Aug 2026· Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2· pp. 3739-3750· 1 citation· 30 references
TL;DR
CanonFedRec is proposed, a novel framework that achieves superior performance while reducing memory costs by up to 40× compared to the best FedRec approach and treats item-wise variance as a proxy for client-side cognitive divergence, and dynamically adapting optimization objectives via elastic decision boundaries.
Abstract
With the increasing demand for privacy protection, the federated recommendation (FedRec) has become a critical research topic. Prevailing frameworks continuously decouple item representation to bypass the heterogeneity of client data, sacrificing the universality of aggregated item embeddings, losing the collaborative benefits of federated learning. This trend not only introduces prohibitive memory costs on resource-constrained edge devices, but also exacerbates the isolation of user interests, leading to a significant loss of recommendation diversity. In this paper, we attribute the failure of aggregated item embeddings to the server-side Global Geometric Misalignment and client-side Local Filtering Bubbles after empirical and theoretical analysis. Then, we propose CanonFedRec, a novel framework that resolves the above challenges through a canonical geometric perspective. First, we shift the optimization landscape from Euclidean space to a unit hypersphere to explicitly decouple local magnitude noise unrelated to popularity from semantic relationships. To rectify the global geometric inconsistency, we propose a Riemannian Projection-based Aggregation mechanism that performs global updates in a geometrically compatible tangent space. Furthermore, to mitigate the local filtering bubbles of clients, we propose a Divergence-Aware Elastic Alignment mechanism. This treats item-wise variance as a proxy for client-side cognitive divergence, and dynamically adapting optimization objectives via elastic decision boundaries. Extensive experiments on four datasets demonstrate that CanonFedRec not only significantly enhances the representational power of aggregated item embeddings, but also achieves superior performance while reducing memory costs by up to 40× compared to the best FedRec approach.
Federated Recommendation (FedRec) enables privacy-preserving recommendation by training models across user devices without sharing raw interaction data. However, personalized FedRec still faces a trade-off between personalization and resource efficiency. In this paper, we revisit personalization from the perspective of the local optimization horizon. We find that sufficiently long local optimization can form user-specific item structures, but frequent synchronization or full replacement can disrupt them. Based on this observation, we propose
FedRKG
, a model-agnostic framework that forms personalized item embeddings through
Long-Horizon Local Optimization
and injects common global knowledge through intermittent
Regularized Knowledge Guidance
. This design incorporates shared collaborative signals without replacing local embeddings or maintaining dual representations. We further introduce
Adaptive Guidance
to control the influence of global knowledge at the user–item interaction level. Extensive experiments show that FedRKG consistently outperforms state-of-the-art FedRec baselines while reducing storage and communication costs. Our code is available at
https://github.com/Jaehyung-Lim/FedRKG_JIIS26_SCIE
.
Jaehyung Lim, Wonbin Kweon, Woojoo Kim et al.· Journal of Intelligence and...· 0 citations
A federated learning, which is merged with GNN to allow decentralised training whilst maintaining the relational structure of data in terms of interaction, is proposed, which demonstrates the usefulness of federated graph-based learning in order to have secure and accurate recommendation systems.
Kolan Helini, P.Salini· International journal of com...· 0 citations
While Cross-Domain Sequential Recommendation (CDSR) has proven effective in mitigating data sparsity and enhancing accuracy, its impact on recommendation diversity remains largely unexplored. We are the first to reveal a counterintuitive phenomenon: while CDSR improves accuracy, it often comes at the cost of diversity, confining users to a narrower scope of interests. Through rigorous empirical experiments and theoretical analysis, we pinpoint two fundamental determinants driving this decline: (1) Domain Homogeneity, where excessive similarity between domains enforces preference redundancy; and (2) Information Asymmetry, where insufficient signal from the source domain fails to meaningfully perturb target-domain distributions. To address these challenges, we propose DivCDSR, a novel model-agnostic framework designed to enhance diversity in CDSR. Specifically, we introduce a dual-prototype semantic constraint mechanism that mitigates the homogenization trap via intra-domain clustering with orthogonalization and inter-domain separation. Furthermore, we devise a dual-guided diffusion module that augments source-domain information by generating new sequences, thereby resolving the information asymmetry issue. Extensive experiments on three public datasets demonstrate that our DivCDSR significantly enhances diversity metrics while maintaining or even improving recommendation accuracy simultaneously. % , offering a robust solution to the informational poverty inherent in conventional CDSR models. All datasets and code are available at https://github.com/Asuei-cs/DivCDSR.
Shu Chen, Yuhan Zhao, Weixin Chen et al.· Proceedings of the 32nd ACM...· 0 citations
Low-Rank Adaptation (LoRA) enables communication-efficient federated fine-tuning of pretrained language models. However, integrating differential privacy (DP) into federated LoRA remains challenging: independently perturbing and aggregating its two low-rank matrices can cause aggregation mismatch and the quadratic noise term. Existing methods mitigate these issues by freezing one low-rank matrix but still rely on Euclidean aggregation, which is basis-dependent and may distort the global update. To address this limitation, we propose FedGSA, a geometry-consistent aggregation framework for differentially private federated LoRA. FedGSA represents each privatized client update as a basis-invariant subspace on the Grassmann manifold. In each communication round, clients extract low-dimensional subspaces capturing dominant update directions and encode them as projection matrices. The server aggregates these representations to estimate a geometry-consistent global update subspace and reconstructs the global LoRA factors within it, reducing distortion caused by basis misalignment, privacy noise, and heterogeneous client updates. We prove that FedGSA incurs no additional privacy loss beyond client-side DP training and establish its convergence under standard assumptions. Experiments on four GLUE tasks and a language generation benchmark demonstrate consistent improvements across privacy budgets and degrees of data heterogeneity. In particular, FedGSA improves average accuracy over the strongest baseline by 2.17% and 2.27% under $\epsilon=6$ and $\epsilon=3$, respectively.
A new FAL framework is proposed that utilizes federated representation learning to align client data in a shared embedding space that achieves performance that surpasses existing FAL methods even when they are given substantially larger annotation budgets, demonstrating the value of centralized coordination under privacy constraints.
Federated recommendation (FedRec) enables personalized modeling without centralizing users'interaction histories, but most existing methods assume a fixed item pool and thus overlook the practical cold-item setting where new items continuously arrive. Under the dual-sided constraint, where the server cannot access clients'interactions while clients cannot access the server's proprietary item attribute features, prior federated cold-start recommendation approaches suffer from three structural limitations: a lack of personalization, compositionality failure caused by forcing heterogeneous semantics into a single embedding space, and training- and communication-inefficiency arising from explicit alignment between separate collaborative and attribute representations. To address these challenges, we propose Personalized and Multi-view Representation for Federated Cold-Start Recommendation (PMFRec). PMFRec learns a personalized representation generator to produce user-specific item representations from attribute features, and introduces a global multi-view encoder with item-adaptive gating and an orthogonality objective to capture complementary semantic views while reducing cross-view redundancy. In addition, PMFRec fuses collaborative and attribute knowledge into a single exchanged item representation, eliminating the need for an explicit client-side regularizer and reducing communication overhead. Extensive experiments on real-world datasets show that PMFRec consistently outperforms strong baselines in cold-item recommendation and further improves user-level fairness, warm-scenario adaptability, and robustness under Local Differential Privacy (LDP).
Jaehyung Lim, Wonbin Kweon, Woojoo Kim et al.· 0 citations