A Relativistic Popularity Adversary for Recommenders

Recommender systems trained on click logs learn that popular items are "good" partly because they were shown more often. One common fix trains a second network, an adversary, to guess an item's popularity from its learned vector, and penalizes the recommender whenever that guess succeeds. The current version of this idea asks an absolute question: is this item popular? That question has two problems. It saturates: once head and tail items sit in separate regions of the embedding space, the adversary becomes confident and its corrective signal fades. It also removes too much: popularity partly reflects real quality, so erasing all of it costs accuracy. We propose asking a relative question instead. Among two items that the same user chose, can the adversary tell which one is the popular one? If it can, the item vectors carry popularity beyond what that user's preference explains, and that excess is what the recommender should lose.

The comparison borrows from relativistic adversarial training, a way of training generative models in which the discriminator scores whether one sample is more realistic than another rather than judging each sample alone. The user-matched pairing borrows its target from the "interactions proportional to like" view of debiasing, which holds that two items a user likes equally should not be separable by exposure. The resulting test is easy to state and to falsify. If the relativistic, user-matched adversary keeps tail-item accuracy gains while giving up less head-item accuracy than an absolute adversary, and if a two-by-two ablation shows which of the two ingredients produces that effect, we learn something about how popularity should be removed from collaborative embeddings, not only that another regularizer helps.

Core Claim

The closest existing system is APWCF (Adversarial Propensity Weighting for Debiasing in Collaborative Filtering). It pairs a dynamically estimated, clipped propensity weight with a discriminator that classifies an item embedding's popularity tier. A gradient reversal layer flips that discriminator's gradient so the encoder learns to hide popularity. Its authors report weaker results on small datasets with shallow models, and the discriminator judges each item alone and unconditionally. Relativistic GANs (The relativistic discriminator: a key element missing from standard GAN) supply the alternative: score a difference of critic outputs so that real and fake terms depend on each other. A follow-up (On Relativistic f-Divergences) shows these objectives are valid divergences and finds that the noisy mini-batch estimator beats the minimum-variance unbiased one. That finding warns us against over-engineering the estimator.

The literature motivates the problem from two sides. PAAC (Popularity-Aware Alignment and Contrast for Mitigating Popularity Bias) documents "representation separation": head and tail items form distinct clusters, which is exactly the condition under which an absolute adversary saturates. IPL (Popularity Debiasing from Exposure to Interaction in Collaborative Filtering) argues that equalizing exposure produces useless exposure. It instead requires interactions to be proportional to latent liking, but it estimates that liking with inverse propensity scoring, which is fragile.

We do not claim that a relativistic loss alone fixes saturation. A confident pairwise sigmoid also saturates. The claim is narrower and testable. Conditioning the comparison on a shared user turns the adversary's target from "popularity" into "popularity not explained by this user's choice." The relativistic form is the natural way to write that conditional comparison, because it scores the same user's two items against each other.

Concretely, let $e_i$ be the encoder's embedding of item $i$. Let $C$ be a small critic network that maps an embedding to a real-valued score. Let $\mathcal{T}$ be the set of triples $(u,i,j)$ in which user $u$ interacted with both $i$ and $j$, and $i$ sits in a higher popularity tier than $j$. The critic minimizes, and the encoder maximizes,

$$\mathcal{L}_{\text{rel}} = -\,\mathbb{E}_{(u,i,j)\sim\mathcal{T}} \log \sigma\big(C(e_i) - C(e_j)\big),$$

where $\sigma$ is the logistic function. The recommendation loss is the usual Bayesian personalized ranking or InfoNCE objective, plus $\lambda$ times the encoder's side of this game. The encoder uses the non-saturating swapped-label form from relativistic GANs rather than a gradient reversal layer.

An averaged variant compares $C(e_i)$ with the mean critic score of the same user's tail items in the batch, mirroring the relativistic average discriminator. That gives us a recommender-side test of an open question the GAN literature leaves unresolved: why averaged comparisons train more stably than paired ones. On datasets with explicit ratings, we tighten $\mathcal{T}$ to pairs with equal ratings, so that "the user chose both" moves closer to "the user liked both equally."

Research Program

We will build the adversary as a plug-in on matrix factorization and LightGCN, the two backbones every cited debiasing paper uses. LightGCN matters especially here. Graph convolution spreads the representations of high-degree popular items across neighborhoods, as shown in How Graph Convolutions Amplify Popularity Bias for Recommendation?. MACR's much larger gains on LightGCN than on MF are consistent with that.

The core analysis is a two-by-two design. One factor is the discriminator form: absolute, as in APWCF's tier classifier, versus relativistic. The other is the comparison set: unconditional, meaning head versus tail across the whole batch, versus user-matched. Every cell uses the same critic architecture, the same $\lambda$ search and the same training budget.

