#> 5 William Jefferson Clinton Saxophone
#> 6 George Herbert Walker Bush Skydiving
#> 7 Biff Tannen Bullying
#> 8 Joe Riley JoggingWe would like a procedure that correctly identifies which records in dfB are likely matches for each record in dfA. The fuzzylink() function performs this record linkage with a single line of code.
library(fuzzylink)
df <- fuzzylink(dfA, dfB, by = 'name', record_type = 'person')
df
#> A B sim jw
#> 1 Joe Biden Joseph Robinette Biden 0.7660045 0.7208273
#> 2 Donald Trump Donald John Trump 0.8388933 0.9333333
#> 3 Barack Obama Barack Hussein Obama 0.8457593 0.9200000
#> 4 George W. Bush George Walker Bush 0.8446479 0.9301587
#> 5 Bill Clinton William Jefferson Clinton 0.8732562 0.5788889
#> match_probability validated age hobby
#> 1 1 Yes 81 Football
#> 2 1 Yes 77 Golf
#> 3 1 Yes 62 Basketball
#> 4 1 Yes 77 Reading
#> 5 1 Yes 77 SaxophoneThe procedure works by using pretrained text embeddings to construct a measure of similarity for each pair of names. These similarity measures are then used as predictors in a statistical model to estimate the probability that two name pairs represent the same entity. In the guide below, I will walk step-by-step through what’s happening under the hood when we call the fuzzylink() function. See Ornstein (2024) for technical details.
Installation
You can install the development version of fuzzylink from GitHub with:
# install.packages("devtools")
devtools::install_github("joeornstein/fuzzylink")You will also need API access to a large language model (LLM). The fuzzylink package currently supports both OpenAI and Mistral LLMs, but will default to using OpenAI unless specified by the user.
OpenAI
You can sign up for an account here, after which you will need to create an API key here. For best performance, I strongly recommend purchasing at least $5 in API credits, which will significantly increase your API rate limits.
Once your account is created, copy-paste your API key into the following line of R code.
Mistral
If you prefer to use language models from Mistral, you can sign up for an account here. As of writing, Mistral requires you to purchase prepaid credits before you can access their language models through the API.
Once you have a paid account, you can create an API key here, and copy-paste the API key into the following line of R code:
Now you’re all set up!
Example
Here is some code to reproduce the example above and make sure that everything is working on your computer.
library(tidyverse)
library(fuzzylink)
dfA <- tribble(~name, ~age,
'Joe Biden', 81,
'Donald Trump', 77,
'Barack Obama', 62,
'George W. Bush', 77,
'Bill Clinton', 77)
dfB <- tribble(~name, ~hobby,
'Joseph Robinette Biden', 'Football',
'Donald John Trump ', 'Golf',
'Barack Hussein Obama', 'Basketball',
'George Walker Bush', 'Reading',
'William Jefferson Clinton', 'Saxophone',
'George Herbert Walker Bush', 'Skydiving',
'Biff Tannen', 'Bullying',
'Joe Riley', 'Jogging')
df <- fuzzylink(dfA, dfB, by = 'name', record_type = 'person')
dfIf the df object links all the presidents to their correct name in dfB, everything is running smoothly! (Note that you may see a warning from glm.fit. This is normal. The stats package gets suspicious whenever the model fit is too perfect.)
Arguments
The
byargument specifies the name of the fuzzy matching variable that you want to use to link records. The dataframesdfAanddfBmust both have a column with this name.The
record_typeargument should be a singular noun describing the type of entity thebyvariable represents (e.g. “person”, “organization”, “interest group”, “city”). It is used as part of a language model prompt when training the statistical model (see Step 3 below).The
instructionsargument should be a string containing additional instructions to include in the language model prompt. Format these like you would format instructions to a human research assistant, including any relevant information that you think would help the model make accurate classifications.The
modelargument specifies which language model to prompt. It defaults to OpenAI’s ‘gpt-3.5-turbo-instruct’, but for more difficult problems you can try ‘gpt-4o’. Note that this will typically increase accuracy at the expense of cost and runtime. If you prefer an open-source language model, try ‘open-mixtral-8x22b’.The
embedding_modelargument specifies which pretrained text embeddings to use when modeling match probability. It defaults to OpenAI’s ‘text-embedding-3-large’, but will also accept ‘text-embedding-3-small’ or Mistral’s ‘mistral-embed’.Several parameters—including
p,k,embedding_dimensions,max_validations, andparallel—are for advanced users who wish to customize the behavior of the algorithm. See the package documentation for more details.If there are any variables that must match exactly in order to link two records, you will want to include them in the
blocking.variablesargument. As a practical matter, I strongly recommend including blocking variables wherever possible, as they reduce the time and cost necessary to compute pairwise distance metrics. Suppose, for example, that our two illustrative datasets have a column calledstate, and we want to instructfuzzylink()to only link people who live within the same state.
