Limits of the Spotify Recommendation Algorithm
In the Insider article, writer Drew Austin discusses the ways in which recommendation algorithms for streaming platforms, particularly Spotify, have failed to reach perfection despite high expectations a decade ago. In essence, these algorithms attempt to understand users’ preferences by analyzing their streaming history and recommending content of a similar nature (similar genre, artist, director, etc.), thereby creating a sort of feedback loop. The problem with this approach, Austin argues, is that it assumes one’s personal taste is inherent and unchanging rather than a constantly evolving product of their surrounding environment. In fact, one’s preferences are heavily influenced by culture and personal relationships; as the writer’s friend once put it, “the way you find out what’s cool is someone cooler than you telling you it’s cool.” In short, Spotify and other platforms’ recommendation algorithms are not highly effective at understanding users’ preferences because they seldom account for these external factors properly.
Some critics even suggest that the true goal of these algorithms is to actually reshape user preferences to fall into more clearly defined archetypes for the sake of maximizing revenue. The logic goes something like this: universally popular content is only created so often, but streaming platforms have a wide variety of content for different specific interests, so if users’ interests are reshaped into more niche categories, then they will likely spend more time on the platforms and serve as better targets for advertisers. Whether intentional or not, the effect of this business model for streaming platforms is highly reminiscent of the Long Tail phenomenon.
As we learned in lecture, the popularity of music (as well as other content) often follows a power law distribution as a result of rich-get-richer dynamics, so few songs tend to become extremely popular, while many more songs see little success. This distribution of popularity gives rise to two types of business models, one in which a company carries a select inventory of the most popular items, while the other model calls for carrying a diverse inventory of less popular items. The latter allows online retailers and streaming platforms to maximize revenue by catering to a wide array of specific interests, since power law distributions have a long tail where a significant portion of the overall demand resides. As a result, there exists an incentive to guide users toward niche genres and interests, which may help explain the ‘ineffectiveness’ of Spotify’s (and other platforms’) recommendation algorithms. Perhaps they are designed in order to mitigate the rich-get-richer effects, thereby pushing more of the demand into the long tail and improving the outcomes of their business model. However, as the article details very well, human desires are heavily influenced by who and what is around us, meaning these rich-get-richer effects are almost insurmountable. So while recommendation algorithms of this nature may boost company revenue, they are unlikely to accurately capture the ever-changing interests of its users and are unlikely to make significant improvements any time soon.