Steering Recommendation Trade-offs Through the Choice of Negatives

Recommenders trained on clicks or purchases never see explicit dislikes. They learn by contrasting each item a user engaged with against "negatives", which are unobserved items treated as dislikes. Which negatives get picked matters a great deal. Hard negatives are items the model currently scores highly but the user did not choose, and they sharpen top-of-list ranking. They also decide which items get pushed down during training. Popular item groups get sampled as hard negatives more often, so their exposure is distorted. Long-tail items the model should surface can also get treated as things to suppress. So the negative sampler is quietly setting the balance between accuracy, fairness and diversity, but almost nobody treats it as the control for that balance.

We propose to do exactly that. We build a single recommender that takes a preference vector as input. The vector says how much the user of the system cares about accuracy versus item-group fairness versus catalog diversity, and it decides how each training batch's negative budget is split among objective-specific samplers. The split is set by a non-linear rule rather than a fixed weighted sum. The rule gives more negatives to whichever objective currently lags most relative to its requested weight, so the training can, in principle, reach trade-offs that weighted sums provably miss. At serving time, sweeping the preference vector should trace a whole trade-off curve from one model. The core question is empirical: does controlling the trade-off through negatives give a better curve than controlling it through loss weights, at the same compute?

Core Claim

Our claim rests on two findings in the file. First, On the Theories Behind Hard Negative Sampling for Recommendation shows that pairwise ranking loss with hard negatives is equivalent to optimizing one-way partial AUC (ranking quality among the top-scored negatives) under distributionally robust optimization. Sampling hardness therefore chooses which region of the ranking the gradient attends to, and it should be matched to the evaluation cutoff. Second, Fairly Adaptive Negative Sampling for Recommendations (FairNeg) shows that the same hardness concentrates negative gradients on majority item groups and damages exposure fairness.

Together these say that the negative distribution is a direct, cheap lever on several objectives at once. It also decides where the push-down falls, which a scalar loss weight does not. We claim that a preference-conditioned model whose negatives are allocated by a hypervolume-style rule will dominate preference-conditioned loss reweighting in hypervolume on accuracy, fairness and diversity. Hypervolume here means the volume of objective space covered by the achieved trade-off points. We also claim the gain will be largest in the regions of the trade-off curve that a weighted mixture cannot reach.

How It Differs From Existing Work

Four papers sit closest to this idea, and each leaves a specific piece undone.

FairNeg blends a hardness-aware and a fairness-aware negative distribution, adjusting group sampling probabilities through bi-level optimization. Its blend is a linear mixup for one fixed operating point over coarse binary groups. We condition one model on a continuous preference over three objectives and replace the linear blend with a non-linear allocation.

A Data-Centric Multi-Objective Learning Framework for Responsible Recommendation Systems (MoRec) also steers trade-offs through the data distribution, using a proportional-integral controller and an adaptive sampler. It reweights training instances rather than negatives. It also targets one operating point, keeping accuracy within a 3% loss band.

TaoSearchEmb replaces offline hard-negative pipelines with on-policy retrieval rewarded by a 42-billion-parameter teacher. It combines relevance, quality and exclusivity by weighted scalarization and drops contrastive losses entirely. We keep the contrastive loss, with its partial-AUC interpretation, and need no teacher.

MultiBalance balances per-task gradients at the shared representation layer to avoid the 70–80% throughput collapse of full-model gradient balancing. It explicitly assumes a fixed negative sampler. Our allocation needs one backward pass per step, since only the batch composition changes, so it competes with MultiBalance on cost.

The multi-objective side supplies the method. Optimal Scalarizations for Sublinear Hypervolume Regret proves that linear scalarization cannot reach concave parts of a trade-off curve. It also shows that the hypervolume scalarization $s_\lambda(y)=\min_i (y_i/\lambda_i)^k$, with $\lambda$ drawn uniformly from the positive unit sphere, can reach any point. Here $y_i$ is the score on objective $i$ and $k$ is the number of objectives. PD-MORL shows that one preference-conditioned network can replace a set of per-preference models with an order of magnitude fewer parameters. We borrow both ideas and move them from reinforcement learning into the choice of negatives.

Research Program

We use three objective-specific negative samplers. The accuracy sampler uses dynamic hard-negative sampling with controllable hardness, the DNS(M,N) form from Shi et al. The fairness sampler follows FairNeg's group-balanced distribution, steered by its group-wise cross-entropy signal. The diversity sampler is popularity-tempered: it lowers the chance that long-tail items are drawn as negatives.

For each objective $i$ we track an exponential moving average $\hat{g}_i$ of a normalized shortfall, meaning how far that objective's training proxy sits from its best value seen so far. The proxies are ranking loss, group cross-entropy, and a coverage loss. The magnitude smoothing follows MultiBalance's response to mini-batch noise.

