Embeddings as a Bridge Between Neural and Classical Models: A Systematic Study of Representation Transfer in Tabular Recommendation Systems

This paper investigates whether neural network–learned categorical embeddings can improve classical machine learning models when transferred as pre-computed features in recommendation systems. A controlled three-condition ablation study is conducted across the MovieLens, Amazon Reviews, and Yelp datasets using Linear Regression, Random Forest, XGBoost, and Artificial Neural Networks. Results show that embedding transfer can improve performance, but the magnitude of benefit depends on model architecture and baseline overfitting severity. Tree-based models benefit most when embedding transfer is combined with hyperparameter optimization, while neural models gain from replacing sparse one-hot encodings with dense representations. The study also highlights the importance of target leakage prevention through Leave-One-Out aggregation. Findings provide practical guidance for applying representation learning in tabular recommendation systems.