In data-intensive environments such as social media, machine learning algorithms have become a standard way of learning to recognize paterns in the data, to discover knowledge, and to predict the likelihood of user actions and tastes. Put another way, machine learning is largely enabled by proliferating data from which models may learn. In the age of so-called big data, having the biggest pool of data available from which to detect paterns is often seen as a competitive necessity. The bigger the database, so the story goes, the beter the conditions for algorithms to detect relevant paterns.
Models are central to the understanding of machine learning as “they are what is being learned from the data, in order to solve a given task” (Flach, 2012: 20). The way it works is that a pool of data is first mined (automatically processed) to identify regularities upon which subsequent decision-making can rely. The accumulated set of discovered relationships, then, produce the model, which subsequently “can be employed to automate the process of classifying entities or activities of interest, estimating the value of unobserved variables, or predicting future outcomes” (Barocas & Selbst, 2016: 7). As Joaquin Quiñonero Candela, Director of Applied Machine Learning at Facebook, says about the company’s machine learning philosophy:
- Get as much data as you can and make sure it is of highest quality 2. Distill your data into signals that will be maximally predictive—a process called feature-engineering 3. Once you have the most awesome data and tools for feature engineering, keep raising the capacity of your algorithms. (Candela, 2016)
Machine learning, then, is about using data to make models that have certain features. Feature engineering, or the process of extracting and selecting the most important features from the data, is arguably one of the most important aspects of machine learning. While feature extraction is usually performed manually, recent advances in deep learning now embed automatic feature engineering into the modeling process itself (Farias et al, 2016). If the algorithm operates on badly drawn features, the results will be poor, no mater how excellent the algorithm is.
How would feature engineering work in calculating such an abstract notion as relevance? One option might be to consider the frequency and types of content on which a specific user has clicked. However, we might also imagine a scenario in which “most important” is not about the frequency of clicks but, rather, the time spent watching or reading certain content. The point here is that the problem of determining what is “most important” depends on the data and the outcome you want to optimize. As Gillespie suggests, “All is in the service of the model’s understanding of the data and what it represents, and in service of the model’s goal and how it has been formalized” (2016a: 20). The understanding of data and what it represents, then, is not merely a mater of a machine that learns but also of humans who specify the states and outcomes in which they are interested in the first place.10 In the case of supervised learning, humans (the data miners and machine learning specialists) first need to specify the target variables (desired states and outcomes), which are then used to define “class labels”—“the diferent classes between which a model should be able to distinguish” (Barocas & Selbst, 2016: 8). Take the task of filtering spam, an area in which machine learning is commonly applied: The goal of the algorithm is to build an internal computer model that will ultimately allow the machine to make automated, accurate classification decisions (Surden, 2014: 91).11 However, learning tasks are far from alike. Classifying spam is one of the “easier” tasks as it typically merely operates on binary categories. Either something is classified as spam or it is not. But most of the tasks with which we are concerned in this book have to do with ranking and recommendations that cannot be broken down into binary categories. For example, the question of what constitutes newsworthiness or relevancy is not a straightforward task. As we shall see—particularly in chapters 4 and 6—there is no way to measure newsworthiness directly because the notion is a function of the particular way in which the respective industries and platforms have constructed their systems.12 Whether it is about ranking content in a feed or classifying email as spam, a useful model depends on the ongoing input of new data to optimize its performance.
Th e M u l tip l icity of A lgo rith ms
25