2022因果推断与机器学习研讨会

题目： Robust and efficient estimation for treatment effect in causal inference

摘要: We develop new methods to evaluate various characteristics for treatment effect in causal inference, including but not limited on average treatment effect, median treatment effect, the Mann-Whitney statistic and etc. The proposed method combines the efficiency of model-based method and robustness of the nonparametric approach, including deep learning methods. It requires few model assumptions and is shown to be efficient if all specifications are correct, and doubly robust if some part is misspecified. Extensive numerical studies have been presented to demonstrate the advantages of the proposed method over others. *Joint work with Ling Zhou, Fanyin Zhou, Qiuxia Wang and Jing Qin

个人简介：西南财经大学教授，统计研究中心主任。国际数理统计学会IMS-fellow，教育部长江学者特聘教授，国家杰出青年科学基金获得者，国家百千万人才工程获得者，享受国务院政府特殊津贴专家。主要研究方向为非参数方法、转换模型、生存数据分析、函数型数据分析、潜变量分析、时空数据分析。研究成果发表在包括国际统计学四大顶级期刊AoS、JASA、JRSSB、Biometrika和计量经济学顶级期刊JOE及JBES上。先后多次主持国家基金项目，包括国家杰出青年基金及自科重点项目。林华珍教授是国际IMS-China、IBS-CHINA及ICSA-China委员，中国现场统计研究会数据科学与人工智能分会理事长，第九届全国工业统计学教学研究会副会长，中国现场统计研究会多个分会的副理事长。先后是国际统计学权威期刊《Biometrics》、《Scandinavian Journal of Statistics》、《Journal of Business & Economic Statistics》、《Canadian Journal of Statistics》、 《Statistics and Its Interface》、《Statistical Theory and Related Fields》的Associate Editor， 国内权威或核心学术期刊《数学学报》（英文）、《应用概率统计》、《系统科学与数学》、《数理统计与管理》编委会编委。

题目：因果启发的稳定学习

摘要：近年来人工智能技术的发展，在诸多垂直领域取得了性能突破。但当我们将这些技术应用于医疗、司法、工业生产等风险敏感领域时，发现当前人工智能在稳定性、可解释性、公平性、可回溯性等“四性”方面存在严重缺陷。究其深层次原因，当前统计机器学习的基础——关联统计自身不稳定、不可解释、不公平、不可回溯可能是问题的根源。相对于关联统计，因果统计在保证“四性”方面具有更好的理论基础。但如何将因果统计融入机器学习框架，是一个开放并有挑战的基础性问题。本报告中，讲者将重点介绍将因果推理引入预测性问题所提出的稳定学习理论和方法，及其在解决OOD泛化问题方面的机会和挑战。

个人简介：崔鹏，清华大学计算机系长聘副教授，博士生导师。研究兴趣聚焦于大数据驱动的因果推理和稳定预测、大规模网络表征学习等。在数据挖掘及人工智能领域顶级国际会议发表论文100余篇，先后5次获得顶级国际会议或期刊论文奖，并先后两次入选数据挖掘领域顶级国际会议KDD最佳论文专刊。担任IEEE TKDE、ACM TOMM、ACM TIST、IEEE TBD等国际顶级期刊编委。曾获得国家自然科学二等奖、教育部自然科学一等奖、CCF-IEEE CS青年科学家奖、ACM杰出科学家。

题目：Deconfounding with the Blessing of Dimensionality

摘要：In this talk, I will give a selective overview of deconfounding methods that exploit the blessing of dimensionality in regression and graphical models with latent variables. For regression problems, I will discuss factor adjustment and spectral deconfounding methods. For graphical models, I will emphasize the low-rank plus sparse matrix decomposition approach. Finally, I will present two recent examples in compositional data analysis where surprisingly simple methods work extremely well.

