跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2611.22716 · v1

High-Coverage Blind Spots in Online Abuse Monitoring: Account Selection, Held-Out Flagging Rates, and Sample Composition

首次提交 2026-11-17 20:55
版本更新 2026-11-17 20:55
雷达观测 2026-11-18 11:33
arXiv ↗PDF ↗

摘要

Platforms with limited monitoring capacity may rank target accounts by the number of abusive replies received or by the share of replies labeled abusive. We test how these rankings affect account selection, historical flagged-reply coverage, and held-out labeling rates. In the Online Abusive Attacks (OAA) data (945,987 replies to 384 accounts), two 25-account lists share just one account. Count-based selection covers 75.56% of flagged replies, while rate-based selection covers 5.04%. In a 500-iteration stratified historical hold-out, the lists have labeling rates of 2.69% and 4.98%, respectively (1.85 times). A volume- and group-preserving null model produces low list overlap without a corresponding rate advantage. We repeat the comparison on a second dataset of replies to UK election candidates. The two 25-account lists overlap in eight accounts, but the rate advantage is smaller in historical splits and reverses in three matched-group chronological tests. The findings distinguish differences in who is selected for monitoring from differences in future labeling rates across populations and time.

作者备注:7 pages, 2 figures, 4 tables. Preprint

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

目前只收录了一个版本。收录下一版后,可以对照标题、作者、分类和摘要。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v1 · 2026-11-17 20:55 · 最新观测版本

High-Coverage Blind Spots in Online Abuse Monitoring: Account Selection, Held-Out Flagging Rates, and Sample Composition

Platforms with limited monitoring capacity may rank target accounts by the number of abusive replies received or by the share of replies labeled abusive. We test how these rankings affect account selection, historical flagged-reply coverage, and held-out labeling rates. In the Online Abusive Attacks (OAA) data (945,987 replies to 384 accounts), two 25-account lists share just one account. Count-based selection covers 75.56% of flagged replies, while rate-based selection covers 5.04%. In a 500-iteration stratified historical hold-out, the lists have labeling rates of 2.69% and 4.98%, respectively (1.85 times). A volume- and group-preserving null model produces low list overlap without a corresponding rate advantage. We repeat the comparison on a second dataset of replies to UK election candidates. The two 25-account lists overlap in eight accounts, but the rate advantage is smaller in historical splits and reverses in three matched-group chronological tests. The findings distinguish differences in who is selected for monitoring from differences in future labeling rates across populations and time.

查看此版本解读 →查看此版本原文 ↗