Fusion AutoEncoder: Feature-Grounded Embeddings for Products with Mixed Structured Attributes

Abstract

We present Fusion AutoEncoder, a masked autoencoding architecture for learning compact, feature-grounded embeddings from heterogeneous structured attributes. Product catalogs in industrial recommender systems combine categorical descriptors, numeric commercial signals, multi-value attributes, vector-valued context, behavioral aggregates, and missing values, making reusable representations difficult to learn and costly to refresh with large text encoders. Fusion AutoEncoder addresses this challenge through feature-specific encoders, residual fusion, missing-value-aware masking, and lightweight reconstruction heads that recover clean feature groups from corrupted inputs. The configurable architecture independently represents hotels and users; evaluation on Amazon Reviews’23 Appliances metadata further validates it on a non-travel catalog with mixed structured attributes. As a downstream evaluation, we test the learned hotel embeddings on temporally bounded bookings using a controlled two-tower ranker with fixed user representations, candidates, and training. Results show that masked reconstruction improves neighborhood feature preservation and that Fusion AutoEncoder improves personalized ranking over structured-text embeddings while retaining a compact, deployment-efficient serving footprint. Our interactive RecSys demo makes these effects tangible through anonymized booking scenarios and side-by-side rankings, while illustrating the embeddings’ applicability to retrieval, ranking, and semantic-ID generation for generative recommendation.