In kailin-lu/recordlinkR: recordlinkR

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(recordlinkR)

# Load sample Iowa data
data(iowa_sample)

This chunk is set to currently not run for convenience. Default parameters of the model are likely not sufficient to achieve a good model fit.

To see parameters and check default settings on the model, see the help manual at ?trainModel.

namesA <- iowa_1915$lname1915[iowa_true_matches$index1915]
namesB <- iowa_1940$lname1940[iowa_true_matches$index1940]

encoder <- trainModel(namesA = namesA, 
                      namesB = namesB, 
                      dim.encode = 64, 
                      dim.decode = 64,
                      dim.latent = 48, 
                      num.encode.layers = 2, 
                      num.decode.layers = 2, 
                      earlystop.patience = 20, 
                      batch.size = 32,
                      epochs = 350, 
                      lr = 5e-4, 
                      save.dir="~/blocking_models/iowa_last_48/")

encoder

save(encoder_iowa_last_48, file="encoder_iowa_last_48.rda")

Outside of model accuracy, models can be evaluated by plotting a sample of names represented as a 2-dimensional projection of the latent space.

# Load prepackaged sample encoders 
loadSampleEncoders()

# Plot sample of 100 first names 
plotNames(names=iowa_1915$fname1915, 
          n.samples = 100, 
          encoder.model.path = encoder_iowa_first_512)