Classifications arrange resources to support discovery, selection, combination, integration, analysis, and other purposeful activity in every organizing system. A classification of diseases facilitates diagnosis and development of medical procedures, as well as accounting and billing. In addition, classifications facilitate understanding of a domain by highlighting the important resources and relationships in it, supporting the training of people who work in the domain and their acquisition of specialized skills for it.
We consider classification to be systematic when it follows principles that govern the structure of categories and their relationships. However, being systematic and principled does not necessarily ensure that a classification will be unbiased or satisfy all users’ requirements. For example, the zoning, environmental, economic development, and political district classifications that overlay different parts of a city determine the present and future allocation of services and resources, and over time influence whether the city thrives or decays. These classifications reflect tradeoffs and negotiations among numerous participants, including businesses, lobbyists, incumbent politicians, donors to political parties, real estate developers, and others with strong self-interests.
Classification Is Purposeful
Categories often arise naturally, but by definition, classifications do not because they are systems of categories that have been intentionally designed for some purpose. Every classification brings together resources that go together, and in doing so differentiates among them. However, bringing resources together would be pointless without reasons for finding, accessing, and interacting with them later.
Classifications Are Reference Models
A classification creates a semantic or conceptual roadmap to a domain by highlighting the properties and relationships that distinguish the resources in it. This reference model facilitates learning, comprehension, and the use of organizing systems within the domain. Standard classifications like those used in libraries enable people to rely on one system that they can use to locate resources in many libraries. Standard business, job, and product classifications enable the reliable collection, analysis, and interchange of economic data and resources.
Another important use of standard classifications created by people is as a “gold standard” for comparison with unsupervised computational classifications carried out on the same collection of resources or in the same domain. Presumably, no unsupervised classifier could exactly reproduce the classifications created by careful experts.
Classifications Support Interactions
A classification creates structure in the organizing system that increases the variety and capability of the interactions it can support. With physical resources, classification increases useful co-location; in kitchens, for example, keeping resources that are used together near each other (e.g., baking ingredients) makes cooking and cleanup more efficient (see “activity-based” classification in “Classification by Activity Structure”).
Classification makes systems more usable when it is manifested in the arrangement of resource descriptions or controls in user interface components like list boxes, tabs, buttons, function menus, and structured lists of search results.
A typical mapping between the logic of a classification scheme and a user interface is illustrated in Figure: Classification and Interactions.
How a business classifies its product or service strongly influences whether a customer can find it; this is the essential task of marketing. The business of “search engine optimization” exists to help a firm with a web presence choose the categories and descriptive terms that will improve its ranking in search results and attract the number of types of customer it desires.
How a customer interacts with a supplier is influenced by how the supplier classifies its offerings in its shopping aisles or catalogs; the “science of shopping” uses creative classifications and co-location of goods to shape browsing behavior and encourage impulse buying.
In business-to-business contexts, standard classifications for business processes and their application interfaces enable firms to more easily build and maintain supply chains and distribution networks that interconnect many business partners.
Classification Is Principled
“Principles for Creating Categories” explained principles for creating categories, including enumeration, single properties, multiple properties and hierarchy, probabilistic co-occurrence of properties, theory and goal-based categorization. It logically follows that the principles considered in designing categories are embodied in classifications that use those categories. However, when we say, “classification is principled,” we are going further to say that the processes of assigning resources to categories and maintaining the classification scheme over time must also follow principles.
The design and use of a classification system involves many choices about its purposes, scope, scale, intended lifetime, extensibility, and other considerations. Principled classification means that once those design choices are made they should be systematically and consistently followed.
Principled does not necessarily equate to “good,” because many of the choices can be arbitrary and others may involve tradeoffs that depend on the nature of the resources, the purposes of the classification, the amount of effort available, the complexity of the domain, and the capabilities of the people doing the classification and of the people using it (see “Category Design Issues and Implications”). Every classification system is biased in one way or another (see “Bibliographic Classification”).
Consider the classifications of resources in a highly-organized kitchen. (See “Organizing a Kitchen”). Tableware, dishes, pots and pans, spices and food provisions, and other resources have dedicated locations determined by a set of intersecting requirements and organizing principles. There is no written specification, and other people organize their kitchens differently.
On the other hand, complex institutional classification systems like those used in libraries or government agencies are implemented with detailed specifications, methods, protocols, and guidelines. The people who apply these methods in the field have studied the protocols in school or they have received extensive on-the-job training to ensure that they apply them correctly, consistently, and in accordance with the specifications and guidelines.
