Dorsaf Sellami

Dorsaf Sellami

Auto-generated profile, based on publicly available information

Dorsaf Sellami

Lorraine University,Inria, LORIA,Nancy,France

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
19 December 2024 in Computer Science Review
Verified
Journal Article
Closed access
DSDorsaf Sellami
Dorsaf Sellami
WIWissem Inoubli
Wissem Inoubli
IFimed riadh farah
imed riadh farah
SASabeur Aridhi
Sabeur Aridhi

8 December 2025
Verified
Conference Paper
Closed access
DSDorsaf Sellami
Dorsaf Sellami
WIWissem Inoubli
Wissem Inoubli
IFimed riadh farah
imed riadh farah
SASabeur Aridhi
Sabeur Aridhi
The exponential growth of knowledge graphs (KGs) poses significant challenges for scalable graph neural network training, primarily due to the memory and computational bottlenecks of centralized processing. Distributed training, which relies on graph partitioning, emerges as a necessary solution. However, traditional partitioning methods like METIS and KaHIP are designed for homogeneous graphs and fail to preserve the rich, multi-relational semantics inherent in KGs, leading to fragmented subgraphs that degrade model performance and increase communication overhead. To address this, we propose SA-KGP (Semantic-Aware Knowledge Graph Partitioning), a novel partitioning framework that incorporates the semantic importance of relations into the partitioning logic. SA-KGP introduces a dynamic relation-weighting scheme and a Semantic Vertex Partition (SVP) score to guide entity placement, followed by iterative refinement through Semantic-Aware Label Propagation (SALP) and a Stabilization Phase. We evaluate our method within a federated learning architecture for training R-GCN models on standard benchmarks. Experimental results demonstrate that SA-KGP significantly outperforms traditional methods, achieving superior link prediction accuracy and near-linear training speedup by effectively preserving semantic cohesion and minimizing inter-partition communication.