Product recommendations are something that nearly every ecommerce retailer has, but very few people understand how they actually work. It’s down to the fact that the technology behind them is complex and complicated to understand. For this reason product recommendations have been shrouded in mystery and packaged up within ecommerce personalization tools and this enables some businesses to twist the truth about how their recommendations work and what technologies they use.   At Qubit we have been developing product recommendation technology for over 8 years and we have been part of every development, from the old to the cutting edge. This gives us a unique perspective on how these systems work and we want to share that with you. In this blog you’ll learn about the differences between machine learning based recommendations and sequential deep learning. It sounds like a lot of buzzwords but stick with us…we’re talking significantly more revenue. 

A lot of the time it can be technology for technology’s sake, however recent developments have genuinely changed the experience for the end customer, in a dramatic way. Businesses are always pushing to drive conversion rates and increase their online revenues. Product recommendations are a brilliant way to do this, hence why they are ubiquitous across most ecommerce websites. Your conventional product recommendations tool will increase revenue by roughly 3%, which is important. These conventional systems have not changed in any significant way until very recently. 

Conventional product recommendations tools are based on machine learning models with the most prominent being collaborative filtering. Cutting-edge systems use sequential deep learning with some collaborative filtering mixed in for good measure.

These conventional systems have sufficed to help try and create dynamic personalized experiences for customers, however as the years have progressed and customer expectations have increased they are no longer sufficient. The main reason for this is that our understanding of the customer and the shopping journey they take has increased, we now truly understand how unique each person's journey is. Customers have distinct preferences, category affinities and price sensitivities. They browse at different times of day,  in various orders and with varying levels of purchase intent. Conventional product recommendations cannot consume and utilize this amount of contextual data and respond with higher levels of relevance. In order to do this you need real time predictive modelling

How do conventional product recommendations engines work?

Largely these engines look for patterns in lots of data. The patterns they look for are, which products are similar to each other in terms of:

  • Viewed together by all customers
  • Bought together by all customers
  • Bought after a different product is viewed by all customers

Even though it’s only 3 key areas that these product recommendations engines use, conventional technology providers will tout hundreds of algorithms to achieve the above, all negating one very important point. The customer. By definition collaborative filtering is not tuned to be personalized 1:1, vendors have tried to find ‘hacks’ to make this possible but they only get you so far. This alongside the fact that most of these machine learning systems, and their extremely data hungry nature, run in batch methods either every few hours or even as slow as daily. Which is far from the real time needs of today’s customer.

Put simply, customers browsing a blue shirt see a “You may also like” carousel of selected similar products, based on browsing behavior of the masses. The products selected in the “You may also like” carousel by conventional recommendations are typically limited to those that are most viewed or purchased within the catalog and may not be the most relevant to the customer. Instead, recommendations should be pulling the best products from the entire breadth of the catalog to effectively inspire customers with truly personalized product recommendations. 

If there’s one thing you can take away from this section it's that: you probably have conventional product recommendations, they are not as personalized as they should be and they are product-to-product.

Why are deep learning recommendations so different

Deep learning is inspired by the makeup of the human brain, using layered sets of algorithms that mimic a neural network. 


neural network

By mimicking the brain, deep learning is able to create important connections between different items, such as products, with much less data than machine learning collaborative filtering. These connections can also be made with lots of different interconnecting signals such as:


  • Every aspect of the product itself (i.e. color, size, category)
  • Understanding how the customer got the the product (i.e. referer)
  • Time spent on a product
  • The order in which products are browsed

The order in which products are viewed is such a fundamental part of the shopping journey that is ignored by vendors offering personalized product recommendations. There’s more on this in our Vuja De paper. 

The other key benefit to using deep learning recommendations over your conventional product recommendations is the speed at which it can make connections between customers and products, using very little data. This is where the neural network really comes into its own, enabling significantly more of the product catalog to be put to work. Soon we will be releasing more research on how deep learning recommendations utilize a lot more of the product catalog than standard recommendations. 

This only matters if it drives better results for businesses and we have found, with our rigorous testing, it really does. We see increases of up to 3% in revenue on top of conventional product recommendations alongside 100%+ increases in engagement. Feel free to take a look at this MandM Direct case study with Google Recommendations AI. 

If there’s one thing you can take away from this section it's that: deep learning recommendations are truly 1:1 and personalized as they’re customer-to-product not product-to-product.

At Qubit we are data-driven, enabling us to be laser focused on achieving high success rates for our customers. Generic solutions that treat every customer the same breed generic results and that’s why we’re always at the forefront of new technologies that enable us to drive a truly personalized ecommerce experience. We’re here to make ecommerce better. 

If you think your product recommendations are underperforming or you’ve seen then plateau then sign up for a free recommendations health check.


Jessica Sta. Maria

Read More

Subscribe to stay up to date.

Receive the personalization newsletter directly to your inbox.