Recommendation algorithm, its types & how big tech companies leverage its power

93
recommendation algorithm of Netflix | iTMunch

We, as humans, make purchasing decisions every day. Most of the products we buy are either something we like or something that our friends like. A study by Harvard Business School assessed a unique set of data from Cyword (a South Korean social networking site) and found that friends indeed influence user’s purchases. This social effect was found positive for 40% of the users and on average, it translated into a 5% boost in revenue [1]. This is so because we tend to buy products that have been tried, tested, loved and recommended by our friends. In today’s digital era, any e-commerce marketplace you visit has some sort of recommendation algorithm. If this recommendation engine, algorithm or system is set up properly, it can accelerate Click Through Rates (CTRs), conversions, revenue and other important Key Performance Indicators (KPIs). Additionally, recommendation algorithms can also have a positive effect on the user’s shopping experience, which ultimately contributes to your revenue and KPIs that are important for customer satisfaction and retention.

In this blog, we will try to explain what is a recommendation algorithm, how it works, what are its types and how different big tech companies, like Amazon, Facebook and Netflix, use these algorithms to their advantage. Let us begin by understanding the basics.

What is a recommendation algorithm?

Recommendation algorithms are, at the core, filtering tools that use data provided by users to recommend them similar or the most relevant items. These are machine learning systems that continuously learn from user behavior and help them discover products that might be the most suitable for them.

To break it down in easier terms, a recommendation algorithm works like a salesman in any offline showroom you visit. You go to a showroom and ask a salesman for a product you’re looking for. If you are a regular customer, he probably recognize you and might be acquainted with your taste and preferences. As a result, he gets you exactly what you want in addition to a few related (probably better) items you might consider purchasing. You keep going back to that showroom because you always get what you want or end up buying something better.

Recommender algorithms are like those salesmen who know what you are looking for and what you’re likely to buy, based on your search history, shopping history and preferences. Recommendation systems aren’t just used in online shopping or e-commerce platforms like Amazon for product recommendations. A few other examples of a recommendation algorithm in play includes suggestions for TV shows and movies on streaming services like Netflix, recommended videos on YouTube, Facebook newsfeed and music on Spotify. 

For these algorithms to work, a huge amount of user data is collected by the platforms. But what exactly is the data collected by them? Let’s find out.

Data collected by platforms for the recommendation algorithm

For any recommendation engine, data is the oil that fuels it up and keeps it running. Be it Amazon, Netflix or Spotify, any tech platform or online shopping website collects a huge amount of data to understand the user’s preferences. Keeping these preferences in ‘mind’, the recommendation engines predict what the user might like and end up spending money on. The recommendation algorithm captures and uses three types of user data:

User Demographics Data

Demographic data includes your personal information such as location, age, gender, education, profession, etc.

User Behavior Data

All the necessary information about how you, as a user, interacts and engages with products showcased online. It takes into account the ratings you provide, the product listings you click on, the products you ‘Like’ and the products which you have actually purchased.

Product Attributes

This data is directly related to the products you interact and engage with. Information like genre and cast (in case of TV shows, movies and books), cuisines (in case of food), type of clothing (in case of an e-commerce website for apparel) and so on, are taken into account.

SEE ALSO: How the Facebook ad algorithm & ad-delivery system skews outcomes

How we provide data to the recommendation algorithm

We feed data and information in the recommendation algorithm by two major methods: Explicit Ratings and Implicit Ratings.

Explicit Ratings are the ratings actually provided by users on the products they have bought online. These ratings directly indicate the preferences of the users. Examples include star ratings, product or service reviews, likes, following and feedback. Relying solely on Explicit Ratings isn’t ideal because these are hard to get and users don’t always rate the products they purchase.

Implicit Rating is based on understanding users’ interaction with the products listed. It is about inferring user behavior and is easier to get as compared to Explicit Ratings. Examples of Implicit Ratings are product views, clicks, the devices you use and location.

