Implementation of Generalist Recommender (GRec), an end-to-end efficient and scalable recommender system designed to train a single model that could generalize across multiple search & recommender tasks while at the same time being highly efficient.

This work was contributed by the Verizon AI Center team.

GRec has the following highlights:

• Input Modalities - Designed to handle inputs of multi-modalities (including categorical and numerical data, texts, and images) by utilizing NLP heads, parallel Transformers, as well as wide and deep in the model architecture.
• Highly Efficient - Adopts a newly proposed task-sentence level routing mechanism to scale the model capabilities on multiple tasks without compromising performance.
• Production Ready - Successfully deployed at Verizon and demonstrates significant performance improvement over the baseline in both offline and online settings.