Personalized Federated Recommendation With Knowledge Guidance
Jaehyung Lim, Wonbin Kweon, Woojoo Kim, Junyoung Kim, Dongha Kim, Hwanjo Yu
TL;DR
This paper introduces FedRKG, a federated recommendation framework that enhances personalization by fusing global knowledge into local embeddings, using adaptive guidance to dynamically optimize user-item interactions, all while maintaining a single-knowledge memory footprint.
Contribution
We propose a novel knowledge guidance framework with adaptive modulation that improves federated recommendation personalization without increasing memory requirements.
Findings
FedRKG outperforms state-of-the-art methods on benchmark datasets.
The adaptive guidance mechanism effectively balances personalization and knowledge integration.
Our approach maintains a low memory footprint suitable for on-device deployment.
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…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1. The paper is well-structured and written fluently, with clear logic in introducing the research background, problem formulation, framework design, and experimental validation. 2. The proposed FedRKG framework is simple and effective: it achieves dual-knowledge-level personalization while maintaining a single-knowledge memory footprint through KG, avoiding the memory overload of dual-knowledge models, which aligns with the practical demand for on-device FedRec deployment. 3. The framework exhi
1. Inconsistencies exist between the performance results of FedRKG in Table 1 and Table 2. For example, in the Amazon-Video dataset: Table 1 reports FedRKG’s N@5=15.74 and N@10=17.50, while Table 2 (ablation study) shows FedRKG’s N@5=14.85 and N@10=19.46; The paper provides no explanation for this discrepancy. 2. The contribution of AG is contradictory across different experiments, and the paper lacks in-depth analysis of the degree of its contribution and applicable scenarios. In the ablation
1. The paper diagnoses the dilemma between suboptimal knowledge replacement practice in single-knowledge (only local item embeddings) and doubling memory in dual-knowledge models (local and global item embeddings). 2. A theoretical account of knowledge guidance is provided.
1. In section 4.3, Eq. (5) inputs to the user-specific gating network are dimensionally incompatible. 2. Experimental Limitations： (1) Datasets vary across experiments. (Amazon-Video for cold/warm users, FilmTrust for LDP). (2) Incomplete ablation (only Amazon-Video and LastFM-2K). (3) Inconsistent FedRKG scores between Table 1 and Table 3.
1. The gating network learns the optimal knowledge fusion strength for each user–item pair, enabling the model to dynamically adapt to user heterogeneity and improve recommendation accuracy. 2. The proposed mechanism can be seamlessly integrated into various federated recommendation frameworks (e.g., FedMF, FedNCF, PFedRec), exhibiting strong model-agnosticism and scalability. 3. Experimental results show that the method consistently improves performance across multiple datasets and outperform
1. Lack of novelty: The proposed concept of “knowledge guidance” essentially fuses the global and local models, which highly overlaps with existing personalized federated learning approaches such as FedALA[1]. The conceptual innovation is therefore limited. Moreover, the notion that directly substituting the local model with the global one is suboptimal in federated settings is already well known; model merging has long been recognized as a standard solution to address client heterogeneity. The
- The paper is easy to follow. - The experiments are solid and extensive.
The claimed novelty of this paper is rather limited. First, the authors repeatedly emphasize the notions of single-knowledge and dual-knowledge models (e.g., Line 13 and Line 41). However, these concepts directly correspond to the well-established global model and personalized model formulations in Personalized Federated Learning (PFL). The paper simply redefines existing PFL terminology without introducing any conceptual advancement. Second, the proposed Knowledge Guidance mechanism is essent
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsRecommender Systems and Techniques · Privacy-Preserving Technologies in Data · Machine Learning in Healthcare