Building online recommendation engines to improve customer experience

Product inventories and availability has grown at a rapid pace in e-commerce. But if customers can’t find what they want on your site, they will leave and shop somewhere else. Better search functionality on your website is one way to help but unfortunately customers don’t have the time to dig through search results or create elaborate filters. But there is another way.

02/ The problem

Product recommendations give the customer options on related products that are relevant to their interest and intentions to buy. They might be interested in one item but their size is out of stock but there is a similar item that has the right size available. Product recommendations can make that connection for them and combat lost sales opportunities for your business. The right product recommendations can also increase basket size and average order value (AOV) through pairing complementary products.

But many companies do not have the time to go through their entire inventory playing matchmaker to make these recommendations.

03/ The opportunity for deep learning

Deep Learning models give machines an ability level to understand the subjects of images and texts. These models allow us to automate repetitive tasks, like matching an entire product inventory, with incredible speed and scale. But understanding what is an image, for example a Blue Shirt, is only half the battle of providing suitable recommendations for customers. How can we train a model to take that understanding and find similar products?

04/ Suggested solution

Product recommendations can be based on a lot of important information for customers. An understanding of who and what your customers like will always be the foundation of building successful AI powered product recommendations. Should recommendations be based on matching similar product descriptions, product images, similar brands/quality or perhaps purchases made by similar customer profiles? 

For this use case, we will focus on looking at the images and text descriptions which deep learning excels at. Using the power of text and image embeddings, a Deep Learning model can understand the similarities between product images or product descriptions. This is done by the model learning creating representations of these texts or images and then mapping to similar representations that the model has learned from.

05/ Data requirements

The best place to start for the data to build a Similarity model is within your product inventory. No matter if this lives in a simple solution like an Excel file or something more advanced like a PIM (Product information management), you want to be able to build a dataset using the product images and/or descriptions you already have. It is important that images are of the same size and descriptions have some detail so that the model gains a good understanding of them. The better the data, the better the recommendations so its is worth the time and effort to build a well formatted dataset.

06/ How does the model work?

Similarity Models work in the world of representation. What this means is that the model goes through the given dataset and assigns a mathematical representation, a vector, to a given product description for example. Then taking this representation, it can calculate the mathematical distance between the other assigned representations to determine how close or similar they are. A successful result would not imitate someone going through a list or exact matches but show some deeper understanding of the text. A good example would be: 

This coat is the best winter coat to buy this season. 

Purchasing this coat will make sure you stay warm in cold weather.

07/ Test it for yourself

To test this yourself, you can build similarity models that find similar Google Questions or Find Similar images of fruits on our Tutorials Page. If you have any questions or get stuck on completing them, please reach out to our support team via the Intercom Bubble inside the platform.

08/ Want to learn more?

If you're interested in learning more, have a look at our record webinars where we offer a deep dive into Text Similarity and Image Similarity.