Existing retrieval (aka candidate generation) system for recommendation is highly fragmented, with typically 100s of heuristic-based candidate generators. Because the majority of the output from each generator is duplicated with those from other generators, most of the computation used for training and serving these generators are wasted. Maintaining such a fragmented highly complex system is also a significant debt. Moreover, because these generators are trained independently with each relying on a specific type of narrowly-focused heuristic, the performance of such a retrieval system is highly suboptimal.

GFR aims to provide a unified modeling paradigm for retrieval. By unifying and integrating a large variety of heuristics, including but not limited to those used by the current generators, in a single framework, GFR enables us to automatically learn expressive retrieval models to improve accuracy. The universality will also greatly simplify the retrieval modeling tech stack, and lower infra cost for training and serving.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.