The vision of the Freadom app is to provide first-generation young learners a fun and engaging medium to learn English. Freadom app offers a vast library of content to learn from – books, audio, video, games, etc. But, when a child launches the Freadom app, she will typically choose a story from a handful of options displayed on the screen.
These ‘displayed’ stories only make up a small portion of an otherwise large library. Since the user chooses what to consume based on what she is shown, the process of selecting these stories is very critical for successful learning outcomes. The problem of picking the right item to display as per the user’s need is a common problem across many apps such as Amazon, Youtube, Netflix and Spotify.
The machine learning based algorithms designed to tackle this problem is known as Recommendation engines. Personalized recommendation engines leverage methods from Computer Science and Statistics to construct tailor-made recommendations based on users’ past behavior. These models attempt to pin down important characteristics which drive interest and figure out users’ preferences based on past interactions.
The Freadom app was launched in 2016. The recommendation system has gone through multiple iterations since. However, 23% of sessions on the Freadom app end up with a child skipping all presented stories. Additionally, we know that a typical user starts four stories before completing one (on average). These data points clearly indicate that there is room for improvement.
We partnered with Golub Capital Social Impact lab lead by Dr. Susan Athey at Stanford University to develop machine learning models and refine content selection. Users tend to have differentiated tastes – Perhaps one child likes stories based on wild animals, while another one prefers stories about sports; A third child may like stories with pictures, while a fourth one prefers moral stories – thus providing bespoke recommendations can substantially boost their engagement and improve the learning pace.
Recommendation Engines -> User Success
Over the past 9 to 12 months our teams analyzed the data and developed various machine learning models in order to understand user preferences and subsequently recommend stories of interest. Collaborative filtering based recommendation engine gave the best results. This model was then implemented in the app as an AB test to evaluate its impact on user engagement.
We found an average increase of 30% in user consumption within a short time span of two weeks. The personalized recommendations have dual benefit – 1) Users consume a higher amount of content per session and 2) Users use the app more frequently as evident in the increase in the number of sessions per week.
More importantly, we observe that gains are significantly higher for students that used the platform frequently in the past. This result suggests that frequent or loyal users benefit to a larger degree as personalization engine has more data to crunch and hence make a more informed decision unlocking a virtuous cycle where more usage leads to better recommendations which in turn means more usage.Boosting-engagement-in-ed-tech-with-personalized-recommendations
A productive screen time app for ages 3 to 12, that focuses on improving English Language skills.
Online English classes for ages 5 to 12. Proven methods for children to improve academic performance and confidence.