In rexdouglass/MeasuringLandscape: Replication file for the paper Measuring the Landscape of CivilWar

Fuzzy Matcher Stage 1

The fuzzy matcher predicts the likelihood that two toponyms are the same place even though their spellings might be different. It has two stages.

This file develops Stage 1 of the fuzzy toponym matcher. Its job is to screen out the vast majority of suggestions that are too dissimilar to ever be a plausible match. It will have a high false positive rate, which can then be further refined in stage 2.

rm(list=ls()); gc()
library(MeasuringLandscape)
library(tidyverse)
#devtools::load_all()

dir_figures <- glue::glue(here::here(), "/paper/figures/")

knitr::opts_knit$set(progress = TRUE, verbose = TRUE)
knitr::opts_chunk$set(fig.width='100%',  warning=FALSE, message=FALSE, cache=TRUE)
options(width = 160)

Load Hand Labeled Place Matches

#Load Hand Labeled Examples
handlabeled <- data.table::fread(system.file("extdata",
                                 "event_flatfile_matches_for_hand_labeling - event_flatfile_matches_for_hand_labeling.csv",
                                 package = "MeasuringLandscape"), data.table=T) %>% distinct() 
dim(handlabeled)

# Remove exact matches 
# Rewrote  with foreach to run in parallel on windows, but it's still relatively slow code with 8 cores.
handlabeled$stemmed_a <- MeasuringLandscape:::strip_postfixes(to_be_striped=handlabeled$name_cleaner_a)[[1]]
handlabeled$stemmed_b <- MeasuringLandscape:::strip_postfixes(to_be_striped=handlabeled$name_cleaner_b)[[1]]

handlabeled_unique <- subset(handlabeled, stemmed_a!=stemmed_b) # very important we're dropping any with identical stems for evaluation
table(handlabeled_unique$rex_match) #1090 matches, 16978 nonmatches

#Stem them
handlabeled_unique$stemmed_a <- MeasuringLandscape:::strip_postfixes(handlabeled_unique$name_cleaner_a)[[1]]
handlabeled_unique$stemmed_b <- MeasuringLandscape:::strip_postfixes(handlabeled_unique$name_cleaner_b)[[1]]
handlabeled_unique$stemmed_ab <- sapply(lapply(strsplit(paste(handlabeled_unique$stemmed_a,
                                                              handlabeled_unique$stemmed_b, sep="_"),
                                                        "_"),
                                               sort),
                                        paste,
                                        collapse="_")


dim(handlabeled)
handlabeled$a <- handlabeled$name_cleaner_a
handlabeled$b <- handlabeled$name_cleaner_b
handlabeled[,ab:=paste(a,b,sep="_")]
handlabeled[,ba:=paste(b,a,sep="_")]

stemmed_ab <- unique(c(handlabeled$stemmed_a, handlabeled$stemmed_b)) ; length(stemmed_ab) #where ab is the unique toponym strings found in the data

Grid search optimal parameters for locality sensitive hashing.

Settling on 20 bands, 5 rows, and qgram of 1 letter

fromsrcatch=F
if(fromsrcatch){
  grid_search_lhs <- list()
  for(q in c(2,5,10,20,25,50)){
    print(q)
    grid_search_lhs[[as.character(q)]] <- lhs_textreuse(minhash_count=100,  bands=q) #good trade off
    print(grid_search_lhs[[as.character(q)]])
  }
  grid_search_lhs_dt <- rbindlist(grid_search_lhs)

  saveRDS(grid_search_lhs_dt,
          glue::glue(getwd(), "/../inst/extdata/grid_search_lhs_dt.Rds"))
}

grid_search_lhs_dt <- readRDS(system.file("extdata", "grid_search_lhs_dt.Rds", package = "MeasuringLandscape"))

There's a big discontinuity between 50 bands and 25, and then diminishing returns with increasingly higher false negative rates thereafter. Choosing 25 as a compromise between low false negative rate and fewer suggestions per case. (Appendix Figure 4)

p_lhs_gridsearch <- ggplot(grid_search_lhs_dt, aes(x=false_negative,y=suggestions_per, label=bands)) + geom_label() + ggtitle("")
p_lhs_gridsearch

ggsave(
  filename = glue::glue(dir_figures, "p_lhs_gridsearch.pdf"),
  plot = p_lhs_gridsearch,
  width = 5.5,
  #height = 8,
  device = cairo_pdf #have to use cairo to correctly embed the fonts
)

rexdouglass/MeasuringLandscape documentation built on May 13, 2019, 6:16 p.m.