个人简介：林伟，现任北京大学数学科学学院概率统计系、统计科学中心长聘副教授。2011年获南加州大学应用数学博士学位，2011至2014年在宾夕法尼亚大学做博士后研究，2014年加入北京大学。主要从事高维统计和统计机器学习的理论与应用研究，代表性成果发表在JASA、Biometrika、Biometrics、IEEE TIT、Operations Research、Environmental Science & Technology、《中国科学：数学》等统计学及相关领域顶级期刊上。2015年入选国家高层次人才计划青年项目，主持国家重点研发计划课题、北京市自然科学基金重点研究专题项目、国家自然科学基金面上项目等。

题目：因果推断，观察性研究和诺贝尔经济学奖

摘要：诺贝尔经济学奖2021年授予Card, Angrist, 和Imbens，以表彰他们在经济学的实证研究和因果推断方法方面的贡献。三位经济学家获奖的科学背景是观察性数据的因果推断。观察性研究有两大挑战，一是混杂因素，二是选择偏差。Card，Angrist和Imbens获得诺贝尔经济学奖的主要成果是使用恰当的自然试验、工具变量解决劳动经济学中的几个百年难题。在此之前，1989年Haavelmo和2000年Heckman获诺贝尔奖的主要贡献都与因果推断、选择偏差密切相关。我将结合这几届诺贝尔经济学奖的科学背景，介绍因果推断和缺失数据领域的一些进展，包括我们提出的代理推断以及在非随机缺失数据和回调数据分析方面的工作。

个人简介：苗旺现为北京大学概率统计系助理教授、研究员，2018–2020曾在北大光华管理学院任助理教授，2008–2017 年在北京大学读本科和博士，2017–2018 在哈佛大学生物统计系做博士后研究。研究兴趣包括因果推断，人工智能，缺失数据分析和半参数统计，以及在生物医药，流行病学和社会经济中的应用。个人网页https://www.math.pku.edu.cn/teachers/mwfy/

题目：Supervised Causal Learning: A New Frontier of Causal Discovery

摘要：Supervised Causal Learning (SCL) aims to learn causal relations from observational data by accessing previously seen datasets associated with ground truth causal relations. The supervision-based method enjoys the benefit of “free” acquisition of training data: with forward sampling techniques, we can generate additional datasets from as many synthetic causal graphs as needed. In this talk, I will first discuss a fundamental question of SCL: What are the benefits from supervision and how does it benefit? And its relationship with causal structure identifiability. Then a general two-phase learning paradigm for SCL is advocated. Following this paradigm, I will share one SCL approach which works remarkably well on discrete data. It shows promising results on validating the effectiveness of supervision for causal learning.

个人简介：Rui Ding is a principal researcher in the DKI (Data, Knowledge, Intelligence) area at Microsoft Research Asia. His main focus is on robust and automated data analytics. Rui Ding is leading the research on AutoInsights. AutoInsights aims at automatically discovering interesting and meaningful data patterns, which are further equipped with techniques from causality and XAI domains to provide deeper insights, explanation and actions to bridge the gap towards automatic and robust data analytics. Key technologies have been/are being shipped to Microsoft Power BI and Microsoft Office (Excel, Forms).

题目： Generalized regression estimators for average treatment effect with multicollinearity in high-dimensional covariates

摘要：To obtain efficient propensity score based estimators for average treatment effect (ATE) with many covariates having multicollinearity, it’s difficult to  formally automate this variable selection rule in one statistical framework. In order to improve efficiency and solve multicollinearity issue, a two-stage estimation procedure is proposed in this paper. In the first stage, we adjust the usual Horvitz-Thompson estimator of the ATE by incorporating IVs to avoid model misspecification and then propose the generalized regression estimator by utilizing the auxiliary information from covariates related to the potential outcomes. In the second stage, we adapt the Elastic-net method to solve the multicollinearity issue and further improve the estimation efficiency based on the selected important covariates. The finite-sample performance of the proposed estimator is studied through simulation, and an application to employees’ weekly wages is also presented.

个人简介：王磊，南开大学统计与数据科学学院副研究员，博士生导师。研究方向是复杂数据分析和统计学习，已在Biometrika、Bernoulli、Statistica Sinica等统计学杂志发表学术论文30多篇，主持国家自然科学基金面上项目、青年项目及天津市自然科学基金各一项。

题目：Combinatorial Causal Bandit

摘要：Combinatorial causal bandit (CCB) is the integration of causal inference, multi-armed bandit and combinatorial optimization techniques. CCB is the following online learning task: A learning agent is given a causal graph with unknown distributions on how each variable X is causally influenced by its parents. The learning task is carried out in T rounds. In each round, the agent selects at most K variables to intervene, the output of a target variable Y is its reward, and the output of all observed variables is the feedback to the agent. The agent aims at learn from the past feedback to help selecting intervention variables in subsequent rounds, so that in the end the cumulative reward of all rounds is maximized. The performance metric is (cumulative) regret, which is defined as the difference between the cumulative reward of always selecting the optimal set of intervention variables and selecting intervention variables according to the learning algorithm. We propose the CCB framework, and study CCB for generalized linear causal models and design learning algorithms with near optimal regret bounds.

