Why rules alone fail
Classic rules catch primitive fraud only. Modern bots rotate residential IPs and mimic scroll and time-on-site. ML finds patterns you cannot encode: correlations across fingerprint, referrer, hour, and navigation speed.
Three ML pillars
- Device / browser fingerprint. Beyond User-Agent — canvas, WebGL, timezone drift, language mismatches. One user with many fingerprints suggests a farm.
- Behavioral biometrics. Mouse paths, scroll rhythm, pause patterns. Synthetic traffic is often too perfect.
- Anomaly detection. Unsupervised models catch spikes without labeled fraud — useful for zero-day schemes.
AI Selena in ClikBy: honest scope
AI Selena is an model ensemble for click and visit quality on smart links: conversion score, confirmed / bot_server / bot_filter statuses. It analyzes technical and behavioral signals on redirect and on-site.
We do not position Selena as bank-grade transaction ML, do not promise to block every bot, and do not claim certifications the product does not hold. The value is clearer junk traffic in the marketing funnel and cleaner retargeting inputs.
Implementing ML without overreach
- Start in observe mode: track bot share 2–4 weeks before hard blocks.
- Validate scoring against business metrics (lead quality, sales), not only bot %.
- Document false positives — they cost more than missed bots in fintech and e-commerce.
Glossary: antifraud terminology dictionary.
Supervised vs unsupervised in practice
Supervised models need labeled bot/confirmed data. Unsupervised catches new clusters with more false positives. Selena combines both in an ensemble.
Drift and retraining
Bots adapt within weeks. ClikBy updates scoring on-platform; revisit Ads exclusions quarterly.