We will also track three diagnostics throughout training. The first is the gradient norm the adversary sends to the encoder, which directly tests the saturation story. The second is the accuracy of a fresh linear probe that predicts popularity tier from frozen embeddings, both unconditionally and within users. The third is the cosine distance between head and tail cluster centroids. Unconditional scrubbing should drive the unconditional probe to chance. Matched scrubbing should drive only the within-user probe to chance, leaving some unconditional popularity signal that tracks quality. If both variants end up with identical probes, the conditioning idea has no bite and we will say so.

Baselines

The naive reference point is the backbone, MF or LightGCN, trained with the plain ranking loss. Every other method is reported as a change relative to it.

The strongest existing method on head-tail fairness is PAAC. It aligns tail items with popular items that the same users interacted with and reweights the contrastive loss by popularity. On Yelp2018 it reports improving NDCG@20 by 30.2% and Recall@20 by 8.7% over the strongest prior baseline, and cutting the popular-versus-unpopular accuracy gap by 87.5%. On Gowalla the gap reduction is 34.18% (Popularity-Aware Alignment and Contrast for Mitigating Popularity Bias). The file gives only relative gains, so the absolute values come from our reruns; no reported absolute Recall@20 for PAAC on Yelp2018 was found on Lacuna. PAAC is the most dangerous competitor because it uses the same user co-interaction signal we use, applied as an attractive alignment rather than an adversarial constraint.

The strongest adversarial competitor is APWCF. It reports up to 5.42% Recall@10 and 8.60% NDCG@10 over PDA, and wins 34 of 40 metrics against ten baselines on Yahoo!R3, KuaiRec, Douban-book, Yelp2018 and ML-100K. No per-dataset absolute numbers were found on Lacuna.

We will also include MACR, which trains popularity and conformity branches and subtracts the direct effect of popularity at test time (Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System). It reports average HR@20 gains of 153% on MF and 241% on LightGCN. Adap-τ completes the set; it normalizes embeddings and sets a closed-form and per-user temperature (Adap-τ : Adaptively Modulating Embedding Magnitude for Recommendation), with a reported 9% average gain.

For fairness, every method runs in one codebase with the same splits, embedding size and epoch budget, and the same number of hyperparameter trials, chosen on validation. The key ablations are the three non-proposed cells of the two-by-two design, plus APWCF with its propensity module removed, which isolates its adversary.

Beating the field means one of two outcomes on the unbiased test protocols. The first is Recall@20 and NDCG@20 at least matching PAAC with a smaller head-item accuracy drop relative to the backbone. The second is matching PAAC on head-tail gap reduction while beating APWCF's adversary on ML-100K, the small-data setting where APWCF's authors report weakness. Either outcome must hold across three seeds with non-overlapping confidence intervals.

Why not just…?

Why not just normalize the embeddings with Adap-τ? If popularity lives only in vector length, normalization removes it and the adversary has nothing left to find. The answer is empirical. We will run the within-user popularity probe on Adap-τ embeddings. PAAC's finding of separated head and tail clusters suggests that popularity also lives in direction, which normalization cannot touch. If the probe is already near chance after Adap-τ, our method is redundant on that dataset and we will report it that way. Running the adversary on top of an Adap-τ backbone tells us whether the two effects add.

Why not just use PAAC's alignment, which already exploits co-interacted head items? This is the real threat. Alignment pulls tail embeddings toward head embeddings, which can blur distinct tail items together. Our adversary removes only the information that tells a user's popular choice from their niche choice and imposes no attraction. The test is head-item accuracy and tail-item diversity, measured as the spread of tail embeddings, at matched gap reduction. If PAAC matches us on both, the adversarial game is not worth its cost.

Why not just fix APWCF's saturation with a hinge or least-squares discriminator, or subtract popularity at test time as MACR does? A non-saturating absolute loss addresses the gradient issue but not the conditioning issue. The absolute-unconditional cell with a hinge loss, included in our search, measures how much of any gain comes merely from avoiding saturation. MACR changes scores but not representations, so it cannot help the downstream reuse of embeddings. On ranking metrics, though, it may simply win, and we will report it alongside everything else.

Proposed Experiments

We will follow PAAC's protocol on Yelp2018, Gowalla and Amazon-book, using unbiased test sets with uniform item distribution and reporting Recall@20, NDCG@20 and the popular-versus-unpopular accuracy gap. We will follow APWCF's protocol on Yahoo!R3 and KuaiRec, whose missing-at-random test data give a less confounded measure, and on ML-100K for the small-data regime. On Yahoo!R3 we will also run the rating-matched version of the pair set.

The decisive check is the two-by-two ablation on LightGCN over Yelp2018 and ML-100K. Our hypothesis predicts a specific pattern. The relativistic form should matter most for gradient health on the small dataset, while user matching should matter most for preserving head-item accuracy. If instead the absolute-unconditional cell with a non-saturating loss matches the full method, the proposal is refuted and the contribution reduces to a loss-function fix for APWCF.

A second check plots adversary gradient norm against epoch for all four cells. The saturation story predicts that the absolute-unconditional curve collapses early while the others do not. A third check reports within-user and unconditional probe accuracy, which tells us whether the method removed the confounded part of popularity rather than all of it.