Principles Embodied in the Classification Scheme
Some of the most important principles that lead us to say that classification is principled are those that guide the design of the classification scheme in the first place. These principles are fundamental in the discipline of library science but they apply more broadly to other domains.
The warrant principle concerns the justification for the choice of categories and the names given to them. The principle of literary warrant holds that a classification must be based only on the specific resources that are being classified. In the library context, this ad hoc principle that builds a classification from a particular collection principle is often posed in opposition to a more philosophical or epistemological perspective, first articulated by Francis Bacon in the seventeenth century, that a classification should be universal and must handle all knowledge and all possible resources.
The principle of scientific warrant argues that only the categories recognized by the scientists or experts in a domain should be used in a classification system, and it is often opposed by the principle of use or user warrant, which chooses categories and descriptive terms according to their frequency of use by everyone, not just experts.
With classifications of physical resources like those in a kitchen, we see object warrant, where similar objects are put together, but more frequently the justifying principle will be one of use warrant, where resources are organized based on how they are used.
A second principle embodied in a classification scheme concerns the breadth and depth of the category hierarchy. We discussed this in “Category Design Issues and Implications” but in the context of classification this principle has additional implications and is framed as the extent to which the scheme is enumerative (“Classification vs. Physical Arrangement”). The decision to classify broadly or precisely depends largely on the variety or heterogeneity of the resources that the system of categories has been designed to organize. Because of the diversity of resources for a sale in a department store, a broad classification is necessary to accommodate everything in the store. Kitchen goods will be grouped together in a few aisles on a single floor. But a specialty kitchen store or a wholesale kitchen supply store for restaurants would classify much more precisely because of the restricted resource domain and the greater expertise of those who want to buy things there. An entire section might be dedicated just to knives, organized by knife type, manufacturer, quality of steel, and other categories that are not used in the kitchen section of the department store.
The precision or enumerativeness of a classification scheme increases the similarity of resources that are assigned to the same category and sharpens the distinctions between resources in different categories. However, when different classifications must be combined, mismatches in their precision or granularity can create challenges (see “Reorganizing Resources for Interactions”).
Principles for Assigning Resources to Categories
The uniqueness principle means the categories in a classification scheme are mutually exclusive. Thus, when a logical concept is assigned to a particular category, it cannot simultaneously be assigned to another category. Resources, however, can be assigned to several categories if they embody several concepts represented by those different categories. This can present a challenge when a physical storage solution is based on storing resources according to its assigned category in a logical classification system. This is not a serious problem for resource types like technical equipment or tools, for which the properties used to classify them are highly salient, and that have very narrow and predictable contexts of use. It is also not a problem for highly-specialized information resources like scientific research reports or government economic data, which might end up in only one specialized class. However, many resources are inherently more difficult to classify because they have less salient properties or because they have many more possible uses.
We face this kind of problem all the time. For example, should we store a pair of scissors in the kitchen or in the office? One solution is to buy a second pair of scissors so that scissors can be kept in both locations where they are typically used, but this is not practical for many types of resources and this principle would be difficult to apply in a systematic manner.
Many books are about multiple subjects. A self-help book about coping with change in a business setting might reasonably be classified as either about applied psychology or about business. It is not helpful that book titles are often poor clues to their content; Who Moved My Cheese? is in fact a self-help book about coping with change in a business setting. Its Library of Congress Classification is BF 637, “Applied Psychology,” and at UC Berkeley it is kept in the business school library.
The general solution to satisfying the uniqueness principle in library classifications when resources do not clearly fit in a single category is to invent and follow a detailed set of often arbitrary rules. Usually, the primary subject of the book is used for assigning a category, which will then determine the book’s place on a shelf.
However, another rule might state that if a book treats two subjects equally, the subject that is covered first determines the classification. For some classifications a “table of preference” can trump other rules at the last minute. Not surprisingly, the rules for categorizing books take a long time to learn and are not always easy to apply.
Principles for Maintaining the Classification over Time
Most personal classifications are created in response to a specific situation to solve an emerging organizational challenge. As a consequence, personal classification systems change in an ad hoc or opportunistic manner during their limited lifetimes. For example, the classification schemes in your kitchen or closet are deconstructed and disappear when you move and take your possessions to a different house or apartment. Your efforts to re-implement the classifications will be influenced by the configuration of shelves and cabinets in your new residence, so they will not be exactly the same.
