Modeling Categorized Consumer Collections with Interlocked Hypergraph Neural Networks

Abstract

Consumers curate collections of items for various reasons and categorize them into subsets or categories based on different criteria as their collections grow. The items in a collection reflect a consumer's preferences, and the categories provide insights into the different contexts in which items are consumed. The authors develop a novel deep generative modeling framework that captures the network structure of consumer collections using multiple interlocked hypergraphs. This model employs message-passing variational autoencoders that leverage hypergraph structures and entity-specific covariates to generate probabilistic deep embeddings for consumers, items, and item categories. Applying this framework to digital music collections and playlists of music consumers, the authors demonstrate that the model outperforms several sophisticated benchmarks in predicting linkages within these collections. They then illustrate how this approach enables firms to generate novel personalized product bundles, recommend relevant items and bundles, and dynamically expand existing bundles with new items. Beyond the music application, this method is broadly applicable to other consumer collections, such as food recipes and content collections on social curation platforms like Pinterest.