"An algorithm can be defined as a series of steps undertaken in order to solve a particular problem or accomplish a defined outcome. Algorithms can be carried out by people, by nature, or by machines. The way you learned to do long division in grade school or the recipe you followed last night to cook dinner are examples of people executing algorithms. You might also say that biologically governed algorithms describe how cells transcribe DNA to RNA and then produce proteins—it’s an information transformation process. While algorithms are everywhere around us, the focus of this paper are those algorithms that run on digital computers, since they have the most potential to scale and affect large swaths of people. Autonomous decision-making is the crux of algorithmic power. Algorithmic decisions can be based on rules about what should happen next in a process, given what’s already happened, or on calculations over massive amounts of data. The rules themselves can be articulated directly by programmers, or be dynamic and flexible based on the data. For instance, machine-learning algorithms enable other algorithms to make smarter decisions based on learned patterns in data. Sometimes, though, the outcomes are important (or messy and uncertain) enough that a human operator makes the final decision in a process. But even in this case the algorithm is biasing the operator, by directing his or her attention to a subset of information or recommended decision. Not all of these decisions are significant of course, but some of them certainly can be. We can start to assess algorithmic power by thinking about the atomic decisions that algorithms make, including prioritization, classification, association, and filtering.
Sometimes these decisions are chained in order to form higher-level decisions and information transformations. For instance, some set of objects might be classified and then subsequently ranked based on their classifications. Or, certain associations to an object could help classify it: Two eyes and a nose associated with a circular blob might help you determine the blob is actually a face. Another composite decision is summarization, which uses prioritization and then filtering operations to consolidate information while maintaining the interpretability of that information. Understanding the elemental decisions that algorithms make, including the compositions of those decisions, can help identify why a particular algorithm might warrant further investigation." (http://towcenter.org/research/algorithmic-accountability-on-the-investigation-of-black-boxes-2/)
"Prioritization, ranking, or ordering serves to emphasize or bring attention to certain things at the expense of others. The city of New York uses prioritization algorithms built atop reams of data to rank buildings for fire-code inspections, essentially optimizing for the limited time of inspectors and prioritizing the buildings most likely to have violations that need immediate remediation. Seventy percent of inspections now lead to eviction orders from unsafe dwellings, up from 13 percent without using the predictive algorithm—a clear improvement in helping inspectors focus on the most troubling cases. Prioritization algorithms can make all sorts of civil services more efficient. For instance, predictive policing, the use of algorithms and analytics to optimize police attention and intervention strategies, has been shown to be an effective crime deterrent. Several states are now using data and ranking algorithms to identify how much supervision a parolee requires. In Michigan, such techniques have been credited with lowering the recidivism rate by 10 percent since 2005. Another burgeoning application of data and algorithms ranks potential illegal immigrants so that higher risk individuals receive more scrutiny.10 Whether it’s deciding which neighborhood, parolee, or immigrant to prioritize, these algorithms are really about assigning risk and then orienting official attention aligned with that risk. When it comes to the question of justice though, we ought to ask: Is that risk being assigned fairly and with freedom from malice or discrimination? Embedded in every algorithm that seeks to prioritize are criteria, or metrics, which are computed and used to define the ranking through a sorting procedure.
These criteria essentially embed a set of choices and value-propositions that determine what gets pushed to the top of the ranking. Unfortunately, sometimes these criteria are not public, making it difficult to understand the weight of different factors contributing to the ranking. For instance, since 2007 the New York City Department of Education has used what’s known as the value-added model (VAM) to rank about 15 percent of the teachers in the city. The model’s intent is to control for individual students’ previous performance or special education status and compute a score indicating a teacher’s contribution to students’ learning. When media organizations eventually obtained the rankings and scores through a Freedom of Information Law (FOIL) request, the teacher’s union argued that, “the reports are deeply flawed, subjective measurements that were intended to be confidential.” Analysis of the public data revealed that there was only a correlation of 24 percent between any given teacher’s scores across different pupils or classes.
This suggests the output scores are very noisy and don’t precisely isolate the contribution of the teacher. What’s problematic in understanding why that’s the case is the lack of accessibility to the criteria that contributed to the fraught teacher rankings. What if the value-proposition of a certain criterion’s use or weighting is political or otherwise biased, intentionally or not?
