For both businesses and consumers, recommender systems have changed the face of online shopping, streaming and social media in the past two decades.

They predict user preferences and make recommendations of products or content they might like. And they are everywhere.

That list of products on Amazon "inspired by your shopping trends"? It's from a recommender system.

The list of videos on YouTube which are similar to the ones you have recently watched? A recommender system.

The "top picks for you" list of movies on Netflix? Also a recommender system.

A significant portion of profits from e-commerce businesses are generated by recommender systems. For example, according to a McKinsey study in 2013, 35% of Amazon's sales came from recommendations. Nine years on, we can assume that proportion is even greater today.

In the retail sector, Amazon is - predictably - very much the market leader when it comes to recommender systems, having successfully introduced this technology in the 1990s.

These systems are constantly evolving as new methods and algorithms are discovered, not to mention data becoming available in increasing volume and variety. 

If retailers are to compete with the likes of Amazon, this means embracing a new paradigm of increasingly sophisticated recommender systems which provide product suggestions that are even more relevant to the user. But what does this involve?

To answer that, we have to understand the two "traditional" models of recommender systems in retail.

The first is content-based systems, which match customers and products based on known attributes to issue recommendations. However, this approach is limited as it can only provide recommendations based on those known features.

The second is collaborative filtering recommender systems, which instead focuses on similarities between a number of customers based on their interactions with products, such as purchases, search history and reviews. If the system knows customer A and customer B like similar products based on historical information, it will assume they will also like similar products in the future, and recommend on that basis.

But a major disadvantage of this method is the "cold start" when a new customer first engages with the business, meaning this system would not initially know if they like or dislike the same things as customers A and B.

To overcome the limitations outlined above, we must look to the next generation of recommender systems - graph-learning models - which provide a hybrid approach.

These models have become the frontier of recommender systems as a powerful way to determine item characteristics and user preferences in an efficient way.

Graph-learning recommenders are a type of deep learning-based system which use techniques and methods from graph and network theory. They combine rich and complex information from a wide variety of sources, such as information from social media, to create customer to product preferences.

It is an efficient way to continuously learn the drivers of customer behaviour to support increasingly relevant recommendations, meaning it's faster, more efficient and more relevant to customers.

It's an approach which has already been adopted by the likes of Uber Eats - and one retailers will increasingly pursue as the global shift to online shopping continues.

For more information visit  Ocula Technologies