Add sketch quality figure

Author: Cecca

analysis/latex.R

@@ -27,8 +27,8 @@ latex_table_best <- function(data) {
                 format = "latex"
             ),
             # more compact names
-            algorithm = str_remove(algorithm, "LSH"),
-            dataset = if_else(dataset == "Livejournal", "LJ", as.character(dataset))
+            # algorithm = str_remove(algorithm, "LSH"),
+            # dataset = if_else(dataset == "Livejournal", "LJ", as.character(dataset))
         ) %>%
         ungroup() %>%
         select(dataset, threshold, algorithm, total_time, recall, k, sketch_bits) %>%
@@ -40,7 +40,7 @@ latex_table_best <- function(data) {
         ) %>%
         kbl(
             format = "latex",
-            align = "ll rrll rrll",
+            align = c("l", "l", "r", "r", "l", "l", "r", "r", "l", "l"),
             escape = F,
             booktabs = T,
             linesep = c("", "", "", "\\addlinespace"),

analysis/plot_k_dep.R

@@ -101,7 +101,7 @@ plot_counters <- function(data, t, ylabels = FALSE) {
 }
 
 (plot_counters(plotdata, 0.5, ylabels = TRUE) | plot_counters(plotdata, 0.7)) / guide_area() +
-    plot_layout(guides = "collect", heights = c(3, 1))
+    plot_layout(guides = "collect", heights = c(10, 1))
 
 ggsave("imgs/counters.png", width = 8, height = 5)
 

analysis/plot_sketches.R

@@ -188,4 +188,21 @@ ggsave("imgs/sketches.png", width = 10, height = 5)
 #         loss512 = (`0` - `512`) / `0`
 #     ) %>%
 #     select(dataset, threshold, loss64, loss128, loss256, loss512) %>%
-#     summarise(across(loss64:loss512, ~max(.x) %>% scales::percent(accuracy=0.001)))
+#     summarise(across(loss64:loss512, ~max(.x) %>% scales::percent(accuracy=0.001)))
+
+table_sketch_quality() %>%
+    filter(sketch_bits > 0) %>%
+    select(dataset, threshold, sketch_bits, lost_pairs, lost_fraction) %>%
+    ggplot(aes(x = factor(sketch_bits), y = lost_fraction)) +
+    geom_point() +
+    geom_segment(aes(xend = factor(sketch_bits)), yend = 0) +
+    geom_hline(yintercept = 0.01) +
+    scale_y_continuous(labels = scales::percent_format()) +
+    labs(
+        x = "sketch bits",
+        y = "false negative rate"
+    ) +
+    facet_wrap(vars(dataset, threshold), ncol = 8) +
+    theme_paper()
+
+ggsave("imgs/sketch_loss.png", width = 10, height = 2)

analysis/tables.R

@@ -31,6 +31,53 @@ recode_algorithms <- function(data) {
     )
 }
 
+table_sketch_quality <- function() {
+    db <- DBI::dbConnect(RSQLite::SQLite(), "danny-results.sqlite")
+    # The load is the maximum load among all the workers in a given experiment
+    all <- tbl(db, "result_recent") %>%
+        collect() %>%
+        mutate(
+            timed_out = is.na(output_size) # && !is.na(total_time_ms)
+        ) %>%
+        filter(hosts == "10.1.1.1:2001__10.1.1.2:2001__10.1.1.3:2001__10.1.1.4:2001__10.1.1.5:2001") %>%
+        filter(profile_frequency == 0) %>%
+        filter(!str_detect(path, "sample-200000.bin")) %>%
+        filter(required_recall == 0.8) %>%
+        filter(threshold %in% c(0.5, 0.7)) %>%
+        filter(!no_verify, !no_dedup) %>%
+        # filter(algorithm != "two-level-lsh" | (repetition_batch >= 1000)) %>%
+        mutate(dataset = basename(path)) %>%
+        inner_join(baseinfo) %>%
+        mutate(
+            dry_run = as.logical(dry_run),
+            total_time = set_units(total_time_ms, "ms"),
+            dataset = case_when(
+                str_detect(dataset, "sift") ~ "SIFT",
+                str_detect(dataset, "Livejournal") ~ "Livejournal",
+                str_detect(dataset, "[Gg]love") ~ "Glove",
+                str_detect(dataset, "Orkut") ~ "Orkut"
+            )
+        ) %>%
+        order_datasets() %>%
+        recode_algorithms() %>%
+        filter(algorithm == "Cartesian") %>%
+        select(-total_time_ms)
+
+    baseline <- filter(all, sketch_bits == 0) %>%
+        select(dataset, threshold, base_output_size = output_size)
+
+    all <- inner_join(all, baseline) %>%
+        mutate(
+            lost_pairs = base_output_size - output_size,
+            lost_fraction = lost_pairs / output_size
+        )
+
+
+    DBI::dbDisconnect(db)
+    all
+}
+
+
 table_search_best <- function() {
     db <- DBI::dbConnect(RSQLite::SQLite(), "danny-results.sqlite")
     # The load is the maximum load among all the workers in a given experiment