个人简介：陈卫是微软亚洲研究院高级研究员，也在清华大学、上海交通大学、中科院等多所院校和研究机构担任客座教授或研究员。他是中国计算机学会理论计算机专业委员会的常务委员，也是全国大数据专家委员会的委员。他是国际电气和电子工程师学会的会士（IEEE Fellow）。陈卫主要的研究方向包括社交和信息网络，在线学习，网络博弈论和经济学，分布式计算，容错等。他在2013年与人合著一本英文专著，在2020年独立撰写一本中文专著。他在多个学术期刊担任编委，也在多个学术会议中担任过技术委员会主席和委员。陈卫于清华大学获得本科和硕士毕业，于美国康奈尔大学获得博士学位。有关陈卫更多的信息，欢迎访问他的主页：http://research.microsoft.com/en-us/people/weic/.

题目：Confederated learning and Inference

摘要: The theory of statistical learning and inference for large-scale/high-dimensional data analysis has recently attracted considerable interest. The central analytic task in the development of confederated statistical learning and inference pertains to the method of integrating results yielded from multiple/sequential data batches. This talk introduced an one-step meta method based on confidence inference functions, a communication efficient method without pooling individual datasets for unbalanced datasets, and an incremental learning algorithm for streaming datasets with correlated outcomes. Integrative causal inference of multiple similar clinical studies conducted at different sites are also investigated.

个人简介：周岭，2004-2010年四川大学数学学院本科和硕士，2014年西南财经大学博士，2018年美国密西根大学生物统计系博士后，2017年钟家庆数学奖获得者，周岭与合作者在数据集成、选择后推断、亚组分析、非参数理论与方法、因果推断等领域取得了一系列研究成果，在Journal of the American Statistical Association (JASA), Journal of Economics (JoE), Journal of Machine Learning Research(JMLR), Annal of Applied Statistics(AOAS), Biometrics等国际统计学、计量经济学、计算机领域期刊上发表论文20余篇。

题目：Nonparametric Estimation of Continuous Treatment Effect with Measurement Error

摘要：We consider estimating the average dose-response function (ADRF) nonparametrically for continuous-valued treatment. The existing literature of continuous treatment effect proposed consistent estimators only for error-free data. However, in observational studies concerned by the literature of treatment effect, the treatment data can be measured with error. There, existing techniques are not applicable and finding a proper modification is not straightforward. We identify the ADRF by a weighted conditional expectation and estimate the weights nonparametrically by maximising a local generalised empirical likelihood subject to an expanding set of conditional moment equations incorporated with the deconvolution kernels. We then construct a deconvolution kernel estimator of the weighted conditional expectation. We derive the $L_2$ and $L_\infty$ convergence rates of our weights estimator and the asymptotic bias and variance of our ADRF estimator. We also provide the asymptotic linear expansion of our ADRF estimator in both the ordinary smooth and the supersmooth error cases, which can help conduct statistical inference. We provide a data-driven method to select our smoothing parameters based on the simulation-extrapolation (SIMEX) idea and propose a new extrapolation procedure to stabilise the computation. Monte-Carlo simulations show a satisfactory finite-sample performance of our method, and a real data study illustrates its practical value.

