Sezdi A., Bilgin M.
Orclever Proceedings of Research and Development, cilt.7, sa.1, ss.16-29, 2025 (Düzenli olarak gerçekleştirilen hakemli kongrenin bildiri kitabı)
Özet
<p>This study models customer–vehicle interactions in an online used-car platform as a bipartite structure, constructing a graph with customer (U) and vehicle (V) nodes. Relations between the two node sets are defined only by edges representing realized purchase events (e=(u,v,t)), thereby focusing on a signal with high business value and relatively low noise. On this graph, inductive node representations (embeddings) are learned with GraphSAGE. During training, link prediction is used solely as a self-supervised proxy task; optimization employs an MLP-based scorer with Binary Cross-Entropy (BCE) loss. Early stopping is triggered when the BCE on a temporally held-out validation set stops improving; together with temporal negative sampling, this prevents leakage of future information.</p>
<p>The objective is to obtain high-quality customer/vehicle embeddings. The learned representations are then used to construct embedding-based customer segments via K-Means. Segmentation quality is evaluated using the Silhouette and Calinski–Harabasz scores. The results show that GraphSAGE embeddings learned on the purchase-induced bipartite graph provide a practical and scalable foundation for recommendation/targeting and customer understanding tasks</p>