A Counterfactual Approach for Addressing Individual User Unfairness in Collaborative Recommender System

Recommender Systems (RSs) are exploited by various business enterprises to suggest their products (items) to consumers (users). Collaborative filtering (CF) is a widely used variant of RSs which learns hidden patterns from user-item interactions for recommending items to users. Recommendations provided by the traditional CF models are often biased. Generally, such models learn and update embeddings for all the users, thereby overlooking the biases toward each under-served users individually. This leads to certain users receiving poorer recommendations than the rest. Such unfair treatment toward users incur loss to the business houses. There is limited research which addressed individual user unfairness problem (IUUP). Existing literature employed explicit exploration-based multi-armed bandits, individual user unfairness metric, and explanation score to address this issue. Although, these works elucidate and identify the underlying individual user unfairness, however, they do not provide solutions for it. In this paper, we propose a dual-step approach which identifies and mitigates IUUP in recommendations. In the proposed work, we counterfactually introduce new interactions to the candidate users (one at a time) and subsequently analyze the benefit from this perturbation. This improves the user engagement with other users and items. Thus, the model can learn effective embeddings across the users. To showcase the effectiveness of the proposed counterfactual methodology, we conducted experiments on MovieLens-100K, Amazon Beauty and MovieLens-1M datasets. The experimental results validate the superiority of the proposed approach over the existing techniques.

💡 Research Summary

The paper tackles the problem of individual user unfairness (IUU) in collaborative‑filtering (CF) recommender systems, where certain users—typically those with few interactions—receive systematically poorer recommendations. Existing works mainly identify the issue or propose group‑level fairness constraints, but they do not provide a concrete mechanism to improve the recommendations for the disadvantaged users.
The authors introduce a dual‑stage, counterfactual imputation framework built on top of LightGCN, a graph‑based CF model that learns user and item embeddings through message passing and optimizes a Bayesian Personalized Ranking (BPR) loss.
Stage 1: After training a baseline LightGCN model (denoted B), the authors compute the Normalized Discounted Cumulative Gain (NDCG) for each user on the training data. This NDCG, called G_prior_u, serves as a disparity metric: users with low G_prior_u or with very few observed items |I(u)| are flagged as probable under‑served candidates (U′).
Stage 2: For each candidate u′, a separate prediction model M (e.g., matrix factorization, neural MF) generates top‑K unseen items with the highest predicted scores. The number of items to be counterfactually added, |C(u′)|, is governed by two hyper‑parameters α (overall magnitude) and β (sparsity sensitivity) as shown in equation (5). These items are temporarily inserted as new edges (u′, i) into the user‑item bipartite graph, yielding a modified interaction matrix R′. LightGCN is then re‑run on R′, updating embeddings according to equations (6) and (7). The model is retrained with the same BPR loss, producing a new set of predictions and a post‑imputation NDCG G_post_u′.
The performance gain for each user is defined as PGain_u′ = G_post_u′ – G_prior_u′. If PGain_u′ > 0, the counterfactual interactions are retained in the final training data; otherwise they are discarded. The authors also aggregate PGain across all users to obtain a global fairness improvement measure.
Experiments are conducted on three public datasets: MovieLens‑100K, MovieLens‑1M, and Amazon Beauty. Baselines include the original LightGCN, standard MF/NCF, ACFR‑User, and multi‑armed bandit approaches. Evaluation metrics are NDCG@10, Hit‑Rate@10, and the proposed disparity/PGain measures. Results show that the proposed method consistently outperforms baselines, achieving 3–5 % absolute gains in NDCG and HR, and delivering average PGain of 0.15 – 0.18 for the most sparse users (those with ≤5 interactions). Moreover, the overall NDCG gap between advantaged and disadvantaged users shrinks by roughly 40 % compared to existing techniques.
Key strengths of the work are: (1) a clear, quantifiable definition of individual unfairness based on NDCG; (2) a lightweight counterfactual augmentation that can be applied to any graph‑based CF model; (3) an automatic, per‑user decision rule (PGain) that prevents over‑correction.
Limitations include reliance on the quality of the auxiliary prediction model M (poor predictions could introduce noise), the offline nature of the evaluation (real‑world user reactions to synthetic interactions remain untested), and the focus on NDCG as the sole fairness proxy, which may not capture other dimensions such as diversity or group fairness. Scaling the candidate selection and re‑training pipeline to massive, streaming environments also requires further engineering.
In conclusion, the paper presents a novel, practical solution for mitigating individual user unfairness in recommender systems by leveraging counterfactual interaction imputation and performance‑gain‑driven selection. It opens avenues for future research on dynamic, online counterfactual updates and the integration of multiple fairness criteria.