From: Vsevolod Stakhov <vsevolod@rspamd.com>
Date: Tue, 20 Jan 2026 16:12:41 +0000 (+0000)
Subject: [Fix] Fix ROC threshold calculation for ham/spam labels
X-Git-Tag: 4.0.0~179^2~10
X-Git-Url: http://git.ipfire.org/gitweb.cgi?a=commitdiff_plain;h=0457fba92bd2997774c040aeec61b56e06859414;p=thirdparty%2Frspamd.git

[Fix] Fix ROC threshold calculation for ham/spam labels

The ROC calculation was checking outputs[i][1] == 0 for ham samples,
but the ceb_neg cost function uses -1.0 for ham and 1.0 for spam.
Changed to check outputs[i][1] < 0 to correctly identify ham samples.
---

diff --git a/lualib/plugins/neural.lua b/lualib/plugins/neural.lua
index 68bdb3c3dc..000a3fc6c6 100644
--- a/lualib/plugins/neural.lua
+++ b/lualib/plugins/neural.lua
@@ -474,7 +474,8 @@ local function get_roc_thresholds(ann, inputs, outputs, alpha, beta)
   spam_count_ahead[n_samples + 1] = 0
 
   for i = n_samples, 1, -1 do
-    if outputs[i][1] == 0 then
+    -- Labels are -1.0 for ham and 1.0 for spam (ceb_neg cost function)
+    if outputs[i][1] < 0 then
       n_ham = n_ham + 1
       ham_count_ahead[i] = 1
       spam_count_ahead[i] = 0
@@ -489,7 +490,8 @@ local function get_roc_thresholds(ann, inputs, outputs, alpha, beta)
   end
 
   for i = 1, n_samples do
-    if outputs[i][1] == 0 then
+    -- Labels are -1.0 for ham and 1.0 for spam (ceb_neg cost function)
+    if outputs[i][1] < 0 then
       ham_count_behind[i] = 1
       spam_count_behind[i] = 0
     else