Rejoinder: “Co-citation and Co-authorship Networks of Statisticians”
Rejoinder: “Co-citation and Co-authorship Networks of Statisticians”
复制标题
DOI：
10.1080/07350015.2022.2055358
复制
发表时间：
2022-04
影响因子：
3
通讯作者：
Pengsheng Ji;Jiashun Jin;Z. Ke;Wanshan Li
中科院分区：
文献类型：
--
作者：
Pengsheng Ji;Jiashun Jin;Z. Ke;Wanshan Li
我们感谢大卫·多诺霍非常鼓舞人心的评论。一如既往，他富有洞察力的视野和深刻的思想极具刺激性。我们很高兴他用几句话概括了早年(例如弗朗西斯·高尔顿时代)的统计学与我们这个时代的统计学之间的重大哲学差异：数据优先与模型优先。我们完全同意他的观点，“统计研究人员为了解一种新的可用数据类型所做的每一次努力都会扩大我们的领域；培养新类型数据集的兴趣，以便能够发现和开发新的方法，这应该是统计学家职业生涯的主要部分”；这些正是我们(几年来)努力收集、清理和分析大规模高质量数据集的动机。我们想补充的是，这两种传统各有长处，结合双方的长处可能会极大地帮助统计学家应对我们今天面临的所谓21世纪的统计危机。让我们先解释一下上面的危机。在模型优先的传统中，考虑到特定的应用问题，我们提出了一个模型，开发了一个方法，并用一些难以证明的定理证明了它的最优性，并找到了一个数据集来支持该方法。在这个传统中，我们非常相信我们的模型和我们的理论：我们希望模型是足够的，我们希望我们的最优化理论保证我们的方法比其他方法更优越。现代机器学习文献(特别是深度学习的最新发展)提供了一种不同的方法来证明一种方法的优越性；我们通过十几个基准数据集的真实数据结果将所提出的方法与现有方法进行了比较。要为他们的数据集选择算法，从业者不一定需要定理的保证；比许多基准数据集更好的性能说明了一切。对于一些理论统计学家来说，这是相当令人失望的，因为他们来自于
We thank David Donoho for very encouraging comments. As always, his penetrating vision and deep thoughts are extremely stimulating. We are glad that he summarizes a major philosophical difference between statistics in earlier years (e.g., the time of Francis Galton) and statistics in our time by just a few words: data-first versus model-first. We completely agree with his comment that “each effort by a statistics researcher to understand a newly available type of data enlarges our field; it should be a primary part of the career of statisticians to cultivate an interest in cultivating new types of datasets, so that new methodology can be discovered and developed”; these are exactly the motivations underlying our (several-year) efforts in collecting, cleaning, and analyzing a large-scale high-quality dataset. We would like to add that both traditions have strengths, and combining the strengths of two sides may greatly help statisticians deal with the so-called crisis of the 21st century in statistics we face today. Let us explain the crisis above first. In the model-first tradition, with a particular application problem in mind, we propose a model, develop a method and justify its optimality by some hard-to-prove theorems, and find a dataset to support the approach. In this tradition, we put a lot of faith on our model and our theory: we hope the model is adequate, and we hope our optimality theory warrants the superiority of our method over others. Modern machine learning literature (especially the recent development of deep learning) provides a different approach to justifying the “superiority” of an approach; we compare the proposed approach with existing approaches by the real data results over a dozen of benchmark datasets. To choose an algorithm for their dataset, a practitioner does not necessarily need warranties from a theorem; a superior performance over many benchmark datasets says it all. To some theoretical statisticians, this is rather disappointing, as they come from a long