Customers, today, want nothing short of personal attention and service. From service to experience and offers, what matters is the personalization that can win customer loyalty. And when it comes to personalized offers, recommendation engine comes into perspective in the way it helps organizations learn what a customer likes and recommends the right product or item to create enriching user experience.

Recommender systems are built to predict product/s which a user might like. With recommendation systems promising more value in terms of personalization, more and more organizations are using recommendation engines across a range of applications including areas like Music, Books, Movies, etc. as well as business areas such as Netflix, Amazon and various other e-commerce applications that make use of different types of recommendation algorithms as per their needs.

Why do we need Recommender Systems?

Recommender system can act like a sales person suggesting different/similar products when we ask for a particular product and also like that sales person cross-selling products, for instance, if we go to a mobile store, the sales person there might come up with the recommendation of a temper glass or good earphones along with the mobile phone that we purchase at the store.

But when it comes to an online store, up-selling and cross-selling can be achieved only by using data around customer activity on the site, interactions and transactions that happen on the site, which then becomes a possibility by implementing recommender engines.

There are different types of recommendation algorithms, wherein we need to select the one or a combination of algorithms that would best satisfy our end goals.

What’s the mandatory feed for recommendation engine?

Data is the primary feed for recommendation engine – data related to products, users and items. Most importantly, customer interactions turn out to be valuable feed for the recommender system. For instance, when we seek to build a recommendation system for an ecommerce site, interaction data related to searches made by users, clicks on products, user visits, ratings, previous purchases, shopping cart items and favourite items, is held as key data feed for recommendation engine.

What are the different types of algorithms and how do they work?

Let’s look at some of the basic and traditional types of algorithms including:

  • Content Based Filtering
  • Collaborative Filtering
  • Hybrid Filtering

The appropriate filtering method can be chosen depending on the type of outcome required. Taking a closer look at these algorithms will help us understand how each of these filtering methods work.

Content Based Filtering

This is one of the simplest filtering methods, which can be used for recommending “similar” products. What does this word “similar” convey in this context?

Let’s take a simple movie recommendation – when a user has shown interest in a particular movie, then movies with similar characteristics are suggested to him. In case of Content Based Filtering, similarity of products is found. When a user has shown interest in a movie whose genre is Comedy, then the user will be recommended with more films from that genre. For this, the type of data needed encompasses information that would describe the product in terms of attributes of the product. For ex: size, colour, price, etc.

Collaborative Filtering

Collaborative Filtering method works by finding similarity between users. The idea behind this is that similar users will have similar liking towards a product. This method would not require lot of features that describes a product like Content Based Filtering, and for this method, we would only require historical data of a user’s interaction with the product.

Take Customer-1 and Customer-2 for instance. Customer-1 likes Orange Mango and Grapes and Customer-2 likes Mango, Grapes and Banana – reading the similar interests exhibited by these two customers, what’s apparent is that Customer-1 ought to like Banana and Customer-2 ought to like Orange. Collaborative filtering based recommender system works based on this notion.

Collaborative Filtering can be of the following types:

  • User Based: Recommend items based on similarity between other users’ purchasing behaviour
  • Item Based: Recommend items, based on similarity between items with which users have interacted

Hybrid Filtering

Hybrid Filtering is an approach where the two types mentioned above are combined. They can use data from user-item interactions and also characteristic data of users or that of items. This type of filtering can overcome drawbacks of the above two when used individually. Combing Content Based and Collaborative filtering methods can be done using multiple techniques.

How can you say if a recommendation engine is offering the best recommendation?

Here are some of the ways used to know if a recommendation engine is working at its best.

Online: A/B testing is implemented to check and monitor whether the user has interacted with the products recommended by the recommendation engine.

Offline: Model is tested before deployment using past data. The Train and Test method is used here.

Prior to building a recommender system, what raises into an indispensable preliminary is the need to pin down the business goal for which the engine is proposed to be built – recommender system for increasing customer engagement or for increasing CLV, for instance. More importantly, it is important to understand data at hand to make data drive robust recommendation engine.