Book Review: Big Data and Social Science
Overview
TL;DR: Broad introduction to data science techniques: data collection, machine learning, and validating conclusions. Read it with a concrete problem in mind and really learn from the references. Look at notebooks for code examples.
Big Data and Social Science: A Practical Guide to Methods and Tools is a book that seeks to give practical advice for applying data science techniques to real-world problems. The book provides a broad overview of a diverse set of techniques in data science, so the book is best consumed with a particular problem in mind.
Reactions
Overall, I felt that the book was a good overview of a large variety of techniques that come to bear on using data science to look at social issues. One of the best features is that they have a set of ipython notebooks located at notebooks. On the other hand, I felt that the book leaves you with little depth in understanding any particular topic (e.g. machine learning is covered in one chapter). I am not sure I have a practical suggestion as to how to remedy this issue, since the book seeks to cover a very broad topic. In retrospect, I think that reading this book would have been much more useful if I had a particular application in mind. That way I could have used this book as a reference for techniques that I could have looked at in more detail later.
More Detailed Summaries
The book contains three major sections, Data Capture and Curation, Modeling and Analysis, and Inference and Ethics.
- Capture and Curation
This section focuses on the different ways that one can obtain data and manage that data. The section starts out with a specific example of using web APIs to get data. The next section discusses combining your raw data and matching different records. This is a challenging and interesting problem because while you want to identify if two pieces of data, for instance, correspond to the same person. The section ends with a discussion of databases.
- Modeling and Analysis
This section starts with an overview of different machine learning methods, text analysis, and network analysis. I already know a lot about these methods, so I didn’t get too much from this part of the book.
- Inference and Ethics
This section contains the chapter that I found to be most interesting: Errors and Inference
Major issues:
Data generation process is often not created for the sake of generating valid conclusions about the data, and can introduce systematic bias, which renders future analysis useless.
Record linking can introduce either random or systematic error in the data.
When your number of features is large in your data, that can lead to spurious correlations and overfitting in your statistical models.
While big data has been said to be useful for studying rare populations, small errors in the data on these rare populations have large impacts on the analyses that you might want to use (e.g. a classifier).
There are a variety of ways to attempt to mitigate errors. The first is to create some rule that meshes with your experience, and to see if that trend is present in your data. This can help identify spurious errors in the data. There are also systematic methods based on various machine learning techniques. One particularly interesting method is called the tableplot. Basically, you take your data and form a histogram for one variable, and then look at the values of another variable in each of the chunks formed by your histogram. This can help identify inconsistencies in those aggregates.
This section also had chapters on information visualization and privacy and confidentiality. A key point is that behavior at the tails of a distribution is both the most interesting in many cases and can be most risky in terms of privacy. I think one major gap in this chapter is addressing the issue that researchers don’t act rationally for a variety of reasons (e.g. taking on implicit biases from culture, or being professionally invested in advancing a particular theory).
Conclusion
Data science seeks to address very challenging problems and as a result, the methods used in data science form a hodgepodge of techniques from a wide variety of different fields. A book on data science is going to reflect that. While to book gives a good introduction to data science, I feel that it suffers from the fact that it doesn’t offer a decisive account of any particular issues. Indeed, as data science as a field grows, this book will need to be updated.