Post-Training Fairness Control: A Single-Train Framework for Dynamic Fairness in Recommendation

Despite growing efforts to mitigate unfairness in recommender systems, existing fairness-aware methods typically fix the fairness requirement at training time and provide limited post-training flexibility. However, in real-world scenarios, diverse stakeholders may demand differing fairness requirements over time, so retraining for different fairness requirements becomes prohibitive. To address this limitation, we propose Cofair, a single-train framework that enables post-training fairness control in recommendation. Specifically, Cofair introduces a shared representation layer with fairness-conditioned adapter modules to produce user embeddings specialized for varied fairness levels, along with a user-level regularization term that guarantees user-wise monotonic fairness improvements across these levels. We theoretically establish that the adversarial objective of Cofair upper bounds demographic parity and the regularization term enforces progressive fairness at user level. Comprehensive experiments on multiple datasets and backbone models demonstrate that our framework provides dynamic fairness at different levels, delivering comparable or better fairness-accuracy curves than state-of-the-art baselines, without the need to retrain for each new fairness requirement. Our code is publicly available at https://github.com/weixinchen98/Cofair.

💡 Research Summary

The paper introduces Cofair, a novel framework that enables post‑training control of fairness in recommender systems without the need for multiple retraining cycles. Traditional fairness‑aware recommendation methods embed fairness constraints directly into the training objective, fixing the fairness level once training is completed. Consequently, any change in stakeholder requirements—such as stricter regulatory demands or different business policies—necessitates a full retraining of the model, which is computationally expensive and impractical for large‑scale production systems.

Cofair tackles this limitation by learning a single set of parameters that can generate user embeddings tailored to a range of fairness levels. The architecture consists of three main components:

1. Shared Representation Layer – A neural network (S) maps the original user embedding (e_u) (produced by any standard recommender backbone such as BPR or LightGCN) to a lower‑dimensional shared embedding (s_u). This layer captures user characteristics that are common across all fairness settings, reducing redundancy and serving as a stable foundation for downstream adapters.

2. Fairness‑Conditioned Adapters – For each predefined fairness level (t \in {1,\dots,T}), a lightweight adapter network (P^{(t)}) receives both the original embedding (e_u) and the shared embedding (s_u) and outputs a correction vector (p^{(t)}_u). The final user representation for level (t) is obtained by concatenating (s_u) and (p^{(t)}_u) and passing them through an output transformation (O). Because adapters are independent, switching to a different fairness level at inference time simply means selecting the corresponding adapter; no additional training is required.

3. User‑Level Regularization – To guarantee that fairness does not deteriorate for any individual user when moving to a stricter level, the authors define a per‑user fairness loss based on the binary cross‑entropy of an adversarial discriminator (D) that tries to predict the sensitive attribute (a_u) from the embedding. A regularization term penalizes any increase in this loss from level (t) to (t+1) using a softplus function, enforcing monotonic improvement of fairness on a per‑user basis.

The fairness objective itself is adversarial: the discriminator (D) is trained to maximize its ability to infer the sensitive attribute, while the recommendation model (shared layer + adapters) is trained to minimize the discriminator’s success. The authors prove that the adversarial loss upper‑bounds the demographic parity (DP) disparity (\Delta_{DP}), establishing a theoretical link between minimizing the adversarial loss and reducing group‑level unfairness.

Because each fairness level has its own loss term, the framework introduces adaptive weighting coefficients (\lambda_t). Rather than fixing (\lambda_t) manually, the authors update them dynamically based on the relative improvement of the fairness loss between consecutive levels. If the fairness loss at level (t+1) does not improve sufficiently over level (t), (\lambda_{t+1}) is decreased, making the fairness term more influential in subsequent updates. This adaptive scheme eliminates costly hyper‑parameter sweeps and yields smooth fairness‑accuracy trade‑offs across many levels.

The overall training objective combines (i) the standard recommendation loss (BPR), (ii) the weighted adversarial fairness loss for each level, and (iii) the user‑level regularization, averaged over all levels. Training proceeds by alternating updates of the discriminator and the shared/adapter networks, while periodically adjusting (\lambda_t). Although each epoch processes all (T) adapters, the shared representation is computed only once, and adapters are intentionally lightweight (single‑layer MLPs), keeping the additional computational overhead modest.

Empirical evaluation spans three public datasets (MovieLens‑1M, Amazon Books, Yelp) and two backbone recommenders (BPR, LightGCN). Fairness is measured by demographic parity difference (\Delta_{DP}) and accuracy by NDCG. Cofair is trained once to produce five fairness levels ranging from no fairness control to maximal control. Results show that Cofair matches or surpasses state‑of‑the‑art baselines (e.g., AFRL, personalized fairness methods) in terms of the fairness‑accuracy curve, especially at higher fairness levels where many baselines suffer steep drops in recommendation quality. Moreover, the memory footprint grows by less than 10 % when increasing the number of adapters, and inference latency is unchanged because only a single adapter is activated per request.

In summary, Cofair contributes five key innovations: (1) a shared‑adapter architecture that efficiently reuses learned user features across fairness settings; (2) an adversarial loss with a provable upper bound on demographic parity; (3) a user‑level regularizer that guarantees monotonic fairness improvements per user; (4) an adaptive weighting mechanism that automatically balances fairness and accuracy across multiple levels; and (5) true post‑training controllability, allowing practitioners to select any desired fairness level at deployment time without retraining. This makes Cofair a practical solution for real‑world recommender systems where fairness requirements evolve over time.