about summary refs log tree commit diff
path: root/src
diff options
context:
space:
mode:
authorPjotr Prins <pjotr.public01@thebird.nl>2026-11-14 23:33:37 +0000
committerPjotr Prins <pjotr.public01@thebird.nl>2026-11-14 23:33:37 +0000
commit12c2a37c3ef3a09f1ab0c5ea69b31b433b15f9f6 (patch)
tree96f87fc751d709e7faad995619d1beafd6248381 /src
parent473ad82b6f4477bbbeeef977e3e9f830f59a9b7f (diff)
downloadcorrelation_rust-main.tar.gz
Expose Spearman, so both methods work without a dataset file main
`Spearman` was a private struct, so only `Pearson` could be constructed
from outside the crate: a caller that already holds two value vectors --
a request handler, an enrichment step, a test -- could not compute the
rank correlation without going through `Compute`, which means writing the
dataset to a file and reading the result back out of a CSV. That is why
`Pearson` was public and `Spearman` was not, and it is not a distinction
worth keeping: the p-value code is the same either way.

`Compute` remains the streaming path for a whole dataset.

tests/in_memory.rs is an integration test (outside the crate, so it can
only use the public surface) covering the three pieces that path needs:
both methods in memory, the positional row parser dropping unparseable
values together with their x, and the |r|-descending/NaN-last ordering.

Also documents `Compute::sorter`, which orders by *signed* r while
`custom_float_sorter` -- the one `compute` uses -- orders by magnitude.
The two disagree for negative correlations; left as is because it is
public API, but no longer silent.
Diffstat (limited to 'src')
-rw-r--r--src/correlations.rs18
1 files changed, 16 insertions, 2 deletions
diff --git a/src/correlations.rs b/src/correlations.rs
index 356bfcc..2b9c0f1 100644
--- a/src/correlations.rs
+++ b/src/correlations.rs
@@ -20,8 +20,16 @@ pub struct Pearson {
     degrees_of_freedom: f64,
 }
 
+/// Spearman's rank correlation, on the same footing as [`Pearson`].
+///
+/// Public for the same reason `Pearson` is: a caller that already holds the
+/// two value vectors (a request handler, an enrichment step, a test) should be
+/// able to compute the rank correlation without going through
+/// [`Compute`] -- which means writing a dataset to a file and reading the
+/// result back out of a CSV.  `Compute` stays the streaming path for a whole
+/// dataset; this is the in-memory one, and the p-value is the same code.
 #[derive(PartialEq, Debug)]
-struct Spearman {
+pub struct Spearman {
     n: usize,
     degrees_of_freedom: f64,
 }
@@ -127,8 +135,14 @@ impl<'a> Compute<'a> {
         }
     }
 
+    /// Sort by the correction coefficient, descending.
+    ///
+    /// NOTE: this orders by *signed* r, while [`crate::sorter::custom_float_sorter`]
+    /// -- the one `compute` actually uses -- orders by *magnitude*, with NaNs
+    /// last. The two disagree for negative correlations, so prefer the latter
+    /// when "same as the file path" is what a caller wants. Kept as it is
+    /// because it is public API, not because the difference is intended.
     pub fn sorter(results: &mut [(f64, f64)]) {
-        //naive sorter
         results.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
     }