Recommendation engines are ubiquitous these days, they are everywhere! Whether you listened to a recommended podcast on Spotify, browsed YouTube during your lunch break, or scrolled through your Instagram explore page, you were the subject of a recommendation system today.
35% of purchases on Amazon come from product recommendations
With that amount of added value, it is clear that recommendation engines are key success factors. With a growing interest in these novel systems in both academia and industry, one thing is certain and that is that recommendation systems are here to stay. One question naturally arises – how do these systems know exactly what I want to see, listen to, and buy; all seemingly by forces of magic?
Defining a recommendation engine
Recommendation engines can be broadly defined by technology systems that encapsulate the ability to recommend items that are relevant to a specific entity based on patterns that can be found in recorded data.
A recommender system is a filtering tool that uses a series of algorithms, data analysis and Artificial Intelligence (AI) to make recommendations online.
Common examples of recommendation systems are:
- Netflix recommends movies and shows to its subscribers
- Amazon recommends products to its customers
- Spotify recommends podcasts, playlists, and songs to its listeners
In each of these cases, the recommenders mentioned use interaction information captured in each of these apps for every user in order to generate a picture of that user's preferences and specific taste. Whether it be the movies and shows you have already watched, the products you have bought in the past, or the songs you listen to the most.
Naturally, two questions arise:
- What happens if I'm a new user on the platform and the system has no gathered information for me?
- How can a recommendation engine relate items to each other?
Multiple recommendation engine types
There exist three main types of recommendation engines:
- Collaborative filtering
- Content-based filtering
- A hybrid model of the two
The main focus of collaborative filtering is to collect and analyze user behavior data as well as user preferences and activities. The idea behind using this method is that if two people show similar behavior on a specific platform, they will have a high probability of enjoying the same sets of items. For example, if you behave very similarly to Joe on Netflix, you will probably like the same movies and shows as Joe.
The main advantage of collaborative filtering is that the content of items themselves doesn't need to be analyzed and understood by the algorithm. This is particularly useful when such content is difficult to compare digitally (movies, books, podcasts, and so forth).
On the other side of the spectrum, content-based filtering will compare items to one another in order to recommend items similar to the ones a user has consumed in the past. This approach is extremely useful when two items are digitally comparable by the recommendation engine.
A major drawback of content-based filtering is the fact that it is very limited to users' past interactions with the system. For example, if a viewer has only watched horror movies on a given platform, a content-based approach will recommend only horror movies even though the viewer might also like romantic comedies. This kind of methodology also struggles when it comes to recommending items for new users who have no previous interactions with the recommendation engine.
Hybrid models are a combination of collaborative and content-based filtering. This methodology leverages both types of data and is known to outperform the two other approaches.
As an example, Netflix is a great example of a hybrid recommender engine model: it takes into account both the interests of the user (collaborative filtering) and the features and description of the movie or show (content-based filtering).
Different kinds of recommendations
There is an immense advantage of leveraging both content-based and collaborative filtering. In fact, using these recommendation engine methods adequately allows you to create two distinct types of recommendations: customer recommendations and order recommendations.
To anticipate the future needs of customers, item recommendations can be made for each customer. This is the most common way to serve recommendations.
Using this technique, recommendations are made for each customer individually.
On the other hand, recommendations can be made with regards to existing, tentative, or recurring orders. This methodology allows the up- and cross-selling of products as it brings together items that complement each other. For example, when a user is buying a new phone, a complementary recommendation could be to add a relevant phone case or screen protector to the order.
Recommender systems are here to stay
The added value of leveraging novel recommendation engines is immense. In fact, they are the key drivers of how many companies make a large chunk of their profit, whether it be large tech giants such as Netflix or Amazon, smaller or medium-sized local e-commerce stores, or even B2B vendors.
Recommendation systems for B2B?
While recommendation engines are mostly leveraged in e-commerce, the methodology they employ can be generalized to other use cases successfully! For example, helping sales representatives with an internal tool to give customer- and order-specific recommendations can have an immensely positive impact on up- and cross-sales. This is extremely useful if the salesperson has to deal with a large set of clients and huge product inventories that are un-memorizable.