dfA <- tribble(~name, ~state, ~age,
'Joe Biden', 'Delaware', 81,
'Donald Trump', 'New York', 77,
'Barack Obama', 'Illinois', 62,
'George W. Bush', 'Texas', 77,
'Bill Clinton', 'Arkansas', 77)
dfB <- tribble(~name, ~state, ~hobby,
'Joseph Robinette Biden', 'Delaware', 'Football',
'Donald John Trump ', 'Florida', 'Golf',
'Barack Hussein Obama', 'Illinois', 'Basketball',
'George Walker Bush', 'Texas', 'Reading',
'William Jefferson Clinton', 'Arkansas', 'Saxophone',
'George Herbert Walker Bush', 'Texas', 'Skydiving',
'Biff Tannen', 'California', 'Bullying',
'Joe Riley', 'South Carolina', 'Jogging')
df <- fuzzylink(dfA, dfB,
by = 'name',
blocking.variables = 'state',
record_type = 'person')
df#> A B sim block jw
#> 1 Joe Biden Joseph Robinette Biden 0.7665511 1 0.7208273
#> 2 Barack Obama Barack Hussein Obama 0.8458046 3 0.9200000
#> 3 George W. Bush George Walker Bush 0.8447483 4 0.9301587
#> 4 Bill Clinton William Jefferson Clinton 0.8731902 5 0.5788889
#> 5 Donald Trump <NA> NA NA NA
#> match_probability state validated age hobby
#> 1 1 Delaware Yes 81 Football
#> 2 1 Illinois Yes 62 Basketball
#> 3 1 Texas Yes 77 Reading
#> 4 1 Arkansas Yes 77 Saxophone
#> 5 NA New York <NA> 77 <NA>Note that because Donald Trump is listed under two different states—New York in dfA and Florida in dfB–the fuzzylink() function no longer returns a match for this record; all blocking variables must match exactly before the function will link two records together. You can specify as many blocking variables as needed by inputting their column names as a vector.
The function returns a few additional columns along with the merged dataframe. The column match_probability reports the model’s estimated probability that the pair of records refer to the same entity. This column should be used to aid in validation and can be used for computing weighted averages if a record in dfA is matched to multiple records in dFB. The columns sim and jw are string distance measures that the model uses to predict whether two records are a match. And if you included blocking.variables in the function call, there will be a column called block with an ID variable denoting which block the records belong to.
Under The Hood
If you’d like to know more details about about how fuzzylink() works, you can read the accompanying research paper. In this section, we’ll take a look under the hood at the previous example, walking through each of the steps that fuzzylink() takes to join the two dataframes.
Step 1: Embedding
First, the function encodes each unique string in dfA and dfB as a 256-dimensional vector called an embedding. You can learn more about embeddings here, but the basic idea is to represent text using a vector of real-valued numbers, such that two vectors close to one another in space have similar meanings.
library(tidyverse)
strings_A <- unique(dfA$name)
strings_B <- unique(dfB$name)
all_strings <- unique( c(strings_A, strings_B) )
embeddings <- get_embeddings(all_strings)
dim(embeddings)
#> [1] 13 256
head(embeddings['Bill Clinton',])
#> [1] 0.08031169 0.07614738 -0.01617801 -0.07958458 -0.09815873 -0.04967427Step 2: Similarity Scores
Next, we compute the cosine similarity between each name pair. This is our measure of how closely related two pieces of text are, where 0 is completely unrelated and 1 is identical. If you include blocking.variables in the call to fuzzylink(), the function will only consider within-block name pairs (i.e. it will only compute similarity scores for records with an exact match on each blocking variable). I strongly recommend blocking wherever possible, as it significantly reduces cost and speeds up computation.
