How We Fought Spam in Reviews

Ivan Gorbunov25 May 2024

Problem

A marketplace review service has to keep reviews trustworthy and safe. On Wildberries, reviews are 50% of consumed content, and 93% of buyers decide on a purchase after reading them. Fraudsters flooded the site with reviews that contain only a picture. Spammers do not upload identical files: they compress, enlarge or otherwise transform one image slightly, so exact matching fails. Millions of reviews arrive per day, so sending everything to assessors is not possible.

Example from the article: a review with the text "Good product" and a photo that leads to Telegram or another outside resource.

Solution in brief

Two parallel branches decide on each review. One catches duplicates by image hashes, then by embeddings of duplicate clusters. The other feeds OCR text, a ViT-based image score and feature-store features into CatBoost. Confident decisions are automatic, and the uncertain gray zone goes to assessors, whose labels also feed training and quality dashboards.

Questions this case answers
  • How to detect near-duplicate images that were slightly transformed to evade exact matching, at a scale of millions of new items per day
  • How to combine automatic decisions with human labeling of uncertain cases under a limited labeling budget
  • How to get training data for a new abuse type when no labeled sample exists
  • How to detect spam photos in marketplace reviews when spammers slightly alter the same image

Tools and methods

  • Image hashes in Redis — a counter per hash finds the same image after small transformations.
  • Embeddings of duplicate clusters in ClickHouse — catch similar images with a different hash.
  • EasyOCR (CRAFT + CRNN) — reads text from spam pictures.
  • ViT embedding with task heads — one embedding per image, a fine-tuned head per task.
  • CatBoost — meta-model over OCR text, CV score and features.
  • Gray zone share, Error Rate — online quality metrics.
  • Assessors — label the gray zone and daily samples.
Continue reading — it's free

Google or a one-time email code. No password.