Benchmarking study demonstrates improved recommendation accuracy across eight datasets, indicating reduced posterior collapse and latent redundancy.
Key Points
To resolve posterior collapse and weak decoder capabilities in variational autoencoders applied to top-N recommendation systems.
Constructed VAEinfox using an Adaptive Compressed Representation module based on the Information Bottleneck principle, Relaxed Bernoulli distributions, and dynamic soft-mask gating.
Optimized the latent space with a Jensen-Shannon mutual information estimator and integrated a Mixture-of-Experts architecture into the generative decoder.
Benchmarked the framework against baseline models across eight real-world implicit feedback recommendation datasets.
Outperformed existing baseline models in top-N recommendation accuracy across all eight evaluated datasets.
Successfully mitigated posterior collapse while reducing latent feature redundancy and effectively processing complex user feedback.