In contrast, the institutional classification schemes for many library resources, culturally or scientifically-important artifacts, and much of the information created or collected by businesses, governments and researchers might have useful lives of decades or centuries. Classification systems like these can only be changed incrementally to avoid disruption of the work flows of the organization. We described maintaining resources as an activity in all organizing systems (“Maintaining Resources”) and the issues of persistence, effectivity, authenticity, and provenance that emerge with resources over time (“Resources over Time”). Much of this previous discussion applies in a straightforward manner to maintaining classifications over time.
However, some additional issues arise with classifications over time. The warrant principle (“Principles Embodied in the Classification Scheme”) implicitly treats the justification for designing and naming categories as a one-time decision. This is reasonable if you are organizing a collection of bibliographic resources or common types of physical resources like printed books, clothing or butterflies. However, in domains where the resources are active, change their state or implementation, or otherwise have a probabilistic character it might be necessary to revisit warrant and the decisions based on it from time to time. Put another way, if the world that you are sampling from or describing has some randomness or change in it, the categories and descriptions you imposed on it probably need to change as well. It often happens that the meaning of an underlying category can change, along with its relative and absolute importance with respect to the other categories in the classification system. Categories sometimes change slowly, but they can also change quickly and radically as a result of technological, process, or geopolitical innovation or events. Entirely new types of resources and bodies of knowledge can appear in a short time. Consider what the categories of “travel,” “entertainment,” “computing,” and “communication” mean today compared to just a decade or two ago.
Changes in the meaning of the categories in a classification threaten its integrity, the principle that categories should not move within the structure of the classification system.
One way to maintain integrity while adapting to the dynamic and changing nature of knowledge is to define a new version of a classification system while allowing earlier ones to persist, which preserves resource assignments in the previous version of the classification system while allowing it to change in the new one. If we adopt a logical perspective on classification (“Classification vs. Tagging”) that dissociates the conceptual assignment of resources to categories from their physical arrangement, there is no reason why a resource cannot have contrasting category assignments in different versions of a classification.
However, the conventional library with collections of physical resources cannot easily abandon its requirement to use a classification to arrange books on shelves in specific places so they can be located, checked out, and returned to the same location.
This constraint does not preclude the versioning of library classifications, but it increases the inertia and limits the degree of change when revisions are made because of the cost and coordination considerations of rearranging books in all the world’s libraries.
A related principle about maintaining classifications over time is flexibility, the degree to which the classification can accommodate new categories. Computer scientists typically describe this principle as extensibility, and library scientists sometimes describe it as hospitality. In any case the concern is the same and we are all familiar with it. When you buy a bookshelf, clothes wardrobe, file cabinet, or computer, it makes sense to buy one that has some extra space to accommodate the books, clothes, or files you will acquire over some future time frame. As with other choices that need to be made about organizing systems, how much extra space and “organizing room” you will acquire involves numerous tradeoffs.
Classification schemes can increase their flexibility by creating extra “logical space” when they are defined. Library classifications accomplish this by using naming or numbering schemes for classification that can be extended easily to create new subcategories.
Classification schemes in information systems can also anticipate the evolution of document or database schemas.
Classification Is Biased
The discipline of organizing is fundamentally about choices of properties and principles for describing and arranging resources. We discussed choices about describing resources in “The Process of Describing Resources”, choices for creating resource categories in “Principles for Creating Categories”, and choices for creating classifications in this chapter. The choices made reflect the purposes, experiences, professions, politics, values, and other characteristics and preferences of the people making them. As a result, every system of classification is biased because it takes a point of view that is a composite of all of these influences.
But first we need to point out that there are at least two quite different senses of “bias” that people reading this book are likely to encounter. The colloquial sense of bias we discuss in this section reflects value-based decisions in organizing systems that implicitly or explicitly favor some interactions or users over others. In contrast, statistical bias is systematic error or distortion in a measurement. (See the sidebar, Statistical Bias and Variance.)
The claim that classification is biased might seem surprising, because many classification systems are formal and institutional, created by governments or firms participating in standards organizations. We expect these classifications to be impartial and objective. However, consider the classification of people as “employed” or “unemployed.” Many people think that any employable person who is not currently employed would be counted as unemployed. But the US government’s Department of Labor only counts someone as unemployed if they have actively looked for work in the past month, effectively removing anyone who has given up on finding work from the unemployed category by assigning them to a “discouraged worker” category. In 2012 this classification scheme allowed the government to report that unemployment was about 8% and falling, when in fact it was closer to 20% and rising. The political implications of this classification are substantial.