Given a preference $\lambda$, the share of the batch's $N$ negatives drawn from sampler $i$ is

$$\pi_i(\lambda)=\frac{\exp(\beta\,\hat{g}_i/\lambda_i)}{\sum_j \exp(\beta\,\hat{g}_j/\lambda_j)},$$

where $\beta$ is a temperature. As $\beta$ grows, the whole budget goes to the objective that maximizes $\hat g_i/\lambda_i$, which is the objective attaining the minimum inside the hypervolume scalarization. A fixed $\pi_i=\lambda_i$ recovers FairNeg-style linear mixup, which is our key ablation.

During training, $\lambda$ is drawn fresh for each batch from the sphere and fed to the user tower through feature-wise scaling. The base models are LightGCN-style collaborative filtering and matrix factorization trained with a pairwise ranking loss, so that Shi et al.'s theory applies. An InfoNCE variant is run as a secondary check.

We also analyze one failure mode directly: false negatives, meaning sampled negatives that are actually future positives. We measure the fraction of sampled negatives that appear in each user's held-out test interactions, as a function of $\lambda$. This tests whether shifting budget toward the diversity sampler reduces false negatives, which is an open problem that Shi et al. flag.

Baselines

The naive reference is uniform negative sampling with ranking loss, which Shi et al. show optimizes full AUC rather than top-heavy partial AUC. Every frontier is reported relative to it.

The strongest accuracy-only competitor is DNS with hardness tuned to the cutoff, as in Shi et al. They report NDCG@50 of 0.1811 on Gowalla. Our model at the accuracy-heavy end of $\lambda$ must come within noise of a DNS model we retrain on our own split before any trade-off gain counts.

The strongest competitor at the negative-sampling level for accuracy versus fairness is FairNeg. It reports a Recall-Disparity reduction of over 90% versus uniform sampling on ML1M-2 and a 65% gain in Recall-Min on Amazon-2. The file reports no absolute Recall or NDCG values for FairNeg, so we rerun it on identical splits. We sweep its mixup factor over the same number of models as our training budget allows, producing a multi-model FairNeg frontier.

For the data-centric competitor we run MoRec on Amazon Electronics and Amazon Movies, where the reported result is accuracy loss held within 3% alongside fairness and revenue gains. Beating it means a point on our curve that matches its accuracy and exceeds it on fairness and diversity.

The gradient-balancing competitor is MultiBalance-style representation-layer balancing of three per-objective losses over a fixed DNS sampler, made preference-conditioned. Its result on its own production systems was a 0.42% throughput drop, and no reported result was found on Lacuna for it on public benchmarks.

The key-ingredient ablations are the same conditioned model with linear allocation ($\pi_i=\lambda_i$), and the same conditioned model with fixed DNS negatives but loss-side weights on fairness and diversity regularizers.

All methods share the data splits, embedding size, epochs, and a fixed hyperparameter search budget of equal trial count. Results cover five seeds. Beating a baseline means a statistically larger hypervolume, with the reference point set by uniform sampling, at no worse wall-clock time.

Why not just…?

Why not just add fairness and diversity regularizers to the loss and weight them? This is the main threat. It is what our loss-side ablation tests directly. Loss weights scale gradients but leave the same majority-group items being pushed down, which is the mechanism FairNeg identifies. If the loss-side conditioned model matches our hypervolume, the sampler-as-control claim fails, and we will report that.

Why not just sweep FairNeg's mixup coefficient and keep the best few models? If the true trade-off curve is convex, a linear mixup sweep reaches all of it. In that case the hypervolume rule adds nothing beyond one shared model. The linear-allocation ablation separates the two contributions. The non-linear rule only earns its place if we find concave stretches where linear sweeps leave gaps, as Zhang's result predicts they can.

Why not just balance gradients with MGDA? Full-model MGDA costs 70–80% of training throughput according to MultiBalance. Representation-layer balancing fixes the cost but takes the negative distribution as given, so it cannot move which items get suppressed. We run it as a baseline rather than dismissing it.

Why not just use uniform negatives and avoid false negatives? Shi et al. and SCONE show that uniform negatives under-train top-K ranking. FairNeg shows they are not fair either. Uniform sampling serves as the reference point that every frontier must clear.

Proposed Experiments

The fairness experiments run on ML1M-2 and Amazon-2 with FairNeg's group splits, reporting Recall@20, NDCG@20, Recall-Disparity and Recall-Min. The three-objective experiments run on Amazon Electronics and Amazon Movies with MoRec's category labels and five-core filtering, adding catalog coverage as the diversity metric. Gowalla and Yelp serve as accuracy-only checks against DNS.

The first decisive check is whether our curve has larger hypervolume than the loss-side conditioned ablation. The second is whether the curve swept by linear-allocation runs shows concave gaps that the hypervolume rule fills. The third is whether the accuracy-heavy end matches tuned DNS. A negative answer to the first check would mean negatives are not a better lever than loss weights, and that would itself be a useful result.