Recommendation Engines— Netflix and Amazon product recommendations techniques.

5 min readJul 31, 2021

In my previous article we looked at the basics of recommendation engines and how they can help improve sales and other metrics. To better understand the concept, we will look at examples of how digital applications use recommendations to drive purchases.

One of my favorite examples of recommendation engines is Netflix. Let’s break it down.

Netflix

Netflix is the leading internet television network and a lot of credit goes to the personalization brought in by their recommendation engine. Netflix has about 6000 movies and TV shows combined in countries like the UK and Canada, and there’s no way users will be able to sift through all of those, to find that perfect blend of their favorite action-comedy.

When I open my Netflix home-page, it welcomes me with an almost full-screen video of one of my favorite web series. The video begins, along with a call-to-action button, and I’m immediately hooked. The recommendation engine had figured that if there’s anything I’d be watching right now it’s this show! To further verify this, I switched to my sister’s account and she had another movie-recommendation on her home-page.

Lolomo Rows

Netflix then displays particular sets of movies in horizontal scrollable lists also known as Lolomo (list of list of Movies), and vertically arranges these Lolomo rows based on its understanding of the user. In my case, I often interact with the most popular movies in my country, so that set is pushed to the top for me. The vertical arrangement of these sets are personalized based on user interaction.

My girlfriend and I share the same account so you can notice how K-drama recommendations find their way into my action-anime world!

How Netflix tackles cold-start!

A common problem with recommendations is the cold-start problem! What should we recommend to a new user, whom we have very little interaction data on? To which users should we recommend a newly released movie, since collaborative filtering requires users to watch a particular movie so that it can be recommended to others?

Netflix tackles these problems by using the simplest methods of recommendation. Popularity and trending based recommendations.

You’re a new user in India, logging in for the first time ever? Here you go with the ‘Top 10 in India today’, or the ‘Popular on Netflix’ Lolomo row.

Netflix also has Lolomo rows for ‘Trending’ as well as ‘New Releases’ to tackle cold-start.

Netflix has 2 major product types, Movies and TV shows. But what about retail websites that deal with endless products and multiple categories

Amazon

Amazon sells more than 12 million products.

Their mission revolves around being customer centric and to help people find and discover anything they might want to buy online. They are able to live up to this mission by improving discovery and helping people search items and related items fast.

This McKinsey report states that 35% of sales come from Amazon’s personalized recommendations.

Amazon has been a pioneer in building recommendation engines, having written multiple papers way before one could imagine. Here’s a patent filed by Amazon in the year 1998 patenting the use of item-to-item collaborative filtering. This patent is one of many more.

Here is the same paper published in 2003 on item-to-item collaborative filtering.

Recommendation Style

Retailers like Amazon mainly base their recommendation on a few important parameters. This includes the items that you purchase, the items in your cart and the ones you have viewed. This coupled with the classic collaborative filtering approach, helps build powerful recommendation engines, helping users to discover relevant items faster and subsequently drive sales. Along with simple popularity based recommendations Amazon also engages in highly personalized product suggestions to users.

I purchased an office chair a few days ago, and subsequently started seeing work-desks recommendations within the mid budget range as the chair I had purchased. It also made sense to recommend items that compliment each other, in this case a work-desk, rather than recommending other chairs.

Amazon is transparent enough to mention that the recommendations were inspired by my purchases.

Personalized recommendations

Amazon also doesn’t miss out on the opportunity to recommend items based on a user’s viewing history. Users spending time browsing through a product category, suggests a strong indication of purchase and all Amazon has to do is keep teasing the user with that product until a successful conversion takes place.

I spent about 30 minutes looking out for a good sports watch and was not too surprised to see recommendations of the watch I was viewing along with some really close candidates. (Well now you know what I want for Christmas)

That along with a few other items I was looking out for, for my home.

Collaborative Recommendations

Besides recommending items on the home page, Amazon also makes classical collaborative filtering recommendations, while a user is on a particular product. Again, Amazon is very transparent about their recommendations, notifying users about what others view after looking at a particular product as well as what others purchase.

Final notes

Organizations like Amazon and Netflix have the bandwidth and resources to work on such complex recommendation engines. Although they are good examples to learn from, building engines at this scale may not be feasible for mid and smaller size development teams.

Amazon and Google recognize this gap and have started building ready to use APIs to help businesses make better recommendations to their users. Google’s Retail API and Amazon Personalize are helping brands create real-time personalized user experiences without having an expert ML team, with some brands reporting about 56% improvement in product purchases!

Thank you for reading. Please share this if you’ve learnt something new today.