Update ThreatFox_MalDomains/README.md

ThreatFox_MalDomains/README.md

@@ -1,5 +1,13 @@
+## Columns
+
 This dataset has 140 columns which:
 - 1 column for input
 - 1 column for classes
 - 1 column for labels
-- 137 columns for collected shared metadata
+- 137 columns for collected shared metadata
+
+## 0 As Null Values vs. Honet Null values
+
+## 0 As Null Values vs. Honest Null Values
+
+When a metadata lookup fails — no RDAP record, no TLS handshake, a domain that doesn't resolve — the value for those features is genuinely *unknown*, not zero. The original collector filled these gaps with `0`, which conflates two very different things: a **measured zero** ("the DNS scan ran and found no MX record") and a **missing value** ("we never got an answer"). This is harmful because the missingness rate itself is class-discriminative — benign domains serve a TLS certificate and resolve to an IP roughly 30 points more often than malicious ones — so a fabricated `0` quietly becomes a proxy for the label, inflating in-distribution scores through an artifact that does not transfer to new data. The **honest-null** convention keeps the value column truthful (empty when unknown, `0` only when actually measured as zero) and instead captures the legitimate "no footprint" signal explicitly through a separate per-category **missingness mask**. The result is a representation where `0` always means the same thing across every dataset, and the model learns the real signal of a missing footprint rather than a leaky, dataset-specific fake zero.