Finding Spam in Short Text Fields
summary
LinkedIn builds a language-agnostic naive Bayes n-gram model that scores names using character n-grams, handling unseen n-grams by aggregation or recursive backoff and running a 5-gram heavy / 3-gram light variant; the full algorithm yields better precision/recall after relabeling some false positives. They report that 54% unique first names and 47% unique last names appear in flagged accounts. A complementary random-forest cluster model aggregates group-level name/email/IP features to find mass-created accounts, and flagged users are asked to update names rather than locked out.
AI-written summary · about
chapters (8)
- 0:00Start
- 5:08Direct vs. Indirect Spam Channels
- 8:10Name Scoring Model - Missing Features
- 9:59Name Scoring Model - Performance
- 10:49Name Scoring Model - User Experience
- 11:49Name Scoring Model - Summary
- 12:47Account Clustering
- 14:57Thank you!
related talks
29m(Benford's) Law and Order (Fraud)
24mHighly Scalable Discriminative Spam Filtering
33mImpact of NLP in AI-powered Bank
27mIs that spam in my ham? A novice's inquiry into classification.
28mEnhancing Fraud Detection
23mText mining online data with scikit-learn
01h 24mMachine Learning with Text in scikit learn
40mBalancing Privacy & Utility: Efficient PII Detection and Replacement in Textual Data