Need of Recommender systems

Businessmen holding thumbs up in approval vector concept

35% of Amazon’s revenue is driven by its recommendation engine. Netflix estimates that its recommendation engine is worth $1 billion a year

A huge number of women are shopping for beauty products based on recommendations from friends, and we really look to be that friend - Emily Weiss

The recommendations of someone we trust are valuable input in making a variety of decisions such as the right outfit for a special occasion, whether to watch the latest DC movie (sorry DC), or deciding which restaurant to visit for that all-important date. Now why is this? We trust the judgment of a close friend or relative simply because we know that they know about our tastes and preferences and will suggest the best option considering these. In today's online environment, a similar function is performed by recommendation engines.

Recommendation engines are algorithms that use past data of customers in order to suggest what products they can buy. They are used by various online platforms such as e-commerce systems (Amazon), social networks (Facebook), streaming entertainment platforms (Netflix) among others. 35% of Amazon's revenue is driven by its recommendation engine. Netflix estimates that its recommendation engine is worth $1 billion a year, driving users to browse more content on the platform as opposed to canceling their subscription.

Recommendations from an algorithm are used as a targeted marketing tool in email campaigns or on websites. The benefits of using recommendation algorithms are listed below:

  • Revenue: Increased revenue by increasing the average cart size per order through algorithms that are optimized for high conversion rates
  • Improved customer experience: Users may revisit a set of recommendations based on their history to explore better deals on products. The algorithm can generate improved recommendations based on this data and thus improve customer experience and customer loyalty
  • Discovery: Features such as "People you may know" on Facebook, "Customers who bought this also bought" on Amazon, "You may also Like" on iTunes enable users to access products and content that they might not have otherwise

Broadly speaking there are three types of recommendation engines that are commonly used today:

  • Content-based systems: These utilize a description of a product and a profile of the user's preferred choices. The features of products are mapped with the features of users to obtain user-product similarity. The top pairs are given as recommendations by the algorithm. The algorithm's focus is on recommending a product that is similar to those a user liked in the past, or is evaluating currently. A major drawback of pure content-based systems is their inability to detect inter-dependencies or complex behaviors of the variable involved.
  • Collaborative filtering: These algorithms utilize user behavior to recommend products. The algorithm works on the assumption that if George likes products A, B & C and Martha likes products C, D & E; then they will like similar kinds of products in the future and therefore George should like products D & E and Martha should like products A & B. Collaborative filtering as an approach is extremely useful since it does not require machine analyzable content and can thus recommend complex items without having to know all features of the item.
Blog1_pic_User_Recommendation

  • User-User collaborative filtering - The algorithm searches for customers with similar profiles and recommends products based on what other users with a similar profile have chosen. While this is extremely effective, it requires large number of computations and is therefore time and resource - intensive
  • Item-Item Collaborative filtering - The algorithm seeks out similar items and recommends them to the user. This requires far fewer resources than user-user collaborative filtering as it does not require that similarity be computed amongst all the users

  • Hybrid Recommendation systems: These combine the approaches of the collaborative filtering and content-based systems. Hybrid systems are typically implemented by adding the content-based system capabilities to a collaborative filtering model or vice versa and unifying them into a single model

 

Recommendation systems are here to stay and will play an increasingly important role in helping customers navigate the complexity of choosing the right options in a sea of choices.

Next Steps

If you're interested in implementing a recommendation system for your firm, do reach out to us at contactus@devisemath.com and we would be delighted to assist you.