rnndescent parameters for search included into the controls_ann() fun…

…ction

R/controls.R

@@ -36,7 +36,10 @@ controls_ann <- function(
                weight_by_degree = FALSE,
                prune_reverse = FALSE,
                progress = "bar",
-               obs = "R"),
+               obs = "R",
+               ##
+               max_search_fraction = 1,
+               epsilon = 0.1),
     hnsw = list(M = 25,
                 ef_c = 200,
                 ef_s = 200,

R/method_nnd.R

@@ -67,12 +67,12 @@ method_nnd <- function(x,
   l_1nn <- rnndescent::rnnd_query(index = l_ind,
                                   query = y,
                                   k = if (nrow(x) < control$k_search) nrow(x) else control$k_search,
-                                  epsilon = 0.1,
-                                  max_search_fraction = 1,
+                                  epsilon = control$nnd$epsilon,
+                                  max_search_fraction = control$nnd$max_search_fraction,
                                   init = NULL,
                                   verbose = verbose,
                                   n_threads = n_threads,
-                                  obs = "R")
+                                  obs = control$nnd$obs)
 
   # if (!is.null(path)) {
   #   if (grepl("(/|\\\\)$", path)) {

vignettes/v2-reclin.Rmd

@@ -157,14 +157,14 @@ It seems that the default parameters of the NND method result in an FNR of `r sp
 ```{r}
 set.seed(2024)
 ann_control_pars <- controls_ann()
-ann_control_pars$k_search <- 60
+ann_control_pars$nnd$epsilon <- 0.2
 
 result3 <- blocking(x = census$txt, y = cis$txt, verbose = 1, 
                     true_blocks = matches[, .(x, y, block)], n_threads = 8, 
                     control_ann = ann_control_pars)
 ```
 
-Changing the `k_search` parameter from 30 to 60 decreased the FDR to `r sprintf("%.1f",result3$metrics["fnr"]*100)`%.
+Changing the `epsilon` search parameter from 0.1 to 0.2 decreased the FDR to `r sprintf("%.1f",result3$metrics["fnr"]*100)`%.
 
 ```{r}
 result3
@@ -184,12 +184,6 @@ It seems that the HNSW algorithm performed better with `r sprintf("%.2f",result4
 result4
 ```
 
-However, this comes at a cost, especially in terms of computation:
-
-1. the HNSW does not handle sparse matrices, so a sparse matrix of tokens must be converted to dense or provided line by line.
-2. The HNSW algorithm is slower than NND. 
-
-Computation times are: 16 seconds for NND and about 60 for HNSW (on M2 MacBook AIR). We can improve the time by changing the parameters `M` and `ef_s` in the `controls_ann()` function (e.g. setting `M=16` and `ef_s=15` leads to about 16 seconds with 1\% FNR).
 
 ## Compare results