As someone who spends far too much time on the TikTok app, I have a rudimentary grasp on how algorithms work. I’m aware that big tech uses algorithms to direct and refine content, in the hope that we always see something we like and therefore continue to return to the platform in question.
Register for free »
Get started now for absolutely FREE, no credit card required.
I did some research and discovered that Spotify, the digital music service, uses algorithms to tailor its site to its 381 million monthly active users (which includes 172 million paying subscribers… read those figures and weep). An algorithm is a finite sequence of well-defined instructions, typically used to solve a class of specific problems or to perform a computation. Spotify creates custom listening experiences for each user. It uses three key machine learning (ML) algorithms to study behaviour on the app. Collaborative filtering creates a back end profile based on a user’s personal taste. Natural language processing analyses music by scanning a song’s metadata and mentions across the web, to help group artists into clusters and connect to relevant listeners. Audio models analyse raw data audio and recommend non-popular new songs, identifying similarities in signature, key, mode, tempo, loudness, tone, mood etc to create a sound profile, in order to match a new song to an existing playlist.1Age, gender and location are also factored in, to ensure no user’s home screen is the same as another’s.
I learned that Spotify’s home screen is governed by an AI system called Bandits for Recommendations as Treatments (BaRT). My Spotify home screen offers me ‘shelves’ that include: playlists I have recently listened to, a new track from one of my favourite artists, and playlist recommendations listed under enticing (to me) headings such as ‘It’s Christmaaaas!!!’, ‘Indie/Alternative’ and ‘Pop a jumper on’. The BaRT system used to curate this home screen combines elements of ‘exploit’ and ‘explore’, as outlined by Spotify Research Director Mounia Lalmas-Roelleke.2 ‘Exploit’ analyses everything a user does on Spotify, including which songs the user likes and which songs they skip, in order to learn about their listening tastes. If you spend less than 30 seconds listening to a song then BaRT thinks you don’t like it, and vice versa. ‘Explore’ compares a user’s activity to that of the rest of the Spotify member hip, analysing trends so that the site can suggest playlists or new artists that the user will probably like.
But while the algorithm can learn a user’s behaviour, it won’t necessarily know how or why the user has those preferences. In his article ‘We might need to make our own playlists again. Our lives might depend on it’, Dave Holmes writes: “As well as the algorithm knows my behaviour, it doesn’t know me… It cannot set you up for self-discovery. It just gives you more of what it knows you like… and pushes you to be more of who it knows you are. It can’t change you. You won’t evolve.”3The message here is enjoy the power of data, utilise it where you can, and understand its limits.
In this issue of World Pipelines, Kongsberg Digital writes about the development of digital twin technology, creating a virtual representation of a real asset and its dynamic performance. Kapil Mukati and Brian Sidle write: “By combining physics-based models, data science approaches and cloud scalability, the Kognitwin digital twin helps operators streamline and test hypothetical scenarios to maximise performance. This allows for improved prediction of impact options and decision-making, leading to overall enhanced productivity, improved safety levels and more sustainable operations.” Turn to p. 15 to read the article.
A continuously learning digital twin model can provide interpretable ML insights and bring about improved decision-making in the pipeline sector. What you decide to play over your truck’s sound system shouldn’t be such an exact science.