Online catalogs are limited by text-based search, which relies on product attributions that aren't always present. Even when attributions are present, they often don't describe a product accurately. To maximize online product discovery, additional search experience features are needed to get around these text-based limitations. Adding other methods of discovery has a positive impact on user experience, helping to raise online conversion rates and increase checkout cart quantities. Therefore, firms are looking at product images and other attributes as a new source of data to create an innovative search experience, as well as help attribute poorly attributed products.
Grid Dynamics has worked for many years on using machine learning and artificial intelligence to create innovative digital use-cases. We are strong believers that the next wave of digital transformation will be through the use of machine learning driven algorithms to advance online services and experiences. Therefore, we have invested heavily into convolutional neural networks, a class of deep neural networks that are the primary component of the ML aspect of computer vision. These models are able to open use cases and discover new ways to please customers.
"More like this" is a new, ML-powered product discovery solution and technical blueprint for large digital catalogs. It uses product images as a source of data to show shoppers products that are visually similar to ones they like, creating a new shopping experience and product discovery tool.
In some domains, such as art or interior design, shoppers want to find products that have similar stylistic feel to one another. To quantify style, which is elusive and hard to describe, our engineers applied an image processing technique called “neural style transfer” to a visual search engine. This deep learning based approach enables a new visual search experience that helps shoppers search products with a similar style to ones they already like.
It's difficult to describe many lifestyle and decor products with text alone, making them challenging to find in a large online catalog. Using reverse image search, shoppers can take photos of products they like from popular sites such as Pinterest, and search for visually similar items in a product catalog. This visual search solution identifies objects in an image, extracts image data and uses trained convolutional neural networks and nearest neighbor search to scour the catalog for visually similar products.
Large online catalogs are plagued with thousands of misattributed products, and an even greater number that are not labeled at all. Product misattribution is a problem that runs downstream to negatively affect search engine relevance and conversion rates, making it a major issue. A product attribute checker both accurately identifies product misattributions and suggests improvements. It can also recommend attributes for unlabeled products, automating this essential yet time-consuming merchandising process.
Using convolutional neural networks, we are able to create an image vectorization model that can represent image contents and style numerically in a
“More like this” can also produce feature similarity search by using natural language processing on the catalog’s product attributes and text. This can automate the product comparison of technical products, like electronics and appliances.
