Algorithmic Fairness (RecSys)
A recommender can be accurate on average and still serve some people badly while starving part of the catalogue of any attention at all. Algorithmic fairness is the study of those gaps: who the system works for, which items get seen, and what to do when an averaged metric hides the answer.
Two sides of every list
Fairness in recommendation has a consumer side and a provider side, and they are not the same problem. Consumer-side fairness asks whether recommendation quality is comparable across groups of users; accuracy drops for users with short histories or tastes outside the mainstream. Provider-side fairness asks how exposure is distributed across the people whose items are recommended, which decides whether a small seller or a new artist is discoverable at all. A third framing, calibration, asks that a list reflect the proportions of a user's actual interests instead of collapsing onto their single strongest one.
Why the gaps compound
Position bias means the top slot absorbs most of the attention, so a small score difference becomes a large exposure difference. Popularity bias then feeds that result back as training data, and the next model is more confident about the same narrow set. Reported as one averaged NDCG, none of this is visible.
Where interventions live
Fairness is mostly enforced after scoring, which leaves the accuracy model untouched:
- Re-ranking under exposure constraints, often amortised across many sessions rather than a single list.
- Fairness regularisers added to the training objective to penalise group disparity directly.
- Per-group reporting of the standard metrics, splitting users by history length and items by popularity tier.
Each trades some measured accuracy for a distribution that can be defended, and the size of that trade is a product decision rather than a modelling one.
Algorithmic fairness in recommendation is the requirement that a ranked list not systematically disadvantage groups of users or groups of items. It has a consumer side, whether recommendation quality holds across user groups, and a provider side, how exposure is spread across item owners. It matters because averaged accuracy metrics hide both: position and popularity bias concentrate attention on a few items, and the usual remedy is an exposure-constrained re-ranking stage.
Think of a shop deciding what goes on the shelf at eye level. Whatever lands there sells, and whatever ends up near the floor barely moves. The shelf is not neutral, and neither is a list of recommendations.
Algorithmic fairness asks two questions about that list. Does it work as well for someone with unusual taste, or almost no history, as it does for everyone else? And do the things being recommended get a reasonable chance of being seen, or do a few big names take every good slot? A system can look excellent on average and still fail quietly on both counts, so these are measured and corrected deliberately rather than left to luck.