Migrating Production Recommenders from Trees to Deep Learning

In fast-evolving service businesses, product catalogs are becoming increasingly dynamic, moving away from static SKUs towards bundled and discount-coupled offerings. This shift challenges traditional tree-based classifiers, which struggle with changing label spaces and multimodal data like tabular information and conversation transcripts. This research presents a practical case study of migrating a production customer support recommendation system from a gradient-boosted multiclass model to a deep learning-based pairwise-binary recommender. The migration prioritized maintaining live recommendation quality, which was achieved through several key techniques. These included reformulating the problem as pairwise binary prediction to jointly learn from user and item features, enhancing representations via negative sampling and noise injection, and efficiently incorporating long conversational context using attention pooling. The deep recommender, evaluated against a CatBoost baseline, demonstrated parity at the start of conversations and superior performance in later stages, proving the viability of deep learning for complex, dynamic recommendation tasks.