Pseudo Label NCF for Sparse OHC Recommendation: Dual Representation Learning and the Separability Accuracy Trade off

A team of researchers has published a new paper, 'Pseudo Label NCF for Sparse OHC Recommendation,' introducing a novel method to solve a critical problem in Online Health Communities (OHCs). These platforms connect patients for peer support but struggle with 'cold-start' recommendations when users have almost no prior interaction data. The authors—Pronob Kumar Barman and Tera L. Reynolds—propose extending standard Neural Collaborative Filtering (NCF) models like Matrix Factorization (MF) and NeuMF with an auxiliary training objective. This objective uses structured survey data (a 16-dimensional intake vector per user) to generate 'pseudo labels' based on feature alignment with support groups, creating a second, semantically meaningful embedding space alongside the main ranking model.

The results on a dataset of 165 users and 498 support groups are significant. Using a leave-one-out protocol that mimics real-world cold-start conditions, all model variants saw major performance gains. The Multi-Layer Perceptron (MLP) model's Hit Rate at 5 (HR@5) doubled from 2.65% to 5.30%. The NeuMF model improved from 4.46% to 5.18%, and Matrix Factorization rose from 4.58% to 5.42%. Furthermore, the pseudo label embeddings were more interpretable, showing higher 'cosine silhouette scores'—a measure of cluster separability—with MF improving from 0.0394 to 0.0684.

Interestingly, the research uncovered a fundamental trade-off: models with more interpretable, well-separated embedding spaces showed a slight negative correlation with top-ranking accuracy. This 'separability-accuracy trade-off' highlights the balance between creating AI that performs well and AI whose decisions we can understand. The work demonstrates that leveraging available, non-interaction data (like surveys) through pseudo-labeling is a powerful technique for bootstrapping personalization in data-sparse environments, moving beyond traditional collaborative filtering's limitations.