因果指导下可解释稳定机器学习研究

数据中变量间复杂的关联关系可分为因果关联和虚假关联，其中因果关联是可解释且在不同数据中是不变的，而虚假关联会随着数据变化而变化。传统关联驱动的机器学习未能区分虚假关联和因果关联，基于虚假关联学习导致模型存在不可解释和预测不稳定等问题。因果推理是洞悉观测数据中因果关联的重要途径，但是在大数据背景下，因果推理面临着全新的挑战。为了提升机器学习的可解释性和稳定性，项目拟研究因果指导下可解释稳定机器学习理论与方法。研究内容包括：针对大数据背景下高维变量、变量差异性和连续干预变量，研究大数据驱动的因果推理算法；针对数据中的虚假关联和因果关联，研究因果关联甄别算法，并用于指导机器学习，实现可解释稳定学习；最后将研究的理论成果与算法应用于电商商品可解释推荐问题中，体现研究成果的优越性和实用性。项目目标是突破传统关联驱动学习的局限，通过融合因果推理和机器学习，实现因果指导下的可解释稳定学习。

The correlations in data can be divided into causation and spurious correlation, where causation is interpretable and invariant across data, while spurious correlation might vary across data. Traditional correlation-driven machine learning fails to distinguish causation from spurious associations. Models learned based on spurious associations would be unexplainable and unstable. Causal Inference is an important way to mine causation from data, but it is facing new challenges in the Big Data era. Aiming at improving the interpretability and stability of machine learning, this project intends to marry causal inference with machine learning and study causality for explainable and stable machine learning. Specifically, in causal inference level, we propose big data-driven causal inference algorithms by addressing new challenges from high-dimensional variables, variable differentiation and continuous treatment variables in the big data era; in machine learning, we propose a series causality based algorithm for screening causation from spurious correlation, and used to guide machines Learning to achieve explainable and stable learning; finally, we apply our study to the e-commerce for explainable and stable recommendation. The goal of this project is to break through the limitations of traditional correlation-driven learning and achieve explainable and stable learning by marrying causal inference and machine learning.