Robust Learning Architecture For Incomplete Data Environments Through Positive-Unlabeled Recommendation Learning

Abstract

Incomplete and biased data environments present significant challenges in domains such as Computer Vision (CV), Natural Language Processing (NLP), and Recommender Systems (RS), where the absence of explicit negative labels reduces prediction reliability and decision-making accuracy. Existing approaches for handling implicit feedback primarily rely on pseudo-negative sampling, confidence screening, and preference inference methods. However, the Missing-Not-At-Random (MNAR) features of incomplete data and the impact of hidden confounding variables are not sufficiently resolved by these methods, leading to poor applicability and limited resilience in practical applications.  An incomplete Interaction Dataset employed for experimental evaluation, 12,000 user-item interactions, and 49 attributes make up the interaction data. Min-Max normalization and missing interaction imputation techniques are used during data preprocessing to improve data quality and consistency. Feature extraction is performed using a Variational Autoencoder (VAE) to learn latent embeddings and interaction-based features for capturing hidden behavioral patterns. A robust Deer Hunting Optimized Light Gradient Boosting (DHO-LGB) model is proposed to address incomplete data environments effectively.  Positive-Unlabeled (PU) with exposure imputation and doubly robust debiasing reduces bias effectively. DHO maximizes performance by finding optimal features and parameters. LGBM allows fast classification with great accuracy and scalability. DHO helps optimize LGBM by effectively choosing the best features and improving predictions. The model is implemented using Python and TensorFlow. The experimental findings show that the DHO-LGB model developed performs better with an accuracy of 0.95, recall of 0.94, ILD of 0.92, and coverage of 0.96 with superior robustness, reliability, and suggestion efficacy in comparison to current cutting-edge methods.