Abstract
Structured data such as tables, graphs, and relational records constitute the backbone of industrial analytics, healthcare informatics, finance, and engineering monitoring systems. However, learning meaningful representations from structured data remains challenging due to heterogeneity, sparsity, and limited labeled samples. Self-supervised representation learning has recently emerged as a promising paradigm to extract informative latent features without heavy reliance on manual annotations. Instead of labels, the learning objective is derived from intrinsic structure of data such as attribute relations, topology, or contextual dependencies. This paper reviews major self supervised learning (SSL) approaches for structured data including tabular data, graph-structured data, relational databases, and time-series tables. We discuss pretext task design, contrastive and generative frameworks, and hybrid architectures combining neural networks with statistical models. Comparative analysis of methods is provided along with application domains such as anomaly detection, recommender systems, predictive maintenance, and healthcare risk modeling. The study also outlines challenges including feature heterogeneity, missing values, scalability, and evaluation metrics. Finally, emerging directions such as multimodal structured SSL and foundation models for tabular learning are discussed. The review aims to provide a consolidated understanding of self-supervised representation learning for structured datasets and its role in future intelligent data systems.
Keywords: Self-supervised learning, structured data, representation learning, tabular data, graph learning, contrastive learning, pretext tasks
Full Issue
| View or download the full issue | PDF 73-89 |