Different types of recommendation systems:

Content Filtering Algorithm

Content-based filtering algorithm is based on the users’ preferences and interactions. It is based on the idea that if you like a particular item, you might also like a ‘similar’ item. By similar, we mean products attributes that are common in certain products. The recommendations presented to the user are based on the common attributes of the items. 

Here’s another way to understand content-based filtering. This filtering method depends on two things: the description of an item and the user’s preferences. In a content filtering algorithm, keywords are used to describe items and a user’s profile is built to understand what they like (or dislike). ‘Products similar to this’ is a typical example of this type of filtering.

One challenge faced in content-based filtering is learning user preferences from their actions about one segment of products (like clothing) and applying or replicating it across other product types (like footwear). ‘Cold start’ is another issue faced by this type of recommendation algorithm. Cold start happens when a recommendation system can’t draw inferences for a query due to insufficient data. It occurs when there are missing sets of data or no data to begin with in the first place.

Collaborative Filtering Algorithm

Collaborative filtering algorithm collects and analyzes information on user’s behavior, their preferences or activities and predicts what might like based on similarity with similar users. It revolves around making recommendations based on users with similar traits and tastes. A big advantage of this filtering algorithm is that it isn’t dependent on machine analyzable content and hence is capable of recommending complex items accurately. For example, it is used to recommend movies without needing an ‘understanding’ of the item itself.

It is simply based on the idea that if an individual A likes items 1,2 and 3 and individual B likes 2,3 and 4, then they have similar preferences and A would like item 4 and B would like item 1. This recommendation system is typically used to recommend items to identify how one product might go well with another product. A ‘Next buy’ suggestion is another usage of collaborative filtering algorithm.

Collaborative filtering recommender can be further classified into the following:

User-User Collaborative Filtering

This type of collaborative filtering recommender focuses on finding similar customers and provides product recommendations based on what their lookalike has picked. This is truly an effective filtering method but consumes a lot of time, especially for platforms with a huge customer base.

Item-Item Collaborative Filtering

This type is similar to user-user filtering, but instead of searching for customer lookalikes, item lookalikes are searched. Once an item lookalike matrix is created, it is easier to recommend similar items to customers that have bought any items from the online store. Far less time and fewer resources are needed for this algorithm to set up. Therefore, the algorithm takes very less time than the user-user collaborative recommendation algorithm to start making recommendations. Amazon uses item-item collaborative filtering on its website to show related products.

Hybrid Recommendation Algorithm

A hybrid recommendation system is a recommender that harnesses the power of both collaborative and content filtering algorithms. This can be implemented by making collaborative-based and content-based predictions individually and finally combining them. It can also be done by adding collaborative-based capabilities to a content-based approach or vice versa. The content-based recommender takes charge when no user data is present and once sufficient data is collected, collaborative filtering methods can be leveraged.

In today’s digital world, the user is spoilt for choice as it is bombarded with product listings and information, and this can be quite overwhelming. Implementing any of the above mentioned recommendation algorithms will surely contribute to your revenue and other KPIs. Now that we know the various types of recommendation engines, let us dig deeper and understand how the big tech platforms like Amazon and Netflix put these algorithms to use.

How Amazon’s recommendation engine works

Research conducted by McKinsey & Company found that 35% of what consumers buy on Amazon happens because of product recommendations [2]. Amazon says collaborative filtering algorithms are the most common and best suited way to do product recommendations in an e-commerce ecosystem. The online shopping giant basically predicts your tastes and preferences based on other similar customers. The approach Amazon has chosen to go about recommending products is item-item collaborative filtering, which means it offers product recommendations based on correlation between products and not customers. With item-item filtering, the algorithm reviews the customer’s purchase history and derives a list of related items.

Using this data, Amazon’s recommendation system works with the following:

  1. Products related to an itemamazon recommendation engine | iTMunch

2. Customers who viewed other similar itemsrecommendation algorithm of an ecommerce website Amazon | iTMunch

Amazon also works with ‘Recommended for you’, ‘Frequently bought together, Customers also considered’ recommendations.

