G-HIN2Vec: Distributed heterogeneous graph representations for cardholder transactions - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

G-HIN2Vec: Distributed heterogeneous graph representations for cardholder transactions

Farouk Damoun
Hamida Seba
Jean Hilger
Radu State

Résumé

Graph related tasks, such as graph classification and clustering, have been substantially improved with the advent of graph neural networks (GNNs). However, existing graph embedding models focus on homogeneous graphs that ignore the heterogeneity of the graphs. Therefore, using homogeneous graph embedding models on heterogeneous graphs discards the rich semantics of graphs and achieves average performance, especially by utilizing unlabeled information. However, limited work has been done on whole heterogeneous graph embedding as a supervised task. In light of this, we investigate unsupervised distributed representations learning on heterogeneous graphs and propose a novel model named G-HIN2Vec, Graph-Level Heterogeneous Information Network to Vector. Inspired by recent advances of unsupervised learning in natural language processing, G-HIN2Vec utilizes negative sampling technique as an unlabeled approach and learns graph embedding matrix from different pre-defined meta-paths. We conduct a variety of experiments on three main graph downstream applications on different socio-demographic cardholder features, graph regression, graph clustering, and graph classification, such as gender classification, age, and income prediction, which shows superior performance of our proposed GNN model on real-world financial credit card data.

Dates et versions

hal-04238052 , version 1 (11-10-2023)

Identifiants

Citer

Farouk Damoun, Hamida Seba, Jean Hilger, Radu State. G-HIN2Vec: Distributed heterogeneous graph representations for cardholder transactions. SAC '23: 38th ACM/SIGAPP Symposium on Applied Computing, Mar 2023, Tallinn Estonia, France. pp.528-535, ⟨10.1145/3555776.3577740⟩. ⟨hal-04238052⟩
24 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More