recommendations | Daniel Antal

Recommendation Systems: What can Go Wrong with the Algorithm?

Thu, 06 May 2021 07:10:00 +0000

Traitors in a war used to be executed by firing squad, and it was a psychologically burdensome task for soldiers to have to shoot former comrades. When a 10-marksman squad fired 8 blank and 2 live ammunition, the traitor would be 100% dead, and the soldiers firing would walk away with a semblance of consolation in the fact they had an 80% chance of not having been the one that killed a former comrade. This is a textbook example of assigning responsibility and blame in systems. AI-driven systems such as the YouTube or Spotify recommendation systems, the shelf organization of Amazon books, or the workings of a stock photo agency come together through complex processes, and when they produce undesirable results, or, on the contrary, they improve life, it is difficult to assign blame or credit.

This is the edited text of my presentation on Copyright Data Improvement in the EU – Towards Better Visibility of European Content and Broader Licensing Opportunities in the Light of New Technologies - download the entire webinar’s agenda.

Assigning and avoding blame.

If you do not see enough women on streaming charts, or if you think that the percentage of European films on your favorite streaming provider—or Slovak music on your music streaming service—is too low, you have to be able to distribute the blame in more precise terms than just saying “it’s the system” that is stacked up against women, small countries, or other groups. We need to be able to point the blame more precisely in order to effect change through economic incentives or legal constraints.

This is precisely the type of work we are doing with the continued support of the Slovak national rightsholder organizations, as well as in our research in the United Kingdom. We try to understand why classical musicians are paid less, or why 15% of Slovak, Estonian, Dutch, and Hungarian artists never appear on anybody’s personalized recommendations. We need to understand how various AI-driven systems operate, and one approach would at the very least model and assign blame for undesirable outcomes in probabilistic terms. The problem is usually not that an algorithm is nasty and malicious; Algorithms are often trained through “machine learning” techniques, and often, machines “learn” from biased, faulty, or low-quality information.

Outcomes: What Can Go Wrong With a Recommendation System?

In complex systems there are hardly ever singular causes that explain undesired outcomes; in the case of algorithmic bias in music streaming, there is no single bullet that eliminates women from charts or makes Slovak or Estonian language content less valuable than that in English. Some apparent causes may in fact be “blank cartridges,” and the real fire might come from unexpected directions. Systematic, robust approaches are needed in order to understand what it is that may be working against female or non-cisgender artists, long-tail works, or small-country repertoires.

Some examples of “undesirable outcomes” in recommendation engines might include:

Recommending too small a proportion of female or small country artists; or recommending artists that promote hate and violence.
Placing Slovak books on lower shelves.
Making the works of major labels easier to find than those of independent labels.
Placing a lower number of European works on your favorite video or music streaming platform’s start window than local television or radio regulations would require.
Filling up your social media newsfeed with fake news about covid-19 spread by some malevolent agents.

These undesirable outcomes are sometimes illegal as they may go against non-discrimination or competition law. (See our ideas on what can go wrong – Music Streaming: Is It a Level Playing Field?) They may undermine national or EU-level cultural policy goals, media regulation, child protection rules, and fundamental rights protection against discrimination without basis. They may make Slovak artists earn significantly less than American artists.

Metadata problems: no single bullet theory

In our work in Slovakia, we reverse engineered some of these undesirable outcomes. Popular video and music streaming recommendation systems have at least three major components based on machine learning:

The users’ history – Is it that users’ history is sexist, or perhaps the training metadata database is skewed against women?
The works’ characteristics – are Dvorak’s works as well documented for the algorithm as Taylor Swift’s or Drake’s?
Independent information from the internet – Does the internet write less about women artists?

In the making of a recommendation or an autonomous playlist, these sources of information can be seen as “metadata” concerning a copyright-protected work (as well as its right-protected recorded fixation.) More often than not, we are not facing a malicious algorithm when we see undesirable system outcomes. The usual problem is that the algorithm is learning from data that is historically biased against women or biased for British and American artists, or that it is only able to find data in English language film and music reviews. Metadata plays an incredibly important role in supporting or undermining general music education, media policy, copyright policy, or competition rules. If a video or music steaming platform’s algorithm is unaware of the music that music educators find suitable for Slovak or Estonian teenagers, then it will not recommend that music to your child.

