What a Recommendation Engine Is Optimising For, and How That Becomes Policy

Recommendation engines are not just neutral filters of content. The objective they optimize for shapes what their audience is exposed to—and, over time, amplifies the reach of the content that is already popular.

Most platform users are aware of recommendation engines' power to influence what content they see. But fewer are aware of how those engines' core objectives shape the content themselves. In recommender systems, the optimization goal is as important as the model. And that goal can become a policy choice with consequences over time, not an inevitability for the platform.

A 2020 paper called Feedback Loop and Bias Amplification in Recommender Systems describes how recommendation algorithms can suffer from popularity bias, where a few popular items are recommended frequently, while many others are ignored. The paper's authors develop a method to augment the recommendation model's objective function with the aim of minimizing disparity in loss values across different groups of items. This shows that optimization isn't neutral: it has to be consciously chosen.

Recommendation engines don't just train on data, exposed to content, and become a neutral conduit: they shape what's seen,, favorite and visit pattern, and data. All of which gets fed back into the model itself, in a "the rich get richer" loop. An August 2023 survey on popularity bias in recommenders shows that feedback loops amplify popularity bias, making already-popular items more attractive for recommendations. A 2019 paper suggests simulating user interactions with recommenders offline to study the impact of popularity-bias amplification.

Popularity Bias and the Rich-Get-Richer Loop

If recommendation engines have a natural tendency to concentrate visibility and engagement on popular content, it skews the system. The recent popularity bias survey on arXiv defines the problem as recommendations focusing on popular items to the extent that they limit the system's value or create harm for some stakeholders. The survey shows that this popularity bias is a natural tendency of the system, not just an artifact of biased or unfair training data.

That's why platforms add guardrails, constraints and mechanisms, to consciously shape the system. And still, feedback loops propagate some of that bias. A 2020 paper says recommended items are consumed, and user reactions are logged, feeding into a feedback loop.

Recommendation systems can create a feedback loop because the system’s exposure mechanism affects user behavior, and those behaviors feed back into the training data, as another Arxiv paper puts it. It’s a closed loop: exposure drives engagement, which feeds back in as data, which affects what’s shown next. This feedback loop can intensify bias over time and produce a "the rich-get-richer" Matthew effect, the 2020 paper concludes. It shows that popularity bias in recommender systems means popular items are recommended more often than their popularity alone would warrant.

What the System Trains On Becomes What It Sees

Optimization goals don't just lead to individual recommendations: they shape the system itself. The exposure mechanism can create or amplify bias, by its own nature, not just because of the starting data. That's why a platform's choice of optimization goal matters. It's a policy, not just a technical detail.

In that 2020 paper, "Feedback Loop and Bias Amplification in Recommender Systems," the authors state that recommended items are consumed, user reactions are logged, and the resulting data form a feedback loop. The same 2020 paper states that its method augments the recommendation model’s objective function with an extra term aimed at minimizing disparity in loss values across item groups.

But what the system starts to recommend - what it trains on and trains to recommend - becomes what it sees over time. Exposing more of some items concentrates traffic on them. That traffic is logged, becoming data, which influences what's shown next. A 2019 paper titled "Degenerate Feedback Loops in Recommender Systems" proposes simulating user interaction with recommenders offline and studying the impact of feedback loops on popularity-bias amplification.

Guardrails Added After the Fact

So, platforms add guardrails and constraints to recommendation systems, to add more diversity or downrank some content.

These guardrails are additions, not the core of the system. The survey on popularity-biased recommenders calls this problem "post-exposure bias amplification". In the 2020 paper, the authors argue that feedback loops can propagate the popularity bias of the content. They propose a method to debias the recommendation's objective, to mitigate post-exposure bias amplification over time, bringing the model to recommend more diverse, rich content.

Why the Objective Matters More Than the Model

The objective that a recommendation system optimizes for determines the shape of the distribution of engagement its audience sees, which is what that audience consumes, which is fed back into the model, impacting the future distribution. This distribution of visible, consumed, repeatable content is an automatic process, not a choice.

That's why prominently putting the data and the model out for public inspection, without highlighting the objective, is misleading at a minimum and fraudulent at its extreme, pushing attention to popular, proven content that's profitable while claiming transparency.

There's nothing lethal about a recommendation algorithm per se. It's not about removing "recommender" algorithms from the platform, or forcing users to browse manually. It's about making the recommender itself a worthy of its name. It's about not making engagement the automatic, conscious objective of a recommendation engine. It's about letting a recommendation algorithm bring its beneficial output — novelty, progression, curiosity — to the fans. And it's about letting those fans decide if they consumed the result.