个人简介：张政，中国人民大学统计与大数据研究院助理教授，2015年于香港中文大学统计系获博士学位。研究方向包括因果推断、缺失数据、污染数据、半参数模型的有效估计、非参数统计推断、随机微分方程、随机分析等。在JRSS-B, JOE, Quantitative Economics, JBES, Statistica Sinica, Stochastic Processes and their Applications等统计、计量经济、概率论国际期刊上发表论文十余篇。主持国家自然科学基金青年基金，北京市自然科学基金面上项目。

题目: A novel stable higher-order influence function estimators for doubly-robust functionals

摘要: In this talk, we will first review the concept of higher-order influence functions (HOIFs) and HOIF-based estimators for a class of smooth statistical functionals commonly encountered in causality. We will demonstrate applications of HOIF-based estimators when deep-learning is being deployed in applied data analysis, including applied causal inference tasks, despite our theoretical understanding about deep learning being still quite limited. Motivated from some of our empirical experience, we recently developed a new class of HOIF-based estimators, which, somewhat surprisingly, enjoy both the nice theoretical properties of the original HOIF-based estimators proposed in 2008 by Robins et al. and possibly more importantly, the numerical stability in finite-sample settings. This new class of HOIF-based estimators (1) paves the way towards making HOIFs practically useful, (2) bridges the HOIF estimators developed in Robins et al. 2008 and the cross-fitting estimators in Newey and Robins 2018, and (3) coincides the HOIF estimator in the fixed-design setting. We envision that this new class of HOIF estimators will be useful in applied works, at least before the myth of deep learning is completely resolved. This talk is based on two working papers, one empirical paper with Kerollos N. Wanis and Jamie Robins and one theoretical paper with Chang Li (a senior undergraduate student at Shanghai Jiao Tong University).

个人简介：Lin Liu is an Assistant Professor in the Institute of Natural Sciences, School of Mathematical Sciences, and SJTU-Yale Joint Center for Biostatistics and Data Science at Shanghai Jiao Tong University, and PI in the Shanghai AI lab. He completed his PhD in biostatistics at Harvard University, under the supervision of Franziska Michor & Jamie Robins. His research interests lie in the intersection between mathematical statistics, causal inference, machine learning, and applied statistics in biomedical sciences.

题目: Estimation and inference for high-dimensional nonparametric additive instrumental-variables regression

摘要: The method of instrumental variables provides a fundamental and practical tool for causal inference in the presence of unmeasured confounding between the treatments and the outcome in various empirical studies. Modern data such as the genetical genomics data in these studies can be high-dimensional. The

high-dimensional linear instrumental-variables model has been considered in the literature due to its simplicity albeit the true relationship may be nonlinear. We propose a more data-driven approach by considering nonparametric additive models between the instrumental variables and the treatments while keeping the linear model assumption between the treatments and the outcome so that the coefficients therein can directly bear causal interpretation. We provide a two-stage framework for estimation and inference under this more general setup. The group lasso regularization is first employed to select optimal instruments for the high-dimensional nonparametric additive model, and then the outcome variable is regressed on the fitted values from the nonparametric additive model to identify and estimate important treatment effects. We provide non-asymptotic analysis for the estimation error of the proposed estimator. A debiased procedure is further employed to establish valid inference. Extensive numerical experiments show that our proposed method can rival or outperform existing approaches in the literature. We finally analyze the mouse obesity data with the proposed method and discuss new discoveries.

个人简介：中国人民大学统计学院，生物统计与流行病学系讲师，北京大学数学科学学院博士。主要研究领域为因果推断、缺失数据、高维统计等。目前已在包括Biometrika, Journal of Econometrics, Biometrics等国际著名统计期刊上发表多篇学术论文。主持一项国家自然科学青年基金项目，参与完成多项国家自然科学基金面上项目。

题目：Improving out-of-Distribution Performance of Machine Learning Models from a Causal Perspective

摘要：Given the remarkable performance of modern machine learning models on various benchmarking datasets, people turn to the next challenges in their wider applications. Among these, out-of-distribution (OOD) generalization is a critical one, since in many real-world tasks, the deploying environment is different from the training one, causing a change in data distribution. Causality provides an insightful approach to analyze and handle the problem. It proposes the model should learn causal relations which represents the fundamental rule governing the data in all environments, in contrast to superficial relations that may only appear in a specific environment accidentally. In this talk, we introduce a model and its variants for prediction/classification tasks which is designed following a causal reasoning process. We show their causal identification guarantees and OOD generalization analysis, and also the improved empirical performance.

个人简介：刘畅，现为微软亚洲研究院机器学习组主管研究员，2019年于清华大学计算机系取得博士学位。主要研究方向包括贝叶斯推断方法，因果模型，生成式模型及其与物理学问题的结合。