Furthermore, metadata is very costly. In the case of cultural heritage, European states and the EU itself have been traditionally investing in metadata with each technological innovation. For Dvorak’s or Beethoven’s works, various library descriptions were made in the analogue world, then work and recording identifiers were assigned to CDs and mp3s, and eventually we must describe them again in a way intelligible for contemporary autonomous systems. In the case of classical music and literature, early cinema, or reproductions of artworks, we have public funding schemes for this work. But this seems not to be enough. In the current economy of streaming, the increasingly low income generated by most European works is insufficient to even cover the cost of proper documentation, which then sends that part of the European repertoire into a self-fulfilling oblivion: the algorithm cannot “learn” its properties and it never shows these works to users and audiences.

Until now, in most cases, it was assumed that it is the artists or their representative’s duty to provide high quality metadata, but in the analogue era, or in the era of individual digital copies, we did not anticipate that the sales value will not even cover the documentation cost. We must find technical solutions with interoperability and new economic incentives to create proper metadata for Europe’s cultural products. With that, we can cover one area out of the three possible problem terrains.

But this is not enough. We need to address the question of how new, better Algorithms can learn from user history and avoid amplifying pre-existing bias against women or hateful speech. We need to make sure that when Algorithms are “scraping” the internet, they do so in an accountable way that does not make small language repertoires vulnerable.

Incentives and investments into metadata

In our paper we argue for new regulatory considerations to create a better, and more accountable playing field for deploying Algorithms in a quasi-autonomous system, and we suggest further research to align economic incentives with the creation of higher quality and less biased metadata. The need for further research on how these large systems affect various fundamental rights, consumer or competition rights, or cultural and media policy goals cannot be overstated. The first step is to open and understand these autonomous systems. It is not enough to say that the firing squads of Big Tech are shooting women out from charts, ethnic minority artists from screens, and small language authors from the virtual bookshelves. We must put a lot more effort on researching the sources of the problems that make machine learning Algorithms behave in a way that is not compatible with our European values or regulations.

This blogpost was first published on our general interest blog Data & Lyrics

Feasibility Study On Promoting Slovak Music In Slovakia & Abroad

Thu, 25 Mar 2021 11:00:00 +0000

How to help promote local music?

The new study opens the question of the local music promotion within the digital environment. The Slovak Performing and Mechanical Rights Society (SOZA), the State51 music group in the United Kingdom, and the Slovak Arts Council commissioned Reprex to created a feasibility study which provides recommendations for better use of quotas for Slovak radio stations and which also maps the share and promotion of Slovak music within large streaming and media platforms such as Spotify.

What should a good local content policy (radio quota, recommendation system, streaming quota) achieve?

The study proposes best practices for the introduction of mandatory quotas for Slovak radio stations and points out how current recommendation systems used by large platforms such as Spotify, YouTube, or Apple hardly consider local music from smaller countries. Local music stands against competition consisting of million songs from the whole world, and for ordinary Slovak musicians, whose music doesn’t belong to the global hits playlists, it is almost impossible to get recommended by the recommendation systems of large platforms.

Listen Local App for discovering new music

We aimed to create a demo version of a utility-based, transparent, accountable recommendation system.

The solution to this problem could be the Listen Local App, built on a comprehensive reference database of local music, which we created as a demo version within the study. The app aims to help listeners discover more local music; the app also presents new and alternative ways for large digital platforms to recommend local artists. Through Listen Local, listeners search for artists and bands based on their taste and the city they are situated in. In this way, listeners can easily search for music by artists from particular cities or from the town they are about to visit. We are releasing today the feasibility study in English and Slovak. We call for an open consultation to evaluate the results of this work and continue developing the Slovak Music Database, the Listen Local recommendation, and the AI validation system.

Check out the Demo Listen Local App. We explain here why.

Screenshot of the first verison of the demo app.

Database

The Slovak Music Database is connected to Reprex’s flagship project, the Demo Music Observatory, an open collaboration-based demo version of the planned European Music Observatory, currently being further developed in the JUMP Music Market Accelerator Programme supported by Music Moves Europe.

The project website contains the demo version of the Slovak Music Database.

Download the Study

You can download the study herein Slovak or in English.

Next steps

In the next phase of the work, we add further data to our Slovak Demo Music Database and carry out more and more experiments and educational activities to understand how Slovak music can become more visible and targeted. We are also bringing this project into an international collaboration for better utilization of R&D efforts and experiences throughout Europe. This agile project method originated in reproducible scientific practice and open-source software development and allows participation in large projects on any scale: from individual musicians and educators to large research universities and music distributors. Anyone can join in on the effort.

