Pandora head of data science for listeners Siddharth Patil said he regrets that his company failed to introduce a premium on-demand music offering until 2017, well after competitors like Spotify and Apple Music.
“I think user habits have evolved, and we were a little late to the game — so we were a little bit caught by surprise, I would say. We’re getting there now, but I wish we had entered the premium market maybe a year before we did.”
Adoption of an on-demand paid subscription music service was inevitable but was delayed in part by the process needed to secure song rights, Patil said in response to a self-described music fan who raised the question.
Patil spoke onstage Monday with VentureBeat founder Matt Marshall at VB Summit 2018, a machine learning gathering for executives held at the Seminary at Strawberry in Mill Valley, California.
Nearly 20 years old, Pandora was founded in 2000 as a personalized radio service and has 70 million monthly active users. Pandora Premium was introduced in March 2017. The music streaming service costs $10 a month and had 6 million paid subscribers as of this summer.
By comparison, a similar service like Spotify has more than 80 million paid subscribers, Apple Music more than 50 million, and Amazon Music Unlimited measures “tens of millions” of paid subscribers around the world.
Pandora was acquired by SiriusXM — a company with 36 million paid subscribers — in September in an all-stock $3.6 billion deal.
Essentially, Pandora recommends music to users through a combination of 75 different algorithms “that act in unison to decide what’s the best song to play,” Patil said. These range from a content-filtering algorithm that dives into a catalogue of 60 million songs to an algorithm trained to understand a person’s taste based on the 100 billion thumbs up or thumbs down signals Pandora has collected over the years.
The foundation for Pandora’s music recommendations is still the Music Genome Project, Patil said, which has been part of the company since the 2000s. And while Pandora applies a lot of machine learning to music recommendations, its personalization wouldn’t be possible without human curators.
“They attribute every song with more than 400 musicological features manually, and these features can be something like vocalists for the song: Is it smooth like Sade or is it gritty like Tom Waits? Or instruments: What are the instruments in the song, and what are the roles of the instruments? Or rhythm: How danceable is a song on a scale? And then melody: Is it highly ornamented like Mariah Carey?” he said.
Pandora is working today to extend its music recommendations to moods and emotions.
Patil spoke about the qualities needed to succeed in data science that go beyond being a numbers wonk. He speaks from experience. While still a student at MIT, he was part of a team of students who learned card counting to beat casino odds.
Books and movies like 21 about the team have made it seem as if they succeeded because of mathematical wizardry, and there were elements of that, but it was largely teamwork and trust that helped them beat the system, he said.
“Because you can’t go by yourself and beat the system, you have to actually play as a team. And so there was a lot of trust, some bookkeeping and boring stuff. There was a lot of strategizing because the rules kept changing, so there was just so much more to it, and very honestly when we look back it was about the team and the people and less about the math,” he said.
Likewise, Patil said being a data scientist requires human qualities that go beyond calculation.
“Data science is not about just applying fancy math techniques. You need a lot of grit. Without a sense of purpose, the day-to-day frustrations can be insurmountable,” he said.