Gaining Deeper Insights into Thin Profile Consumers via Attributed Graph Contrastive Learning

Abstract

An essential task for marketers involves accurately understanding consumer preferences to enhance engagement. This becomes challenging when data on consumer attributes and interactions are limited, as is often the case for newly established businesses and nonprofit organizations. The growing emphasis on consumer privacy and the restriction of granular data from third-party sources further exacerbate these limitations. To address these, the authors propose the Attributed Graph Contrastive Learning framework, a graph representation learning approach tailored to thin-profile consumers - those with limited information - without requiring additional data collection. The proposed approach leverages Graph Convolutional Networks to augment limited information by leveraging information from both directly and indirectly connected consumers and products. By incorporating available attributes, the approach captures contextual connections that improve learning. Contrastive learning further enhances the model’s ability to distinguish key similarities and differences within the data, maximizing the utility of existing information. Applied to real-world data from a nonprofit organization, the framework uncovers distinct consumer segments otherwise unattainable from sparse observable information alone. By borrowing richer information from similar consumers and products through graph structures and refining it with contrastive learning, the framework generates more comprehensive embeddings, leading to actionable insights for better engaging thin-profile consumers.

Keywords: Deep learning, Graph representation learning, Graph convolutional networks, Contrastive learning, Data sparsity, Thin-profile consumers, Segmentation, Nonprofit organizations