Reprex is looking for further international partners; Reprex is currently part of the Dutch AI Coalition and the European AI Alliance project. SOZA and Reprex are committed to opening this project for international collaboration while ensuring that a significant part of the R&D activities remains in the Slovak Republic.

We are preparing informal, online information sessions for artists, promoters, researchers, and developers to join our project.

Contributors

The Reprex team who contributed to the English version:

Budai, Sándor, programming and deployment
Dr. Emily H. Clarke, musicologist
Stef Koenis, musicologist, musician
Dr. Andrés Garcia Molina, data scientist, musicologist, editor
Kátya Nagy, music journalist, research assistant;

and the Slovak version:

Dáša Bulíková, musician, translator
Dominika Semaňáková, musicologist, editor, layout.

Special thanks to Tammy Nižňanska & the Youniverse for the case study.

Demo Slovak Music Database

Thu, 17 Dec 2020 17:10:00 +0200

We are finalizing our first local recommendation system, Listen Local Slovakia, and the accompanying Demo Slovak Music Database. Our aim is

Show how the Slovak repertoire is seen by media and streaming platforms
What are the possibilities to give greater visibility to the Slovak repertoire in radio and streaming platforms
What are the specific problems why certain artists and music is almost invisible.

In the next year, we would like to create a modern, comprehensive national music database that serves music promotion in radio, streaming, live music within Slovakia and abroad.

To train our locally relevant, alternative recommendation system, we filled the Demo Slovak Music Database from two sources. In the opt-in process we asked artists to participate in Listen Local, and we selected those artists who opted in from Slovakia, or whose language is Slovak. In the write-in process we collected publicly available data from other artists that our musicology team considered to be Slovak, mainly on the basis of their language use, residence, and other public biographical information. The following artists form the basis of our experiment. (If you want to be excluded from the write-in list, write to us, or you want to be included, please, fill out this form.)

Click here to view the table on a separate page

Modern recommendation systems usually rely on data provided by artists or their representatives, data on who and how is listening to their music, and what music is listened to by the audience of the artists, and certain musicological features of the music. Usually they collect data from various data sources, but these data sources are mainly English language sources.

The problem with these recommendation systems is that they do not help music discovery, and make starting new acts very difficult. Recommendation systems tend to help already established artists, and artists whose work is well described in the English language.

Our alternative recommendation system is a utility-based system that gives a user-defined priority to artists released in Slovakia, or artists identified as Slovak, or both. The system can be extended for lyrics language priorities, too. Currently, our app is demonstration to provide a more comprehensive database-driven tool that can support various music discovery, recommendation or music export tools. Our Feasibility Study to build such tools and our Demo App is currently under consultation with Slovak stakeholders.

Listen Local is developing transparent algorithms and open source solutions to find new audiences for independent music. We want to correct the injustice and inherent bias of market leading big data algorithms. If you want your music and audience to be analysed in Listen Local, fill this form in. We will include you in our demo application for local music recommendations and our analysis to be revealed in December.

Listen Local: Why We Need Alternative Recommendation Systems

Mon, 14 Dec 2020 17:10:00 +0200

The first version of our demo application

Recommendation systems utilize knowledge about music content and their audiences while also pursuing the objectives or needs of recommenders.

The simplest recommendation systems just follow the charts: for example, they select from well-known current or perennial greatest hits. Such a system may work well for an amateur DJ in a home party or a small local radio that just wants to make sure that the music in its programme will be liked by many people. They reinforce existing trends and make already popular songs and their creators even more popular.

If the recommendation engine is supported by big data and a machine learning system – or increasingly, a combination of several machine learning algorithms – the general modus operandi is to exploit information about both content and users in order to achieve certain goals.

How algorithmic recommendation systems work?

How recommendation systems work?

Spotify’s recommendation system is a mix of content- and collaborative filtering that exploits information about users’ past behaviour (e.g. liked, skipped, and re-listened songs), the behaviour of similar users, as well as data collected from the users’ social media and other online activities, or from blogs. Deezer uses a similar system that is boosted by the acquisition of Last.fm – big data created from user comments are used to understand the mood of the songs, for example.

Spotify makes 16 billion music recommendations each month in 2020.

YouTube, which plays an even larger role in music discovery, uses a system comprised of two neural networks: one for candidate generation and one for ranking. The candidate generation deep neural network provides works on the basis of collaborative filtering, while the ranking system is based on content-based filtering and a form of utility ranking that takes into consideration the user’s languages, for example.

