Federated Personal Knowledge Graph Completion with Lightweight Large Language Models for Personalized Recommendations

Fernando Spadea; Oshani Seneviratne

cs.LGFeb 25, 2026

Federated Personal Knowledge Graph Completion with Lightweight Large Language Models for Personalized Recommendations

Authors: Fernando Spadea, Oshani Seneviratne

Organizations: Rensselaer Polytechnic Institute, Troy, NY, USA

Abstract

Personalized recommendation increasingly relies on private user data, motivating approaches that can adapt to individuals without centralizing their information. We present Federated Targeted Recommendations with Evolving Knowledge graphs and Language Models (FedTREK-LM), a framework that unifies lightweight large language models (LLMs), evolving personal knowledge graphs (PKGs), federated learning (FL), and Kahneman-Tversky Optimization to enable scalable, decentralized personalization. By prompting LLMs with structured PKGs, FedTREK-LM performs context-aware reasoning for personalized recommendation tasks such as movie and recipe suggestions. Across three lightweight Qwen3 models (0.6B, 1.7B, 4B), FedTREK-LM consistently and substantially outperforms state-of-the-art KG completion and federated recommendation baselines (HAKE, KBGAT, and FedKGRec), achieving more than a 4x improvement in F1-score on the movie and food benchmarks. Our results further show that real user data is critical for effective personalization, as synthetic data degrades performance by up to 46%. Overall, FedTREK-LM offers a practical paradigm for adaptive, LLM-powered personalization that generalizes across decentralized, evolving user PKGs.

Figures & tables

Figure 1: Overall performance results of our TREK-LM models against the baselines (KBGAT, HAKE, and FedKGRec).

Figure 2: Overview of the FedTREK-LM architecture.

Panel of Figure 3 — Figure 3: Detailed view of the FedTREK-LM local operations: (a) shows the feedback-driven fine-tuning using KTO, and (b) illustrates how the LLM reasons over the PKG to derive recommendations.

Figure 4: Illustration of KTO data construction in the form of prompt–completion–label triplets from user-LLM interactions.

Model	Trainable Parameters	Per-Round Client Cost	128-Round Total Server Cost
Qwen3-0.6B	10.093 M	38.50 MB	38.5 GB
Qwen3-1.7B	17.433 M	66.50 MB	66.5 GB
Qwen3-4B	33.030 M	126.00 MB	126.0 GB

Table 1: Communication costs for FedTREK-LM models between the individual client overhead per active round and the total aggregated server load.

Model	Setting	Precision	Recall	F1-score	MRR	Hits@1	Hits@3	Hits@10
TREK-LM	Centralized	0.149	0.301	0.199	0.247	0.192	0.299	0.321
	Centralized +Syn	0.122	0.240	0.162	0.209	0.177	0.238	0.255
	Federated	0.238	0.251	0.244	0.284	0.280	0.290	0.290
	Federated +Syn	0.090	0.247	0.132	0.218	0.183	0.254	0.270
	Local-only	0.015	0.068	0.025	0.069	0.056	0.081	0.081
	Local-only +Syn	0.015	0.068	0.025	0.069	0.056	0.081	0.081

Table 2: Comparative performance on the Movie PKG recommendation task. We compare our TREK-LM variants (using the Qwen3-4B model) across centralized, federated, and local-only settings, with and without synthetic data ( +Syn ). Baselines include traditional KG completion models (KBGAT and HAKE) and a federated recommendation model (FedKGRec). The best result in each column is bolded.

Figure 5: TREK-LM Performance by Setting (Qwen3-4B)

Figure 8: Impact of Synthetic Data Across All Models

Model	Setting	F1-score	MRR	Hits@10
TREK-LM	Centralized	0.000	0.000	0.000
	Federated	0.002	0.002	0.000
	Local	0.000	0.000	0.000
KBGAT	Cen./Fed.	< 0.004	< 0.017	< 0.019
HAKE	Cen./Fed.	< 0.002	< 0.003	< 0.007
FedKGRec	Federated	0.000	0.000	0.000

Table 3: Performance on the Recipe PKG recommendation task. All TREK-LM variants and baselines achieved near-zero performance across all metrics.

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 514 papers

Guiding Federated Graph Recommendation with LLM-encoded knowledge
Jun 13, 2026Thi Minh Chau Nguyen, Hien Trang Nguyen, Duc Anh Nguyen +3Federated Learning Aggregation Graph Representation Learning
L2Rec: Towards Dual-View Understanding of LLMs for Personalized Recommendation
May 26, 2026Pingjun Pan, Tingting Zhou, Peiyao Lu +3Mixture-of-Experts Language Models LLM Personalization
Netease Cloud Music Hangzhou, China
AdaptFuse: Training-Free Sequential Preference Inference via Externalized Bayesian Computation
Apr 5, 2026Fangzhou Lin, Zilin Dai, Shuo Xing +8Preference Learning Large Language Model-Based Recommendation
Texas A&M University · Worcester Polytechnic Institute · Tohoku University +1

Jun 13, 2026J/K move · Enter open · S save

Guiding Federated Graph Recommendation with LLM-encoded knowledge

Thi Minh Chau Nguyen, Hien Trang Nguyen, Duc Anh Nguyen, Van Ho-Long+2

Federated Learning Aggregation Graph Representation Learning Non-IID Federated Learning Federated Graph Learning

Open paper AskPDF arXiv

Graph-based recommender systems are highly effective at extracting collaborative signals from user--item interactions, and federated learning (FL) allows these models to be trained while preserving user privacy. However, aggregating graph representations across distributed, non-IID clients remains a challenge; structural embeddings learned locally often misalign, and naive averaging fails to capture meaningful cross-client relationships. Most existing federated graph methods rely exclusively on structural aggregation, neglecting the rich, global semantic context available in large language models (LLMs). In this paper, we propose a novel framework that uses LLM-encoded knowledge to guide federated graph recommendation. Specifically, clients learn structural representations from local graphs while simultaneously summarizing their typical interaction patterns into compact semantic vectors via a frozen LLM. The central server then uses these LLM-encoded semantic signals to discover related preference patterns across clients, guiding the selective aggregation of their structural representations. This enables semantically informed cross-client collaboration without exposing raw data. Extensive experiments on standard benchmarks show that guiding structural alignment with LLM-encoded knowledge consistently improves recommendation accuracy over existing federated graph baselines.

Abstract

Recommender Systems

Explore similar work

Guiding Federated Graph Recommendation with LLM-encoded knowledge

L2Rec: Towards Dual-View Understanding of LLMs for Personalized Recommendation

AdaptFuse: Training-Free Sequential Preference Inference via Externalized Bayesian Computation