Skip to content

Latest commit

 

History

History
172 lines (110 loc) · 8.13 KB

README.md

File metadata and controls

172 lines (110 loc) · 8.13 KB

FastLink.jl

Fast Probabilistic Record Linkage for the Julia Language

What is FastLink.jl

The purpose of FastLink.jl is to bring a fast record linkage package to the julia language. When attempting to match large datasets using existing libraries in R and Python, I found they can be very slow and succumb to issues with memory pressure. This implementation of the fastlink algorithm is intended to scale effeciently in parallel and be able to easily handle matches between tabular data that span millions of rows.

Run tests


Using the fastLink function

The basic arguments for the fastLink function to run are

  • dfA: A DataFrame table of records to be matched.

  • dfB: A DataFrame table of records to be matched.

  • config: A Dict{String, Any} that specifies how the two dataframes ought to be matched.

Match Configurations

The match configuration for a FastLink match needs to contain certain in the base dictionary (nested dictionaries will be discussed later).

The Base Dictionary needs to contain:

  • link_type: Either link_only, dedupe_only, or link_and_dedupe.

  • idvar: A Vector{String} of length 2 that specifies the ids of the two dataframes (in [dfA, dfB] order).

  • comparisons: a Dict{String, Any} a that defines the type of matching to be done and the variables that will be matched.

Comparisons dictionary

The comparison dictionary defined above can be located in the base Dictionary or can be substituted instead of a varname dictionary in the variables vector. The effect of nesting the comparisons in the variables vector will lead it to be matched first using the fastlink algorithm and then treated as a single variables in the parent comparisons dictionary. You can substitute multiple varnames for comparisons at the same level of nestedness.

Each comparisons dictionary much have:

  • name: should be "total" in the base dictionary and then can be any name for the nested dictionaries.

  • variables: a Vector{Dict{String, Any}} that contains the individual variable dictionaries and/or a comparisons dictionaries.

The optional parameters for the comparisons dictionary are:

  • w_lambda::Float64: Default 0.0.

  • prior_lambda::Float64: Default 0.0.

  • threshold_match: Lower bound for the posterior probability that will act as a cutoff for matches. Default [0.85].

  • tol_em: Convergence tolerance for the EM Algorithm. (default 1e-05)

  • prior_pi::Float64: Default 0.0.

  • w_pi::Float64: Default 0.0.

Variables

Individual variables can be declared in a dictionary and must contain both a varname and method.

  • varname: name of the variable in dfA and dfB to be compared.

  • method: the method to match the variable. The current accepted methods are (exact, fuzzy, string, numeric, float, int any of the distmethod options).

Methods

Each method has a number of arguments that can be specified for that matching method.

fuzzy

string

numeric

  • term_freq_adjustment: A Bool that determines whether you want the term frequencies for each comparision for a given variable. Note: does not adjust match weight.

  • tf_adjustment_weight: how much to weight on the term_freq_adjustment vs the predicted match value.

  • tf_minimum_u_value: minimum term frequency value to adjust by.

  • partial: A Bool that specifies whether you want to do 2 (true) or 1 (false) comparison levels for a given variable. Default value true.

  • upper_case: A Bool that specifies whether a strings column value is upper or lower (only if method=true. Default value is true.

  • stringdist_method: A String that specifies the desired string distance method ("jw" Jaro-Winkler (Default), "dl" Damerau-Levenshtein, "jaro" Jaro, "lv" Levenshtein, and "ham" Hamming). Default "jw".

  • cut_a: A Float that specifies the first lower bound for string distance cutoff for each comparison. Default 0.92.

  • cut_b: A Float that specifies the second lower bound for string distance (if varnames in partial) for each comparison. Default 0.88.

  • w: A Float that specifies the Winkler weight for jw string distance for each comparison. Default 0.1.

Example FastLink configuration with 1 embedded comparisons dictionary

{
    "link_type": "link_only",
    "idvar": ["id", "id2"],
    "comparisons": {
        "name": "total",
        "prior_lambda": 0.000001,
        "w_lambda": 0.5,
        "threshold_match": 0.88,
        "variables": [
            {"varname": "firstname", "method": "fuzzy", "partial": true, "cut_a": 0.92, "cut_b": 0.88, "upper": true, "tf_adjust": true, "w": 0.1},
            {"varname": "middlename", "method": "exact"},
            {"varname": "lastname", "method": "jarowinkler", "tf_adjust": true},
            {"varname": "birthyear", "method": "exact"},
            {
                "comparisons": {
                    "name": "address",
                    "threshold_match": 0.92,
                    "variables": [
                        {"varname": "housenum", "method": "exact", "tf_adjust": true},
                        {"varname": "streetname", "method": "jarowinkler", "w": 0.1, "tf_adjust": true, "tf_adjustment_weight":0.25, "tf_minimum_u_value": 0.001},
                        {"varname": "city", "method": "jarowinkler", "tf_adjustment_weight":0.15, "tf_adjust": true}
                    ]
                }
            }
        ]
    }
}

fastLink's output

For ease of extracting matches, the getMatches function was added. You can call it on the fastLink output as the single argument getMatches(FastLinkOutput) or with a specified threshold getMatches(FastLinkOutput, threshold_match).

The FastLink output is: A Dict{String,Any} with these vars:

  • ids: A vector of vectors of tuple pairs of ids for each match pattern.
  • idvar: ID variable from configuration
  • resultsEM: The results of the Expectation Maximization algorithm

If term frequency is specified then

  • resultsTF: term frequencies for each variable with specified term frequency by pattern if relevant for the pattern (if no term frequency is applied then tf_adjusted is false).

If benchmark is specified:

  • benchtimes: times for each variable to be matched.

Within resultsEM in the EM output, there is:

  • iter_converge - number of iterations for expectation maximization algorithm to converge.

  • obs_a - observations in dfA

  • obs_b - observations in dfB

  • p_m - posterior match probability

  • p_u - posterior not match probability

  • number_of_unique_patterns - equivalent to number of rows in patterns_w

  • number_of_comparisons - For convenience nrow(dfA) * nrow(dfB)

  • patterns_w - a DataFrame of:

    • gamma_* - An Int64 with the gamma values for each variable (similar to patterns_b)
    • counts - An Int64 with counts for each agreement pattern
    • weights - An Int64 with partial match weights for each agreement pattern
    • p_gamma_jm - A Float64 that has the posterior probability that a pair matches for each agreement pattern
    • p_gamma_ju - A Float64 that has the posterior probability that a pair does not match for each agreement pattern
    • is_match - A Bool that specifies whether the given pattern is above the input parameter threshold_match
  • patterns_b - vector of all patterns observed. each pattern as a scored number for each variable (0 nonmatch, 1 partial, 2 exact, 3 missing)

  • pgamma_km - A Vector{Vector{Float64}} with posterior probababilities for each variable in the EM algorithm. Ordered (0,1,2).

  • pgamma_ku - A Vector{Vector{Float64}} with posterior probababilities for each variable in the EM algorithm. Ordered (2,1,0).

  • p_gamma_jm - A Float64 that has the posterior probability that a pair matches for each agreement pattern (see patterns_w).

  • p_gamma_ju - A Float64 that has the posterior probability that a pair does not match for each agreement pattern (see patterns_w).

  • varnames - A Vector{String} of the input variable names