Work

Adaptive Lagrangian Control of Long-Tail Exposure in Graph Recommendation

Year: 2026
Type: preprint
Abstract: Graph neural network (GNN) recommenders can amplify popularity bias, limiting long-tail exposure. Existing in-training remedies rely on a penalty weight that must be tuned to achieve a desired exposure level, and the same weight can produce very different exposure levels across datasets. We propose Adaptive Lagrangian Fairness (ALF), a primal–dual controller that makes the desired topK share of reachable long-tail items the control variable of GNN training: a dual variable driven by the measured validation tail share sets the strength of a separate, interchangeable differentiable surrogate. We instantiate it with a softmax tail-mass proxy (ALF-S) and the pairwise popularity-bias loss PBiLoss (ALF-P). On MovieLens-100K, MovieLens-1M, and Yelp2018 with a LightGCN backbone, fixed weights transfer poorly: the same softmax-penalty weight yields tail exposure differing up to elevenfold across datasets, and the same PBiLoss weight up to 2.6-fold. ALF instead attains the requested share closely, with median errors of 0.002–0.004 for ALF-S on MovieLens-1M and Yelp2018 and for ALF-P on MovieLens-100K within its range. Against their static counterparts over shared exposure ranges, the ALF variants have higher common-range hypervolume in five of six comparisons (up to 13%), with 95% bootstrap intervals excluding zero in four. The surrogates differ in range: ALF-P saturates near tail shares of 0.55, 0.59, and 0.21, whereas ALF-S remains responsive across the full tested range. Cosine scoring also raises unconstrained long-tail exposure over inner-product scoring on all three datasets.
Language: English
Open Access status: green
OpenAlex ID: W7225999176
Cites: 32
Cited by:
Related to: 10