MIT: A Multi-Tower Information Transfer Framework Based on Hierarchical Task Relationship Modeling

With the advancement of e-commerce platforms and online recommendation systems, conversion objectives have evolved from singular to diverse goals. To simultaneously enhance the performance of multiple conversion objectives, multi-task learning has become a classic solution. However, existing multi-task models either neglect task dependencies or impose manually defined sequential dependencies, subsequently transferring information based on task output vectors. Such methods fail to capture the intricate relationships between tasks in real-world scenarios and underutilize the logits outputs of tasks. In this paper, we propose the Multi-tower Information Transfer(MIT) framework, which introduces hierarchical modeling of task relationships for the first time. Specifically, we employ a data-driven approach to abstract tasks and their dependencies as nodes and edges in a graph, thereby constructing a Bayesian network of task relationships. To enhance information transfer, MIT not only refines existing vector-based methods but also pioneers a novel logits-based transfer mechanism. Extensive experiments on public and industrial datasets demonstrate the effectiveness of MIT. Furthermore, online A/B tests conducted in Huawei's online advertising platform reveal that MIT achieves RPM and CPM improvements of +2.00% and +8.87%, respectively. MIT has been fully deployed in Huawei's online advertising platform, delivering superior services to hundreds of millions of users.Our implementation is available at https://github.com/USTC-StarTeam/MIT.