Data Scientists
Retail & e-Commerce
Industry
Amsterdam, The Netherlands
Location
10,000+
Company size
We had a lot of user-generated content and were struggling to make sense of it given the lack of consistency in user-generated tags. Our product recommendation engine was limited in its functionality because of this. Using the super.AI platform has allowed us to define a consistent taxonomy, label hundreds of thousands of images in record time, and significantly enhance our recommendation engine.
Ricardo C
Data Scientist

Starting point: inconsistent product tags generated by end users

The customer is a multinational Internet group which is known for its principal operations in Internet communication, entertainment, gaming and e-commerce. For one of their e-commerce sites, they wanted to enhance the product search and the recommendation engine. The problem they were facing was that most of the product details (images, description, tags) were provided directly by the users. This resulted in an inconsistent tagging system, with different labeling categories used for the same product types. The resulting models they built using the user-generated tags was therefore based on noisy data, which affected the overall quality of the recommendations offered to the other customers.

Improving tagging and categorisation with super.AI

When the company contacted us, we started by helping the customer build a consistent product taxonomy. This was important to the overall success of the recommendation engine. We then went on to agree with them on the edge cases: tags for subjective categories such as style. Once there was full alignment on the categories used and we ensured labeling consistency, we transcribed hundreds of thousands of user-generated images with the help of our image transcription data program.

Results

The company was able to tag several hundred thousand images within a record timeframe. The labeled data also met their quality thresholds and will be the basis for training a new product recommendation engine with a much higher level of granularity.

Related Case Studies