The Not-So-Secret Sauce Powering Our Recommender System
Data Science

The Not-So-Secret Sauce Powering Our Recommender System

Over 2 million apps are available on the Google Play Store and App Store. The average number of apps used daily is 9 and monthly is 30.1

So the question is:

Does your app figure in the top 9 – or at least the top 30 – apps among the millions of apps available?

Is your primary focus on just acquisition or are you working on retention?

Better Experience, Better Retention

It will come as no surprise to any product manager or marketer that improving the customer experience will lead to: better customer satisfaction rates, increased customer retention, higher revenues, and lower costs.

As per a report by American Express, 7 in 10 Americans are willing to spend 13% more on better customer experience.2 Further, a research by Temkin group claims that a moderate increase in customer experience could lead to an increase of $775 million for a company with $1 billion revenue.

Revenue Increase Customer Experience

Source

If your app offers hundreds, or even thousands, of products, and you want to improve the customer experience, you face some monumental challenges, namely:

  • How can you offer better awareness and easier discovery of products?
  • How can you make engagement smarter and hyper-personalized?
  • How can you make your app sticky enough to become a habit?

The Power of Recommendations

This is where personalized recommendations based on user interaction can become immensely useful to improve customer experience.

Here are a few examples of the impact that recommendations can have on business KPIs:

  • 80% of content consumed on Netflix is due to recommendations.3
  • 35% of Amazon.com’s revenue is generated by its recommendation engine.4
  • 60% of video clicks on Youtube’s homepage are attributed to recommendations.5
  • The increase in Spotify’s MAUs in the face of competition is attributed to its recommendation engine.

The 2 Types of Algorithms Behind Recommendations

So how do these companies power their recommendations?

Key weapons in the armory of these companies are powerful machine learning algorithms that understand user behavior and suggest appropriate personalized products.

And how do these algorithms work?

There are traditionally two types of algorithms:

  • Content based
  • Collaborative filtering based

There are also hybrid methods that combine content and collaborative filtering to arrive at recommendations.

  1. Content Based
    A content based recommender tries to extract insights from meta information about the product to learn the preferences of the users and suggest appropriate products.

    Suppose you want to predict which movies to suggest to a user. You could very well look at the type of movies the user has viewed in the past to make a suggestion drawn from previously watched genres, actors, series, and more.

    For example: if the user has viewed an Avengers movie, you could suggest the next Avengers movie in the series. You could also infer that this user may be interested in superhero movies or action movies in general.

    Pros and Cons
    One of the key advantages of using content based recommender is the ability of recommending new and unpopular items based on a user’s past interests.

    But its biggest drawback is it never recommends items outside the user’s content profile. Just because a user has viewed content in the action genre, should the user only be recommended other action movies?

  2. Collaborative Filtering
    This method tries to leverage the underlying hidden insight in a user’s behavior.

    The manner in which users interact with the app reveals a lot about their preferences. The key idea behind Collaborative Filtering (CF) is that similar users share the same interests and that similar items are liked/purchased by users. How the users or items are paired can be explicit (based on ratings/likes) or implicit (based on purchases, views).

    Collaborative Filtering

    Pros and Cons
    Though CF may face difficulty recommending new items, it is able to recommend items outside a user’s content profile.

    But even more, CF does not need additional meta information to build the recommendation system, making it frictionless to build from a marketer’s standpoint. The marketer need not worry about what meta information is needed.

Why We Chose to Use CF

At CleverTap, we wanted our recommenders to be relevant and also lead to much better discovery of products. This would lead to better diversity in product sales/views in the long run.

We use collaborative filtering to power our own recommendation system. Though there are many CF algorithms, we chose item2vec, which is based on a popular algorithm called word2vec used for Natural Language Processing (NLP).

Check out these posts on neural network and item2vec, if you aren’t familiar with the underlying algorithms.

The Goals of Our Recommendation System

Before we started running trials (proof of concept/POC) with our clients across industries, we had four goals in mind for our recommender system:

  • Relevancy: Suggest relevant products
  • Novelty: Delight users with surprising recommendations
  • Serendipity: Wow users with surprising options, but also trigger an impulsive behavior leading to purchase/consumption
  • Diversity: Increase diversity in product sales

Validating recommender systems based on the above goals can be tricky since back-testing the efficacy of recommendations is fraught with danger.

Backtesting evaluates the efficacy of the model by discovering how it would play out using historical data.

The problem with using historical data is that the recommendations generated have never been served. Hence, relying solely on evaluation metrics of models from backtesting can be risky.

The best option was to try it live in production environment to measure the impact. And we weren’t wrong. We started to observe immediate impact on engagement and conversion statistics. This behavior was observed in multiple industries that used our system. This gave us sufficient confidence to use the algorithm in production. Initial results do indicate the performance in production is either better or equal to the performance observed in the POC, which confirmed that our choice of algorithm would achieve the goals we wanted.

Conclusion

In this post, we discussed the need for a recommendation system to enhance the customer experience leading to better retention and engagement, and also described how we used collaborative filtering and specifically item2vec to achieve our goal.

At CleverTap, we aim to constantly innovate and improve our capabilities. In keeping with this spirit, we plan to use hybrid recommenders in future iterations to combine the meta information of items and/users and make the recommendation even stronger.

Just as the brain constitutes just 2% of the body weight, similarly the machine learning algorithm that powers the recommendation constitutes a small portion of the architecture.

If you’re interested to know how CleverTap’s recommender system is engineered to give a seamless experience from the marketer to the app user, stay tuned for our next post.

8,000+

$2 Billion

1 Billion

20 Million