Multi-task learning (MTL) has been successfully implemented in many real-world applications, which aims to simultaneously solve multiple tasks with a single model. The general idea of multi-task learning is designing kinds of global parameter sharing mechanism and task-specific feature extractor to improve the performance of all tasks. However, sequential dependence between tasks are rarely studied but frequently encountered in e-commence online recommendation, e.g. impression, click and conversion on displayed product. There is few theoretical work on this problem and biased optimization object adopted in most MTL methods deteriorates online performance. Besides, challenge still remains in balancing the trade-off between various tasks and effectively learn common and specific representation. In this paper, we first analyze sequential dependence MTL from rigorous mathematical perspective and design a dependence task learning loss to provide an unbiased optimizing object. And we propose a Task Aware Feature Extraction (TAFE) framework for sequential dependence MTL, which enables to selectively reconstruct implicit shared representations from a sample-wise view and extract explicit task-specific information in an more efficient way. Extensive experiments on offline datasets and online A/B implementation demonstrate the effectiveness of our proposed TAFE.
翻译:多任务学习(MTL)在许多现实世界应用中得到成功实施,目的是同时用单一模式解决多重任务。多任务学习的一般想法是设计各种全球参数共享机制和具体任务特征提取器,以改善所有任务的绩效。然而,任务之间的相继依赖性很少研究,但在在线电子启动建议中却经常遇到,例如对显示产品的印象、点击和转换。关于该问题的理论工作很少,大多数MTL方法中采用的有偏向优化对象使在线性能恶化。此外,在平衡各种任务之间的取舍和有效学习共同和具体的代表性方面仍然存在挑战。在本文件中,我们首先从严格的数学角度分析连续依赖性 MTL,并设计依赖性任务学习损失,以提供一个公正的优化对象。我们还提议了一个任务了解地段提取(TAFE)框架,以便从抽样角度有选择地重建隐含的共享表达方式,并以更有效的方式提取明确的任务特定信息。关于离线数据集和在线执行的实验展示了我们提议的TA/BFE的实效。