Digital › Software2 min read

Unequal bins skew Dempster conflict K in fusion of 37,975 CVE-2024 severity records

Authors tested on 37,975 CVE-2024 records: with unequal intervals, Dempster's conflict K tracks the evidence position and reverses the apparent link between conflict and severity.

Unequal bins skew Dempster conflict K in fusion of 37,975 CVE-2024 severity records
Photo: ​中華民國總統府, Wikimedia Commons, CC BY 2.0 · License: CC BY 2.0

A study in the Journal of King Saud University - Computer and Information Sciences shows that when a continuous measurement is split into intervals of unequal width, Dempster's conflict K depends on the position of the evidence, not only on disagreement between sources springer.comJournal of King Saud University - Computer and Information Sciencesspringer.comOpen source ↗. The authors tested this on 37,975 CVE-2024 records from three severity databases. The analytical sample was 13,221 dually assessed pairs with 70 confirmed-exploited positives; the third source entered only descriptively.

Bin geometry and conflict

Dempster-Shafer fusion of continuous measurements requires a frame of discernment, normally built by binning. Under any fixed-bandwidth basic probability assignment (BPA), two identical readings conflict more near a narrow interval than near a wide one. On the authors' data this confound reverses the apparent relationship between conflict and severity.

The authors describe two remedies. An excess-conflict statistic removes the confound at negligible cost, and Jousselme's distance is free of it. On such a frame both reduce to a monotone reparametrisation of the plain difference between the readings, so they are diagnostics, not new signals.

A second failure mode is the encoding. A point mass on the interval containing each reading turns K into a two-valued disagreement indicator, with r = 0.998 across 13,221 records springer.comJournal of King Saud University - Computer and Information Sciencesspringer.comOpen source ↗. With singleton focal elements, Dempster's rule becomes a discounted naive-Bayes product over intervals, so the choice of combination rule barely matters.

Fusion results

The comparison covered a 4 × 6 × 4 grid of BPA constructions, combination rules and measures, evaluated under nested cross-validation. Fusion produced a null result for efficacy, which the authors report together with its statistical power. The gain in average precision was directionally consistent over twenty partitions but not significant under every resampling scheme.

The relationship between agreement of the sources and exploitation holds on the records whose scores predate their listing. The authors describe it as a descriptive association, not an operational rule for triage. For pipelines that bin continuous data, they recommend correcting or replacing Dempster's K over a binned frame, and checking the partition by sweeping two identical readings across it to see whether conflict stays flat.

Sources · 2

Found a mistake?
  1. ↩

Read also

Up next

The next story opens automatically

More
Follow us

Type at least two letters — results appear instantly

↑↓to choose↵to openescto close