It is thought-provoking to read the pair of articles on 10 challenges in data science by Xuming He and Xihong Lin from a statistics perspective and Jeannette Wing from a computer science perspective. Unsurprisingly, there is a good overlap of important topics including multimodal and heterogenous data, data privacy, fairness and interpretability, and causal inference or reasoning. This overlap reflects and confirms the foundational and shared roles of statistics and computer science in data science, which is the merging of statistical and computing thinking in the context of solving domain problems. The challenges in both articles are presented as separate, not integrated, topics, and mostly decoupled from domain problems, possibly because of the mandate of “10 challenges.”
In my mind, the most exciting 10 challenges in data science are to solve 10 pressing real-world data problems with positive impacts. For example, how is data science going to help control covid-19 spread while allowing a healthy economy? To mitigate climate change so that its negative impact on human and economics can be minimized and in time? To bring precision medicine to every patient safely and timely? To unlock the mysteries of the unconscious brain? To design genomic therapies for Alzheimer’s? To design wearables that interact with multiple sclerosis patients to keep them safe? To help discover chip materials for the next generation of computers? To understand the origins of universe? To prevent cyberattacks on democracies all over the world? To self-regulate interactions of digital media with kids? To help people retool skills needed by the rapidly changing economy while allowing them to stay in familiar physical environments of friends, families, mountains, and rivers? Such real-world problems have to be the mission, the anchor, and the goal of data science, while methodologies/algorithms, approaches, and theories have to be at their service and appraised relative to how well they help solve them.
To solve any of these 10 real-world challenges and more, an integrated- and system-framing of data science needs to be embraced. Real-world data science problems are multidisciplinary, multidimensional, and multiphased. Each data science life cycle (DSLC) consists of domain problem formulation, data collection, data cleaning/preprocessing, visualization, analytical problem formulation/modeling, interpretation, evaluation/validation, data conclusions and decisions, and communication of decisions and conclusions. The steps are not at all linear but nonlinear and iterative. The challenges in He and Lin (this issue) fall mostly in the analytical problem-formulation or modeling stage and some on data preprocessing and one on issues in decision making. They do not touch other important steps such as data cleaning, problem formulation, and communication of decisions. Wing (this issue) covers emerging conceptual topics such as trustworthy AI and automating data preparation/preprocessing. Even though I believe some automation in the data cleaning step is necessary, I believe humans have to be in the loop to monitor, check, and make judgment calls in ambiguous situations flagged by machines. That is, I see a human–machine collaboration future, not automation, for “front-end stages of the data life cycle” (Wing 2020).
The challenges in both articles are important, yet incomplete, components of a data science life cycle or system. Unless the entire system or all the components are integrated and connected together and owned as the traditional topics, there is no insurance that real-world problems such as the 10 challenges above will be solved with positive impacts. In particular, neither article recognizes the many human judgment calls in DSLC or discusses the stability or robustness or reproducibility issues in, say, the choices of data leaning and algorithm in solving a data problem. Data cleaning/preprocessing and coding irreproducibility has led to grave consequences in the past. An article called “Growth in a Time of Debt” was published by economists Carmen Reinhart and Kenneth Rogoff (2010). They concluded that public debt is not good for growth. Such a conclusion was widely used as evidence to argue for austerity policies in Europe and the United States after the 2008 financial crisis. Four years later, Thomas Herndon, Michael Ash, and Robert Pollin (2014) invalidated this conclusion when they included the few data points from New Zealand and corrected the coding errors. (It is not clear why these data points were omitted in the first place.)
When we embrace the data science life cycle as a system, it is clear that the elephant in the room is the human judgment calls made in every step. That is, stability (or robustness) relative to reasonable or appropriate perturbations to the system, including human judgment calls on data-cleaning choices, data perturbation, and model choices, has to be among the core considerations and a key metric for success. This is to makes sure that these perturbations and judgment calls are not driving the data conclusions and decisions, unless justified with well-explained documents. Equally important is to ensure a reality check through prediction into the future (or its good surrogate). Stability is a fundamental and common-sense principle in knowledge seeking and decision making. In fact, when I asked philosopher colleague Branden Fitelson at Northeastern whether considerations of stability of belief/judgment go back to the Greeks, his answer was an affirmative yes and he pointed me to Plato’s quotes, here.
Read more: Read the full article here.