Growth of online shopping presents new challenges
Over the past decade, online retail has grown exponentially, accounting for a larger and larger share of market revenue with no sign of stopping. Companies that have traditionally relied on physical locations are expanding into ecommerce, which brings its own set of challenges. One of the biggest is onboarding new products rapidly without sacrificing profit or searchability. More than 30% of online shoppers abandon their carts before purchase, due primarily to poor product descriptions and difficulty locating items via search.
Getting the quality of a product catalog to a point where a product from a catalog of millions can be easily found isn’t as trivial a task as it sounds. If done manually, adding structured product attributes in a taxonomy defined by a retailer can take several weeks. Weeks that the product is sitting in the inventory not being sold. Considering that many new products do most of their sale in the first few weeks, these lost weeks can be detrimental to a retail business giving significant advantage to a competitor like Amazon that already has a swift onboarding process.
A top 3 retailer automates on-boarding data for millions of products using AI
While the company already had a nationwide physical presence, they needed to grow their online catalog from 1 million to more than 200 million individual products. Since the number of SKUs would continue to grow daily, adding products by hand was impracticable. The retailer hoped to find a way to add products automatically, while still tagging them intuitively. Accurate tagging would ensure that customers could find the items they sought and would help the company maintain accurate inventory across hundreds of stores and warehouses.
We realized very early that a generic Deep Learning model to extract all tags from an image or text document just wouldn’t work. The precision would be nowhere near that expected by the Client and moreover since the data would be extracted as a bunch of keywords, we would lose the context behind the tags.
So, when you extract the word “Blue,” knowing whether it’s a color, a brand or just an emotion becomes very important when it comes to finding the right product from a catalog with hundreds of millions of products. You don’t want someone searching for bed sheets see paper sheets in their search results!
So, our approach was to build hundreds of micro-models for each unique product attribute in the client’s catalog so a model for extracting color and another one for extracting the pattern and so on. This approach allowed us to increase the precision of our models to an average of 90%+.
An added advantage of this approach was that by having a model dedicated to each attribute, even if a model had a precision of less than 90%, the rest of the attributes of a product would still be at an acceptable accuracy and any tuning would be needed only for the model that had a precision lower than 90%. Unlike with a generic model where either the whole model works or doesn’t!
Our solution has allowed the retailer to grow it’s catalog from 1 million to over 200 million in a short period of 18 months. And they’ve done this while simultaneously reducing their dependence on human resources from over 3000 data entry executives to just 50 that validate the outcome of our models. It wouldn’t be an exaggeration to say that building a robust product catalog of this size would not have been possible without the help of CrowdANALYTIX.