Discussion of Professor Bradley Efron’s Article on “Prediction, Estimation, and Attribution”
Discussion of Professor Bradley Efron’s Article on “Prediction, Estimation, and Attribution”
复制标题
Bradley Efron 教授关于“预测、估计和归因”的文章的讨论
DOI：
10.1111/insr.12415
复制
发表时间：
2020
影响因子：
2
通讯作者：
Zheng, Zheshi
中科院分区：
文献类型：
--
作者：
Xie, Min‐ge;Zheng, Zheshi
通过注意到“纯预测算法”的快速增长趋势，埃夫隆教授将 20 世纪的统计数据（估计和归因）与当前快速发展的 21 世纪的统计数据（预测）进行了比较和连接。精彩的讨论提供了许多根深蒂固的见解和评论。正如他关于费舍尔对现代统计学影响的前瞻性文章（Efron 1998）一样，该文章帮助塑造了统计推断的许多最新发展（包括我们自己在置信分布方面的工作（Singh、Xie 和 Strawderman 2005；Xie 和 Singh 2013）），Efron 教授的这篇同样鼓舞人心的文章必将激励现代统计学和数据科学基础的许多当代和强大的发展。在这篇文章中，我们呼应并提供了对 Efron 教授提出的两个重要观点的额外支持：（1）预测是“比归因或估计更容易的任务”；（2）IID 假设（例如训练和测试数据集的随机分割）在当前预测的发展中至关重要，但我们还需要针对不满足 IID 假设的情况做更多的工作。根据我们自己的研究，我们提供了额外的证据来支持这些讨论。我们发现预测具有稳态特性，并且即使使用的学习模型完全错误，在 IID 设置下也能很好地工作。我们还强调了良好的建模和推理实践的重要性：具有良好估计的良好学习模型对于提高 IID 情况下的预测效率非常重要，并且对于保持非 IID 情况下的有效性也至关重要。我们仍然需要努力在预测中建立良好的学习模型和估计算法，即使预测比估计更容易。从一开始，我们就想指出，考虑非独立同分布测试数据并不是一个稻草人的论点。相反，此类数据在数据科学中很普遍。除了 Efron 教授提供的那些显示“漂移”的示例之外，我们还可以轻松想象许多典型应用中的非 IID 示例。例如，预测算法是在患者病历数据库上进行训练的，我们希望预测对症状比普通患者更严重的新患者的潜在治疗结果。症状更严重的新患者并非典型患者
By noting the rapid growing trend of “pure prediction algorithms,” Professor Efron compares and bridges the statistics of the 20th Century (estimation and attribution) to that of the current fast growing development of the 21st Century (prediction). The outstanding discussion offers many deep-rooted insights and comments. As did his forward thinking article on Fisher’s influence on modern statistics (Efron 1998), which helped shape many recent developments on statistical inference (including our own work on confidence distribution (Singh, Xie, and Strawderman 2005; Xie and Singh 2013)), this equally inspiring article by Professor Efron will certainly galvanize many contemporary and powerful developments for modern statistics and for the foundations of data science. In this note, we echo and also provide additional support to two important points made by Professor Efron:(1) prediction is “an easier task than either attribution or estimation”;(2) the IID assumption (eg random splitting of training and testing datasets) is crucial in the current developments on predictions, but we also need to do more for the case when the IID assumption is not met. Based on our own research, we provide additional evidence to support these discussions. We discover that prediction has a homeostasis property and works well under the IID setting even if the learning model used is completely wrong. We also highlight the importance of having a good modeling and inference practice: a good learning model with good estimation is important to improve prediction efficiency in the IID case and it becomes essential to maintain validity in the non-IID case. The message remains: we still need to make effort to build a good learning model and estimation algorithm in prediction, even if prediction is an easier task than estimation.From the outset, we would like to point out that it is not a straw-man argument to consider non-IID testing data. On the contrary, such data are prevalent in data science. In addition to those examples provided by Professor Efron that showed “drift,” we can easily imagine non-IID examples in many typical applications. For instance, a predictive algorithm is trained on a database of patient medical records and we would like to predict potential outcomes of a treatment for a new patient with more severe symptoms than what the average patient shows. The new patient with more severe symptoms is not a typical
影响因子：
1.6
作者：
Barber, Rina Foygel;Candes, Emmanuel J.;Tibshirani, Ryan J.
通讯作者：
Tibshirani, Ryan J.
影响因子：
4.5
作者：
Barber, Rina Foygel;Candes, Emmanuel J.;Tibshirani, Ryan J.
通讯作者：
Tibshirani, Ryan J.