Personalized Federated Recommendation With Knowledge Guidance

TL;DR

FedRKG tackles the memory-personalization trade-off in Federated Recommendation by replacing full knowledge replacement with Knowledge Guidance, updating local item embeddings toward a globally learned embedding via $\\mathbf{P}_{u} \\leftarrow \\beta \\mathbf{P}_{u} + (1-\\beta) \\mathbf{P}_g$ and interpreting this as a gradient step on a global-knowledge regularized loss $\\mathcal{L}(\\mathbf{P}) = \\frac{\\lambda}{2} \\|\\mathbf{P} - \\mathbf{P}_g\\|_2^2$. The framework is extended with Adaptive Guidance, using a per-interaction gate $g_{ui}$ to produce a guided embedding $\\mathbf{G}_{ui} = 2 \, g_{ui} \, \\mathbf{P}_{gi}$ and a nested gate-training procedure, enabling model-agnostic improvements across backbones. Theoretical justification (Proposition 1) confirms the update corresponds to a single gradient step on the quadratic penalty, while empirical results across four benchmarks show stepwise performance gains and robustness under Local Differential Privacy. Overall, FedRKG achieves the personalization benefits of dual-knowledge models with the memory footprint of single-knowledge approaches, enabling practical on-device federated recommendations with strong privacy considerations.

Abstract

Federated Recommendation (FedRec) has emerged as a key paradigm for building privacy-preserving recommender systems. However, existing FedRec models face a critical dilemma: memory-efficient single-knowledge models suffer from a suboptimal knowledge replacement practice that discards valuable personalization, while high-performance dual-knowledge models are often too memory-intensive for practical on-device deployment. We propose Federated Recommendation with Knowledge Guidance (FedRKG), a model-agnostic framework that resolves this dilemma. The core principle, Knowledge Guidance, avoids full replacement and instead fuses global knowledge into preserved local embeddings, attaining the personalization benefits of dual-knowledge within a single-knowledge memory footprint. Furthermore, we introduce Adaptive Guidance, a fine-grained mechanism that dynamically modulates the intensity of this guidance for each user-item interaction, overcoming the limitations of static fusion methods. Extensive experiments on benchmark datasets demonstrate that FedRKG significantly outperforms state-of-the-art methods, validating the effectiveness of our approach. The code is available at https://github.com/Jaehyung-Lim/fedrkg.

Figures (7)

• Figure 1: A conceptual comparison between (a) Single Knowledge, (b) Dual Knowledge, and (c) Guided Knowledge (Ours)
• Figure 2: Convergence on Amazon-Video (HR@10 vs. rounds) with a FedMF backbone. Methods: Full Replacement (single-knowledge), Local Only, Dual-Knowledge (FedRAP), and Knowledge Guidance (ours). Guidance every 100 rounds yields stepwise improvements and the best final HR@10 while retaining a single-embedding inference footprint.
• Figure 3: Overall framework of FedRKG. On each client $u$: (i) Before Guidance, the local model is trained with $\mathcal{L}_{rec}$ (Eq.\ref{['eq:bce']}); (ii) Nested Training, a lightweight gate is fit (main parameters frozen) to produce an adaptive guidance vector; (iii) During Guidance, global knowledge is injected into the single local embedding set using guidance vector (applied twice here), avoiding full replacement and preserving personalization.
• Figure 4: Convergence curves of FedRKG.
• Figure 5: Performance under Local Differential Privacy on FilmTrust.

Theorems & Definitions (1)

• proof