sim <- get_similarity_matrix(embeddings, strings_A, strings_B)
sim
#> Joseph Robinette Biden Donald John Trump Barack Hussein Obama
#> Joe Biden 0.7665511 0.5532816 0.5309110
#> Donald Trump 0.4316448 0.8388565 0.4477471
#> Barack Obama 0.5171871 0.4756583 0.8458046
#> George W. Bush 0.4942466 0.4877767 0.5681277
#> Bill Clinton 0.4886684 0.5037386 0.5174360
#> George Walker Bush William Jefferson Clinton
#> Joe Biden 0.5094280 0.5426318
#> Donald Trump 0.4805295 0.4462709
#> Barack Obama 0.4854634 0.5131122
#> George W. Bush 0.8447483 0.6113368
#> Bill Clinton 0.6232500 0.8731912
#> George Herbert Walker Bush Biff Tannen Joe Riley
#> Joe Biden 0.4701124 0.3016349 0.3906386
#> Donald Trump 0.3942993 0.3438548 0.2328768
#> Barack Obama 0.4243879 0.2546999 0.3480991
#> George W. Bush 0.7335260 0.2459214 0.3606344
#> Bill Clinton 0.5950811 0.2214671 0.3194544Step 3: Create a Training Set
We would like to use those cosine similarity scores to predict whether two names refer to the same entity. In order to do that, we need to first create a labeled dataset to fit a statistical model. The get_training_set() function selects a sample of name pairs and labels them using the following prompt to GPT-3.5 (brackets denote input variables).
Decide if the following two names refer to the same {record_type}. {instructions} Think carefully. Respond "Yes" or "No".'
Name A: {A}
Name B: {B}
Response:
train <- get_training_set(list(sim), record_type = 'person')
train
#> # A tibble: 40 × 5
#> A B sim jw match
#> <fct> <fct> <dbl> <dbl> <chr>
#> 1 Donald Trump "Biff Tannen" 0.344 0.399 No
#> 2 Donald Trump "William Jefferson Clinton" 0.446 0.372 No
#> 3 Bill Clinton "George Herbert Walker Bush" 0.595 0.371 No
#> 4 Donald Trump "George Walker Bush" 0.481 0.5 No
#> 5 George W. Bush "Joe Riley" 0.361 0.410 No
#> 6 George W. Bush "George Herbert Walker Bush" 0.734 0.870 No
#> 7 Bill Clinton "William Jefferson Clinton" 0.873 0.579 Yes
#> 8 Joe Biden "William Jefferson Clinton" 0.543 0.524 No
#> 9 Bill Clinton "Donald John Trump " 0.504 0.465 No
#> 10 Donald Trump "George Herbert Walker Bush" 0.394 0.344 No
#> # ℹ 30 more rowsStep 4: Fit Model
Next, we fit a logistic regression model on the train dataset, so that we can map similarity scores onto a probability that two records match. We use both the cosine similarity (sim) and a measure of lexical similarity (jw) as predictors in this model.
model <- glm(as.numeric(match == 'Yes') ~ sim + jw,
data = train,
family = 'binomial')Append these predictions to each name pair in dfA and dfB.
# create a dataframe with each name pair
df <- sim |>
reshape2::melt() |>
set_names(c('A', 'B', 'sim')) |>
# compute lexical similarity measures for each name pair
mutate(jw = stringdist::stringsim(A, B, method = 'jw', p = 0.1))
df$match_probability <- predict(model, df, type = 'response')
head(df)
#> A B sim jw match_probability
#> 1 Joe Biden Joseph Robinette Biden 0.7665511 0.7208273 1.000000e+00
#> 2 Donald Trump Joseph Robinette Biden 0.4316448 0.4217172 2.220446e-16
#> 3 Barack Obama Joseph Robinette Biden 0.5171871 0.4191919 2.220446e-16
#> 4 George W. Bush Joseph Robinette Biden 0.4942466 0.5200216 2.220446e-16
#> 5 Bill Clinton Joseph Robinette Biden 0.4886684 0.4797980 2.220446e-16
#> 6 Joe Biden Donald John Trump 0.5532816 0.4444444 2.220446e-16Step 5: Validate Uncertain Matches
We now have a dataset with estimated match probabilities for each pair of records in dfA and dfB. We could stop there and just report the match probabilities. But for larger datasets we can get better results if we conduct a final validation step. For each name pair within a range of estimated match probabilities (by default 0.1 to 0.95), we will use the GPT-4 prompt above to check whether the name pair is a match. These labeled pairs are then added to the training dataset, the logistic regression model is refined, and we repeat this process until there are no matches left to validate. At that point, every record in dfA is either linked to a record in dfB or there are no candidate matches in dfB with an estimated probability higher than the threshold.
Note that, by default, the fuzzylink() function will validate at most 100,000 name pairs during this step. This setting reduces both cost and runtime (see “A Note On Cost” below), but users who wish to validate more name pairs within larger datasets can increase the cap using the max_validations argument.
