Music Recommendations @ Wynk
One of the thrills of being a data scientist are the possibilities to work with diverse domains. Each domain brings its own niche set of challenges and perspectives. This short note is about such exciting challenges and innovations that we are working on in the area of “Music Personalisation and Discovery” at Wynk Music.
Wynk is a music streaming app dedicated to bring high-quality music to millions of users in India. Increasing penetration of internet and smartphones in the fabric of Indian society has allowed for the growth of streaming music consumption at an unprecedented rate. Correspondingly Music Industry has been very appreciative of these developments and has been pro-actively using music streaming platforms to disseminate their content. This has led to a variety of music on our platform meant to be listened by a variety of users, which means a big responsibility of being a matchmaker between songs and user.
From the standpoint of music recommendation, it is a hard problem due to the following reasons:
1. Subjectivity of consumer (Heterogeneity of tastes)
Humans have varied musical tastes and different reasons for music consumption. Some of us like a soothing background soundtrack for reading a book or coding, while some of us need music while driving to reduce driver fatigue and stay alert.
It is also worth to note that criterion of liking/disliking of a song is not objective e.g., whilst a heavy metal song might be the epitome of relaxation at the end of a hard day at work for some, it might be a headache-causing nuisance for others. We have been listening to music in some form or other ever since we were born, in a way we have been conditioned by melodies. Thus our reaction to music is very innately related to the kind of persona that we really are.
At Wynk, we have a real-time data engineering pipeline that captures user’s behaviour throughout the app in real-time, be it playing tracks or skipping them, searching for specific songs/curated playlists, following artists, time of play etc. All of these events, along with the shared profile info is used in building a User Persona, which along with latent features from interactions with tracks becomes an input to our ML models. A lot of work is put in the understanding of user’s intent by crafting features from the implicit behaviour on app.
Understanding the user is a cross-organizational effort and with a unique opportunity of being backed by a leading telco in India, we are collaborating with Airtel’s digital innovation lab “Airtel XLabs” to solve problems to help with challenges like improving new user experience.
2. Scale of content
It has been estimated that there are ~97 million songs in the world, with a new song being created every 2 minutes. This underscores the gigantic size of inventory that we are working with. Thus the user-item interaction matrix is at a big data scale in both vertical (users) and horizontal (songs) dimensions, and sparse in nature (making traditional Recsys techniques like collaborative filtering exceedingly difficult). Also, new songs are added every Friday, for which we don’t have any user interaction information.
These issues are alleviated by either using different variants of modelling techniques with our secret spice. We heavily make use of our spark cluster to perform these tasks in a scalable manner.
3. Intricacies of content
Unlike a physical item available at an e-commerce site which has well-defined dimensions for utility and quality (e.g., an 8GB ram phone is better than 2GB), a piece of music, no matter how simple in its structure is incredibly complex in its perceived worth/value.
Not very long(few years) ago, the only way possible to algorithmically analyse a piece of music was through audio processing techniques using DSP wizardry. However, building on the advancements in deep learning, research community have been very busy in using deep neural networks to extract perceptual features from raw music signals.
At Wynk, we are laboriously working in these areas with a mission like zeal to algorithmically describe components of music using Deep Learning. A team of inhouse music experts help us along by creating training sets and testing the models.
The fuel that fires the aforementioned innovations is “data”. At Wynk we have a data-driven culture and a great deal of effort is put by the engineering team to maintain its immaculate quality, so that it can be directly used by Machine Learning team, enabling them to focus more modelling and design.
So, by now, you must have got a glimpse of the problems and innovations that we are working on @Wynk.
Currently, We are hiring for the Machine Learning team which will work on the problems that we have described. Feeling Excited? Check out the open positions at https://www.airtel.in/careers/opportunities