Taboola

Developed an AI-powered image similarity model
based on human input

Image
Accelerate your
e-commerce AI
Talk to our AI expert
Accelerate your
e-commerce AI
Talk to our AI expert

Client

Content recommendation platform that works with news publishers (CNN, NBC, MSN.com, and others) and advertisers who pay to have their ads displayed in the newsfeed.

Challenge

Required an efficient model for ad moderation that would reduce the likelihood of unwanted content appearing in the news feed.
  • Inappropriate or incorrectly classified ads occasionally appeared in the news feed as a result of errors in automatic or AI-based moderation.
  • Similar images were difficult for automated moderation and AI models to recognize and process automatically.
  • Multiple iterations of manual moderation were performed on identical or very similar images, increasing the workload.

Solution

The new model combines automation and AI technologies with human labeling.
  1. Human labeling serves as the foundation for defining ground truth labels.
  2. To find similarities, the model calculates the distance between image elements and considers them "similar" if the distance is less than a predefined threshold.
  3. The threshold value was determined through experiments and comparing the results to ground truth data.

Business results

  • AI-powered automation covers 11% of all advertisements compared to 4% earlier.
  • Manual review was reduced from 31% to 20% of ads, allowing in-house content moderators to focus on more important tasks.

“The majority of our models are not actually developed internally at Taboola. They are from the open web, and you can use existing models and just manipulate them to feed them your own products. So you don't have to build your own model each time you face a problem.”

– Gal Cohen, Product Manager, Taboola

Similar success stories

  • Trained a voice assistant with language data

    Results:

    x5

    accuracy in the new languages

    Read the storyRead the story
  • Set up a moderation process for items on sale to ensure legal and cultural compliance

    Results:

    50%

    reduced cost per item with x500 more items verified daily

    Read the storyRead the story
  • Improved the accuracy of a predictive tool using local shopping patterns

    Results:

    30%

    improvement in app accuracy after data collection, reaching 95%

    Read the storyRead the story
  • Improved crowdsourced translations of product descriptions

    Results:

    17%

    budget reduction while achieving optimal quality

    Read the storyRead the story

Accelerate your e-commerce AI

Let's talk about the ideal solution for your data needs.
Fractal