# find all unlabeled name pairs within a range of match probabilities
matches_to_validate <- df |>
left_join(train, by = c('A', 'B', 'sim')) |>
filter(match_probability > 0.1,
match_probability < 0.95,
is.na(match))
while(nrow(matches_to_validate) > 0){
# validate matches using LLM prompt
matches_to_validate$match <- check_match(matches_to_validate$A,
matches_to_validate$B)
# append new labeled pairs to the train set
train <- train |>
bind_rows(matches_to_validate |>
select(A,B,sim,match))
# refine the model
model <- glm(as.numeric(match == 'Yes') ~ sim + jw,
data = train,
family = 'binomial')
# re-estimate match probabilities
df$match_probability <- predict(model, df, type = 'response')
# find all unlabeled name pairs within a range of match probabilities
matches_to_validate <- df |>
left_join(train, by = c('A', 'B', 'sim')) |>
filter(match_probability > 0.1,
match_probability < 0.95,
is.na(match))
}Step 6: Link Datasets
Finally, we take all name pairs whose match probability is higher than a user-specified threshold and merge them into a single dataset.
matches <- df |>
# join with match labels from the training set
left_join(train |> select(A, B, match),
by = c('A', 'B')) |>
# only keep pairs that have been validated or have a match probability > 0.1
filter((match_probability > 0.1 & is.na(match)) | match == 'Yes') |>
right_join(dfA, by = c('A' = 'name'),
relationship = 'many-to-many') |>
left_join(dfB, by = c('B' = 'name'),
relationship = 'many-to-many')
matches
#> A B sim jw
#> 1 Joe Biden Joseph Robinette Biden 0.7665511 0.7208273
#> 2 Donald Trump Donald John Trump 0.8388565 0.9333333
#> 3 Barack Obama Barack Hussein Obama 0.8458046 0.9200000
#> 4 George W. Bush George Walker Bush 0.8447483 0.9301587
#> 5 Bill Clinton William Jefferson Clinton 0.8731912 0.5788889
#> match_probability match state.x age state.y hobby
#> 1 1 Yes Delaware 81 Delaware Football
#> 2 1 Yes New York 77 Florida Golf
#> 3 1 Yes Illinois 62 Illinois Basketball
#> 4 1 Yes Texas 77 Texas Reading
#> 5 1 Yes Arkansas 77 Arkansas SaxophoneA Note On Cost
Because the fuzzylink() function makes several calls to the OpenAI API—which charges a per-token fee—there is a monetary cost associated with each use. Based on the package defaults and API pricing as of March 2024, here is a table of approximate costs for merging datasets of various sizes.
| dfA | dfB | Approximate Cost (Default Settings) |
|---|---|---|
| 10 | 10 | $0.02 |
| 10 | 100 | $0.02 |
| 10 | 1,000 | $0.02 |
| 10 | 10,000 | $0.02 |
| 10 | 100,000 | $0.07 |
| 10 | 1,000,000 | $0.6 |
| 100 | 10 | $0.15 |
| 100 | 100 | $0.15 |
| 100 | 1,000 | $0.15 |
| 100 | 10,000 | $0.16 |
| 100 | 100,000 | $0.21 |
| 100 | 1,000,000 | $0.74 |
| 1,000 | 10 | $1.5 |
| 1,000 | 100 | $1.5 |
| 1,000 | 1,000 | $1.5 |
| 1,000 | 10,000 | $1.51 |
| 1,000 | 100,000 | $1.56 |
| 1,000 | 1,000,000 | $2.09 |
| 10,000 | 10 | $15.01 |
| 10,000 | 100 | $15.01 |
| 10,000 | 1,000 | $15.01 |
| 10,000 | 10,000 | $15.01 |
| 10,000 | 100,000 | $15.06 |
| 10,000 | 1,000,000 | $15.59 |
| 100,000 | 10 | $30.06 |
| 100,000 | 100 | $30.06 |
| 100,000 | 1,000 | $30.06 |
| 100,000 | 10,000 | $30.06 |
| 100,000 | 100,000 | $30.12 |
| 100,000 | 1,000,000 | $30.64 |
| 1,000,000 | 10 | $30.59 |
| 1,000,000 | 100 | $30.59 |
| 1,000,000 | 1,000 | $30.59 |
| 1,000,000 | 10,000 | $30.59 |
| 1,000,000 | 100,000 | $30.64 |
| 1,000,000 | 1,000,000 | $31.17 |
Note that cost scales more quickly with the size of dfA than with dfB, because it is more costly to complete LLM prompts for validation than it is to retrieve embeddings. For particularly large datasets, one can reduce costs by using GPT-3.5 (model = 'gpt-3.5-turbo'), blocking (blocking.variables), or reducing the maximum number of validations (max_validations).