Algoholics: How Streaming Platforms Know Our Taste Better Than We Do
by Zachary Marshall
Isn't it impressive how Spotify, Pandora, or Apple Music can supply their consumers with seemingly endless new music recommendations? Data has become the name of the game for any company trying to market a product or service. How information is gathered, interpreted, and then curated towards the consumer is crucial for companies' success. This is why there has been a rapid increase in the development of algorithms. Spotify’s algorithm BART (Bandits for Recommendations as Treatment) is no pushover to say the least. BART comes well prepared with multiple tools to give you the perfect recommendation.
The first component of the BART’s algorithm is a reward model. An algorithms reward model has the role of predicting the user response to a recommendation when given the context. Bart uses your listening data as a benchmark for making successful recommendations. For instance, BART takes one artist and compares it to similar artists that fall along similar genres and themes. BART’s reward model assesses a wide range of music data points and ranks their probability of success. After analyzing countless samples of data BART can rank these data points and base its recommendations off that information. If you are listening to Snuff by Slipknot, BART knows it can have success recommending music from other mid 90’s heavy metal bands. Bart can distinguish Korn as having a higher success rate than the Grateful Dead. This distinction also does not rule out the Grateful Dead from being recommended. Bart acknowledges that low percentage recommendations can still succeed.
The next component of BART is a stochastic policy. A stochastic policy is responsible for finding new music to recommend. A stochastic policy is designed to seek out uncertain recommendations that a user could potentially like. This stochastic policy uses a simple algorithm to randomly choose between exploring or exploiting. Exploration is the process of finding a new region of music to search through. Exploitation is the process of updating previously successful recommendations. BART’s stochastic policy could choose to explore new content by using the reward model mentioned above. BART could also choose to exploit its preexisting data by recommending a new album from a user's favorite artist. As the algorithm gains more confidence in the quality of items, it explores them less.
Finally, BART implements a propensity score. In statistical matching, a propensity score attempts to estimate the effect of a suggestion by accounting for what brought you to that suggestion. In layman's terms, BART measures a first song and a recommended song based on that first song. BART compares the relation of those two songs across thousands of users. BART wants to know how strong the connection between A and B is through your listening habits. If BART sees a decreasing success rate of recommendations between Sublime and Portugal, the Man, it will reevaluate the strength of the connection. BART will decrease the Portugal, the Man recommendations it makes to known Sublime fans. BART wants to know how to make the best recommendations to users, and in order to do that it must know everything about a users listening habits.
The modern digital ecosystem revolves around digital curation and commodifying information. It stands to reason that these algorithms have become so developed for the sake of collecting and capitalizing on gathered data. While we may not be able to particularly control our favorite music platforms methods gathering this information we can at least understand the technology behind it.