From Sequential to Recursive: Enhancing Decision-Focused Learning with Bidirectional Feedback

The introduction of recursive decision-focused learning (R-DFL) marks a significant advancement in decision-making frameworks by enabling bidirectional feedback between prediction and optimization, a feature lacking in the traditional sequential decision-focused learning (S-DFL). This innovation not only enhances the final decision quality but also exhibits robust adaptability across diverse scenarios, particularly in closed-loop decision-making problems. R-DFL employs two differentiation methods: explicit unrolling via automatic differentiation and implicit differentiation based on fixed-point methods. Extensive experiments conducted on both synthetic and real-world datasets, including the newsvendor problem and the bipartite matching problem, validate the effectiveness of R-DFL. The implicit differentiation method, in particular, offers superior computational efficiency, making R-DFL a promising alternative to conventional predict-then-optimize pipelines.