Unified ranking framework for feature grouping, heterogeneity, and cross-scenario scalability

Develop a unified, general-purpose ranking framework for recommender systems that simultaneously addresses feature grouping, heterogeneous input modeling, and scalable deployment across multiple scenarios, enabling robust cross-scenario transfer without relying on fragmented, scenario-specific architectures.

Background

The paper surveys the fragmentation in current recommender system ranking models, where different feature modalities (dense numerical inputs, categorical embeddings, behavioral sequences, graph relations, and multimodal content) are handled by disparate modules (MLPs, Transformers, GNNs, etc.). This architectural heterogeneity complicates scaling and hinders parameter sharing and transfer across scenarios.

Within the Related Work section, the authors explicitly state that creating a general-purpose ranking framework that jointly resolves feature grouping, heterogeneous feature handling, and cross-scenario scalability remains an open problem. MTmixAtt is proposed as a step toward this goal by integrating automatic grouping (AutoToken) and a mixture-of-experts design (MTmixAttBlock), but the broader challenge is framed as open in the literature.