15. Machine Learning

In this chapter we explore machine learning. This topic is closely related to statistical modeling, which we considered in Chapter 14, in the sense that both deal with using data to describe and predict outcomes of uncertain or unknown processes. However, while statistical modeling emphasizes the model used in the analysis, machine learning sidesteps the model part and focuses on algorithms that can be trained to predict the outcome of new observations. In other words, the approach taken in statistical modeling emphasizes understanding how the data is generated, by devising models and tuning their parameters by fitting to the data. If the model is found to fit the data well and if it satisfies the relevant model assumptions, then the model gives an overall description of the process, and it can be used to compute statistics with known distributions and for evaluating statistical tests. However, if the actual data is too complex to be explained using available statistical models, this approach has reached its limits. In machine learning, on the other hand, the actual process that generates the data, and potential models thereof, is not central. Instead, the observed data and the explanatory variables are the fundamental starting point of a machine-learning application. Given data, machine-learning methods can be used to find patterns and structure in the data, which can be used to predict the outcome of new observations. Machine learning therefore does not provide an understanding of how data was generated, and because fewer assumptions are made regarding the distribution and statistical properties of the data, we typically cannot compute statistics and perform statistical tests regarding the significance of certain observations. Instead, machine learning puts a strong emphasis on the accuracy with which new observations are predicted.