What makes these systems common is that they maximize the algorithm creators’ corporate key performance indicators. Spotify wants to be ‘your playlist to life’ and increase the amount of music played during work or sports in the background, during travelling, or active music listening –- i.e. maximizing the number of hours spent using it, and do not let empty timeslots for other music providers, such as radio stations. YouTube and Netflix have similar targets. They are in many ways like commercial radio targets, which want to maximize the time spent listening to the broadcast stream. Radios and YouTube, in particular, have similar goals because they are mainly financed through advertising. For Spotify or Netflix, their key financial motivation is to avoid users’ cancelling their subscriptions or changing it to different providers, such as Amzon, Apple or Deezer.

What is the problem with black box recommendation systems?

What they also have in common is that they do not aim to give a fair chance to each uploaded song, serve equally every artist, or provide whatever equality of chances for English, Slovak or Farsi language content.

They tend to reinforce trends similarly to music charts, but with far bigger efficiency. As the Dutch comedian, author and journalist Arjen Lubach explains the YouTube algorithm, to keep their personal recommendations engaging all the day and all of the night, they create a comfortable universe for the user allows little distraction in. If the user wants to listen to global hit music, or stoner rock, it will never be distracted with anything else.

Zondag met Lubach on Dutch public broadcaster VPRO. Click settings sign to change the language of the captions.

The problem with such hyper-personalized media is that they leave no room for public activities. Public broadcasters, which had a monopoly to television broadcasting in most European countries until the early 1990s, for example, were aiming to air a diversity of news, knowledge and access to local culture. Many countries on all continents have maintained local content guidelines for broadcasting on public, commercial and community television and radio channels, for example, local music and films, and reliable news as a public service. Personalized media-, social media- and streaming platforms do not have such obligations.

Black box recommendation systems usually maximize a corporate key performance indicator, and they are not subject to usual public new service or local content regulations that traditional broadcast media is.
The goal and the steps that the algorithm is pursuing is not know to content creators, and they do not know when will the algorithm work for their benefit or against them.

Transparent and regulated AI

In our view, utility-based recommendation system can provide a bridge between current, corporate-owned systems that maximize a media or streaming platforms’ business indicators.

Public new service requirements or local content requirements (“national quotas”) set for commercial broadcasting are similar to utility or knowledge-based recommendation systems. A utility-based recommendation system, for example, would prefer from two candidates for a playlist the one that has a Slovak composer, or a performer from Wales, or which has Farsi lyrics.

Our Demo App creates recommendations on the basis of a pre-existing radio or personal streaming playlist that, by choice, contains a pre-defined ratio of music produced in Slovakia, or performed by Slovak artists. We will soon add Dutch and Hungarian choices to this demo, but naturally, we could add any city’s, regions’, province’s our countries preferences into the app.

Our Demo App and accompanying Feasibility Study in Slovakia shows how can a regulator create better broadcasting regulations, learning from the experience with AI-driven streaming platforms, and how can it apply the goals of local content requirements (such as a certain visibility for Slovak or a city-based music) and public service requirements (for example, spreading reliable information or stopping hateful music.)

Access for all

We do not believe that the current heated discussion on the re-regulation of AI and music streaming will solve all the problems of independent artists, bands from ethnic or racial minorities, or otherwise vulnerable producers. New regulation can limit the unintended collateral damage of big data algorithms deployed by big corporations, but they will not bring down the benefits of AI to these creators.

Take the example of the controversial new initiative that let’s artists, labels and publishers to promote their music in exchange for a cut in royalty rates. While felt by many artists injust and even corrupt, it is an answer to the growing need to influence how the recommendation algorithms are promoting certain music at the expense of tens of millions of sound recordings that are not recommended.

We believe that algorithms create value for the users, and if artists are not happy to pay for corporations to influence black box algoritms, thatn they must collaborate and share data, and build large enough data pools so that they can deploy white, transparent algorithms that work for them.

Our Feasibility Study shows why it is important that creators have a control over what data describes their music, their biographies and other information online, because corporate streaming platforms use this information for their algorithms.
We should that with relatively little effort creators can pool enough information to create alternative recommendation systems that follow a more agreeable goal, that is sensitive to local content requirements and more access to new artists, women or black performers, or which suppress hateful lyrics.
The benefit of an open algorithm and pooled data is that artists can actively look for audiences in various age groups or in cities that are accessible for them on a performing tour after the pandemic.

Overall, we want to show that regulating black box, private algorithms and data monopolies is only a first step to damage control. Deploying white, transparent algorithms and building collaborative or open data pools can only guarantee fairness in the digital platforms, in recommendations, and generally in the use of AI.