Reproduce “Personalized next-best action recommendation with multi-party interaction learning for automated decision-making” on OpenResearch

About this paper

Longbing Cao, Chengzhang Zhu

Researchers at the University of Technology Sydney developed a Reinforced Coupled Recurrent Network (CRN) to learn personalized next-best actions by explicitly modeling complex multi-party interactions and long-term dependencies. Evaluated on a real-world debt collection dataset, the model, augmented with hierarchical imbalance handling, achieved a 56.92% higher action average reward lift and reduced prediction error compared to leading deep learning baselines and domain-driven rules.

Key insights

• Modeling explicit long-term and coupled dependencies within multi-party interactions is crucial for effective personalized next-best action recommendation, moving beyond simplified Markovian assumptions.
• A dedicated Coupled Recurrent Unit (CRU) can effectively capture intricate interdependencies between different sequential data streams (e.g., decision-maker actions and client responses).
• Addressing hierarchical data imbalances through targeted weighting and sampling strategies is essential for training robust models on real-world interactive datasets, significantly boosting performance.

Summary by alphaXiv.