Product Attributes
The Challenge: Without accurate product attributes, product descriptions lack detail, leaving customers frustrated as they sift through broad search results to find exactly what they want.
The Solution: An AI-driven system that extracts attributes from images and data, from colors and materials to target audiences and companion products, providing accurate information for better descriptions and search results.
Schema mapping vs freestyle
Going from images and other data sources to full attributions on products is one of those things that adds tremendous value and saves endless hours of work. So we decided to build that.
Without proper attribution, you might end up in a situation where a search on “black dress” will return 6,000 black dresses, but no further way to drill down. Your copy can end up vague, and on Google search, even luck won’t help you reach your potential customers.
So, we decided to build a product attribution system that leverages AI to provide comprehensive attribution for all your products in the background, so it’s just there and ready for you. This data can be used in copywriting, in your on-site search, or even to complete the data in your PIM system.
We built a four-step method for this, so you can rely on the results without having to validate and check everything. The first step is where we “find” the product in all the supplied images. The more data we get, the better our results will be.
From there, we do an in-depth analysis of the images, guided by extensive schemas, in freestyle attribution mode, gathering all the impressions we can. This includes everything from color or product type to design styles, target audiences, and occasions for use. These schemas are, of course, easy to customize and extend.
The third step is converting this freestyle attribute extraction to specific data schemas. We might have a dress categorized as a traditional prom dress, but your business is not in a region that uses the concept of prom. Well, since we store all the extracted data as multi-dimensional vectors, we can map the attributes in “space” and use geometric operators to extract the closest related options from your specified attribute option in your custom schema. That way, your data will be both highly relevant and conform to your data structure.
With all this done in the background, without you noticing it, we validate the extracted data with the images and whatever other data we have available. This will provide you with a confidence score for the attributes and allow us to make the last few corrections, if any are needed.