Data science ethics (concepts, techniques and cautionary tales), by David Martens, was published in 2022 by Oxford University Press. The book is inspired by the author’s course on Data Science and ethics he has been teaching at the University of Antwerp. (With a link to his slides.) The 255p book proceeds by decomposing the ethics of data science into its different steps: data gathering (Chap. 2), data preprocessing (Chap. 3), modelling (Chap. 4), evaluation (Chap. 5), and deployment (Chap. 6). Following the `FAT Flow Framework´, where FAT stands for fairness, accountability, and transparency.
Do not expect much maths, stats, or anything quantitative: this book is mostly about concepts, even though some (mostly well-known) illustrations are provided. Chapter 2 includes a description of encryption (with homomorphic encryption treated in Chapter 4). And somewhat improbably quantum computing. Differential privacy gets a few pages with not a single formula (until Chapter 4, again).
Chapter 3 covers k-anonymity, record linkage, reidentification, (through cautionary tales) and discrimination through biases in the (learning) dataset. Chapter 4 is defining ε differential privacy with the Laplace randomization as a possible implementation and with no critical stance on the limitations of the concept. The computation limitations of homomorphic encryption are more clearly pointed out. Federated learning is only quickly mentioned. The section about measuring fairness and reducing bias implies that some prior knowledge is available about whom is potentially discriminated and which covariates to add to the model. The last section on explicability of predictions is worthwhile in signalling the difficulty with most (black box) AI but the example opposing an SVM model to a logistic model is not tremendously convincing in that neither model is true.
Chapter 5 addresses the crucial challenge of ethical evaluation in a rather verbose and vague manner. For instance, with no instruction on how to resist adversarial attacks. Or criticising p-hacking and multiple testing while missing the elephant in the room (p-values!). Drifting from the topic when discussing the misdeeds of Diederik Stapel. Most of the same goes about Chapter 6 and its take on ethical deployment, when going through examples such as Google’s policies in China. Or general musing on the impact of AI on societal inequalities (with mentions of companies and CEOs who have since then back-pedalled on their ethical engagement). These chapters are lacking in tools and (more) practical recommendations.
One interesting aspect of the book is the attention paid to the EU(ropean) aspect of these concerns, through the GDPR (General DAta Protection Regulations) adopted by the European Parliament in 2016. (There is also a brief mention of China’s regulations, but no details beyond a reference. Maybe the Chinese edition differs.)
