The developer of the email validation service Cleancontact has implemented a machine learning-based scoring system to identify non-human email addresses. The tool aims to detect synthetic addresses generated by scripts, which are often used for mass registrations, even when the addresses themselves are technically valid and deliverable.
The solution utilizes the CatBoost library for gradient boosting, chosen for its efficiency and support for ONNX export, allowing the model to run within a Node.js backend without requiring a separate Python process. The model evaluates only the local part of the email address, excluding the domain to avoid bias.
The training dataset was constructed by combining confirmed valid email addresses with synthetic negative samples. These synthetic addresses were generated using six distinct methods, including random character combinations, keyboard walks, and dot-trick variations, to simulate common spammer tactics.
To improve accuracy, the developer incorporated n-gram analysis to evaluate character sequences and integrated name databases covering 55 regions. This addition helps the model recognize non-Russian names, which were previously misidentified as random strings, significantly improving the scoring of international email addresses.