SEE ALSO: What tech companies are doing to control the spread of fake news

The recommendation system of Netflix

Netflix is a subscription-based content streaming platform that deploys a hybrid recommender. The same McKinsey research also found that % of what consumers watch on Netflix comes from watch-next recommendations [3]. At first, when you join Netflix, the platform asks you to pick a few titles that you’ve seen in the past and liked, to give the recommendation system a jump start. In the beginning when you watch movies on the platform, your initial preferences will dominate the algorithm to show you the titles you’re like to enjoy. Netflix presents its recommendations in rows (Trending Now, Continue Watching, New Releases, Critically-acclaimed Witty US TV Shows, etc.). This is where the content filtering algorithm is at play. As you continue to watch titles on Netflix over time, the tiles you’ve watched more recently will supersede the titles you’ve watched in the past to keep you going and their recommendation more relevant.

After you’ve watched a couple of titles, the collaborative filtering algorithm takes over. At this point of time, the Netflix recommender starts recording and analyzing your interactions with the service, like viewing history and ratings you’ve provided. To present more personalized recommendations, it also looks at things such as what time of the day you watch, the duration you watch for and the devices you’re watching Netflix on. Moreover, it looks at things like other members with similar preferences and tastes on its service and information about the titles itself like categories, genres, release year, cast, etc. Surprisingly, Netflix doesn’t take demographic information, like age and gender, about you into consideration as a part of its algorithm. 

Following are the recommendations provided by Netflix:

1. Recommendations based on genres you’ve liked before

netflix recommendation engine | iTMunch

2. Recommendations based on what titles you’ve liked before

Netflix recommendation system | iTMunch

Facebook’s recommendation algorithm and social graph

Social networking giant Facebook also uses collaborative filtering recommender systems to help its users discover people and items that are the most meaningful to them. It uses CF to recommend groups, pages, games, events and much more. It uses previous item ratings of similar people to predict how they would rate an item in the future.

Apart from the recommendation system, Facebook’s revenue model is also largely based on creating a user’s social graph. Facebook captures a huge amount of data about you when you create a profile on its platform. It knows your name, your age, your gender, where you live, your education, where you work and who your friends and family are. It also knows abut your interests, tastes and preferences. The social media company uses all of this data and much more to create your social graph. It believes that you’ll probably like what your friends and family like. While a collaborative filtering algorithm helps the platform itself in providing you relevant recommendations, a social graph is created to deliver an overall personalized experience. Social graph is also created by Facebook and other social networking platforms to allow advertisers reach out to their target audience. You can read more about social graphs in this blog.

A recommendation algorithm is essentially the system that keeps you suggesting what to watch next on Netflix, the next video you should watch on YouTube and which song you should play on Spotify. Recommendations systems, engines or algorithms guide you towards the items you are most likely to buy or watch or listen to. The algorithm analyzes what you as a user are looking for and suggests related products that you are most likely to click on. There are three types of recommendation systems – Content-based filtering, collaborative filtering and hybrid filtering algorithms. The reason behind the big tech companies putting huge chunks of money on continuously developing these recommendation systems is to keep you going on their platforms and ultimately increase their revenue.

SEE ALSO: The Social Dilemma: Social Graph VS Interest Graph

For more updates and latest tech news, keep reading iTMunch

Sources

[1] Harvard Business School (2009) “Do Friends Influence Purchases in a Social Network?” [Online] Available from: https://hbswk.hbs.edu/item/do-friends-influence-purchases-in-a-social-network [Accessed December 2020]

[2] [3] McKinsey & Company “How retailers can keep up with consumers” [Online] Available from: https://www.mckinsey.com/industries/retail/our-insights/how-retailers-can-keep-up-with-consumers [Accessed December 2020]

Subscribe to our Newsletter!