Running summarizing tools on a new article

Abstract: Uplift modeling, which estimates individual treatment effects, is increasingly important in domains such as marketing and healthcare. Existing approaches primarily focus on single outcomes, overlooking the sequential nature of multiple real-world outcomes, leading to challenges in estimation due to extreme imbalance and a funnel-shaped decrease in outcome probabilities. This paper introduces ADSeq-Uplift, an attention-based deep uplift model for imbalanced sequential outcomes. Inspired by object detection frameworks, our architecture first builds a backbone network to efficiently construct foundational feature representations and capture higher-order interactions. It then incorporates feature fusion in the neck, drawing inspiration from FPN and PAN, to enable multi-scale feature extraction. Furthermore, it integrates specialized prediction heads for causal inference tasks and introduces an attention mechanism in the sequential prediction head to model probabilistic relationships between outcomes, thereby enhancing complex dependency capture. Moreover, we employ an adaptive multi-task learning framework with dynamic loss weighting and regularization to ensure stable training and optimal performance across multiple learning objectives. To address class imbalance, we adopt a comprehensive strategy that combines two-stage stratified sampling and modified focal loss. Extensive experiments on real-world datasets demonstrate that ADSeq-Uplift significantly outperforms existing approaches, particularly in scenarios with severe class imbalance and strong temporal dependencies. Ablation studies further validate the contributions of two key components.