Classification decisions involve categorizing a particular entity as a constituent of a given class by looking at any number of that entity’s features. Classifications can be built off of a prioritization step by setting a threshold (e.g., anyone with a GPA above X is classified as being on the honor roll), or through more sophisticated computing procedures involving machine learning or clustering. Google’s Content ID is a good example of an algorithm that makes consequential classification decisions that feed into filtering decisions12. Content ID is an algorithm that automatically scans all videos uploaded to YouTube, identifying and classifying them according to whether or not they have a bit of copyrighted music playing during the video. If the algorithm classifies your video as an infringer it can automatically remove (i.e., filter) that video from the site, or it can initiate a dialogue with the content owner of that music to see if they want to enforce a copyright. Forget the idea of fair use, or a lawyer considering some nuanced and context-sensitive definition of infringement, the algorithm makes a cut-and-dry classification decision for you. Classification algorithms can have biases and make mistakes though; there can be uncertainty in the algorithm’s decision to classify one way or another. Depending on how the classification algorithm is implemented there may be different sources of error. For example, in a supervised machine-learning algorithm, training data is used to teach the algorithm how to place a dividing line to separate classes. Falling on either side of that dividing line determines to which class an entity belongs. That training data is often gathered from people who manually inspect thousands of examples and tag each instance according to its category.
The algorithm learns how to classify based on the definitions and criteria humans used to produce the training data, potentially introducing human bias into the classifier.
In general, there are two kinds of mistakes a classification algorithm can make—often referred to as false positives and false negatives. Suppose Google is trying to classify a video into one of two categories: “infringing” or “fair use.” A false positive is a video classified as “infringing” when it is actually “fair use.” A false negative, on the other hand, is a video classified as “fair use” when it is in fact “infringing.” Classification algorithms can be tuned to make fewer of either of those mistakes. However, as false positives are tuned down, false negatives will often increase, and vice versa. Tuned all the way toward false positives, the algorithm will mark a lot of fair use videos as infringing; tuned the other way it will miss a lot of infringing videos altogether. You get the sense that tuning one way or the other can privilege different stakeholders in a decision, implying an essential value judgment by the designer of such an algorithm14. The consequences or risks may vary for different stakeholders depending on the choice of how to balance false positive and false negative errors. To understand the power of classification algorithms we need to ask: Are there errors that may be acceptable to the algorithm creator, but do a disservice to the public? And if so, why was the algorithm tuned that way?
Association decisions are about marking relationships between entities. A hyperlink is a very visible form of association between webpages. Algorithms exist to automatically create hyperlinks between pages that share some relationship on Wikipedia for instance. A related algorithmic decision involves grouping entities into clusters, in a sort of association en masse. Associations can also be prioritized, leading to a composite decision known as relevance. A search engine prioritizes the association of a set of webpages in response to a query that a user enters, outputting a ranked list of relevant pages to view. Association decisions draw their power through both semantics and connotative ability. Suppose you’re doing an investigation of doctors known to submit fraudulent insurance claims. Several doctors in your dataset have associations to known fraudsters (e.g., perhaps they worked together at some point in the past). This might suggest further scrutinizing those associated doctors, even if there’s no additional evidence to suggest they have actually done something wrong. IBM sells a product called InfoSphere Identity Insight, which is used by various governmental social service management agencies to reduce fraud and help make decisions about resource allocation. The system is particularly good at entity analytics, building up context around people (entities) and then figuring out how they’re associated. One of the IBM white papers for the product points out a use case that highlights the power of associative algorithms.15The scenario depicted is one in which a potential foster parent, Johnson Smith, is being evaluated. InfoSphere is able to associate him, through a shared address and phone number, with his brother, a convicted felon. The paper then renders judgment: “Based on this investigation, approving Johnson Smith as a foster parent is not recommended.” In this scenario the social worker would deny a person the chance to be a foster parent because he or she has a felon in the family. Is that right? In this case because the algorithm made the decision to associate the two entities, that association suggested a particular decision for the social worker. Association algorithms are also built on criteria that define the association. An important metric that gets fed into many of these algorithms is a similarity function, which defines how precisely two things match according to the given association. When the similarity reaches a particular threshold value, the two things are said to have that association. Because of their relation to classification then, association decisions can also suffer the same kinds of false positive and false negative mistakes.
The last algorithmic decision I’ll consider here is filtering, which involves including or excluding information according to various rules or criteria. Indeed, inputs to filtering algorithms often take prioritizing, classification, or association decisions into account. In news personalization apps like Zite or Flipboard news is filtered in and out according to how that news has been categorized, associated to the person’s interests, and prioritized for that person. Filtering decisions exert their power by either over-emphasizing or censoring certain information. The thesis of Eli Pariser’s The Filter Bubble16is largely predicated on the idea that by only exposing people to information that they already agree with (by overemphasizing it), it amplifies biases and hampers people’s development of diverse and healthy perspectives. Furthermore, there’s the issue of censorship. Weibo, the Chinese equivalent to Twitter, uses computer systems that constantly scan, read, and censor any objectionable content before it’s published. If the algorithm isn’t sure, a human censor is notified to take a look." (http://towcenter.org/research/algorithmic-accountability-on-the-investigation-of-black-boxes-2/)