Classification bias is often intentionally or unintentionally shown in data visualizations, including choropleth maps, in which map regions are colored, patterned, or otherwise distinguished according to a statistical variable being displayed on the map. Choropleths are commonly used to display election results, with the districts or states won by each candidate shown in different colors; in the United States, the convention is to show those won by Democratic Party candidates in blue, and those won by Republicans in Red. These election choropleths are often misleading because coloring an entire state in the winner’s colors ignores population density and the regional concentrations of votes that differ from the majority.
A more subtle way in which choropleths encode bias reflects the decisions made to organize the data into the categories that are represented by different colors or patterns. Choropleth categories might present data divided into equal range intervals, into sets with the same number of observations, or into categories that reflect clusters or natural breaks in the observed data. Small changes in the data ranges or proportions that are then assigned to each category can communicate entirely different stories with the same data. To learn “how to lie with maps” or how to prevent being lied to, refer to the classic book with that title by Mark Monmonier.
Friedman and Nissenbaum’s Bias In Computer Systems offers a framework for conceptualizing the various types of bias that may be present in technical systems. Friedman and Nissenbaum define bias as “a system that systematically and unfairly discriminates against individuals or groups of individuals in favor of others” Their taxonomy includes pre-existing, technical, and emergent bias.
Pre-existing bias is the type people are most familiar with: it occurs when an organizing system’s design embodies personal or societal biases that exist at the time of its creation, either intentionally or inadvertently, and sometimes despite one’s best intentions to prevent it.
Technical bias arises from limitations and constraints of technical systems that result in unfairness when the system is applied to the real world. Automated decision-making is especially ripe for this sort of bias: alphabetical ordering, processes that rely on pseudo-random number generation, and other automated ways of sorting or grouping resources may systematically create different opportunities for different user groups (e.g., people or companies whose names begin with “A”).
Emergent bias is related to the interplay between actual users and a technical system. Problems of this type arise when, due to the designer’s incomplete understanding of the user population, or a change in that population over time, there is a mismatch between users and the system. User interfaces are especially susceptible to this form of bias, given their need to reflect the habits and capacities of intended users. Unfairness can emerge when an unexpected user group uses the system, or as new societal knowledge arises that the system is not able to incorporate or respond to.
Both pre-existing and emergent bias may be difficult to assess accurately; the former may be difficult for the biased to see or admit to, and the latter, arising due to unanticipated circumstances after implementation, is hard to predict.
Bowker and Star have written extensively about biases in classification systems but acknowledge that many people do not see them:
Information scientists work every day on the design, delegation and choice of classification systems and standards, yet few see them as artifacts embodying moral and aesthetic choices that in turn craft people’s identities, aspirations and dignity.
Bowker and Star describe many examples where seemingly neutral and benign classifications implement controversial assumptions. A striking example is found in the ethnic classifications of the United States Census and the categories to which US residents are required to assign themselves. These categories have changed nearly every decade since the first census in 1790 and strongly reflect political goals, prevailing cultural sensitivities or lack thereof, and non-scientific considerations. Some recent changes included a “multi-racial” category, which some people viewed as empowering, but which was attacked by African-American and Hispanic civil rights groups as diluting their power.
A more positive way to think about bias in classification is that the choices made in an organizing system about resource selection, description, and arrangement come together to convey the values of the organizers. This makes a classification a rhetorical or communicative vehicle for establishing credibility and trust with those who interact with the resources in the classification. Seen in this light, an objective or neutral classification is not only unrealistic as a goal; it may also consume valuable time and energy when instead it might be more desirable to seize the opportunity to interpret the resources in a creative way to communicate a particular message to a particular user group. Melanie Feinberg makes the point that “fair trade” or “green” supermarkets differentiate themselves by a relatively small proportion of the goods they offer compared with ordinary stores, but these particular items signal the values that their customers care most about.
Bias is clearly evident in the most widely used bibliographic classifications, the Library of Congress and the Dewey Decimal, which we discuss next.
The application of classification and organizing principles more generally to the design of user interfaces to facilitate information access, navigation, and use is often called “Information Architecture.” See (Morville and Rosenfeld 2006).
The RosettaNet standards are used by thousands of firms as specifications and implementations of business-to-business processes in several industries, especially component manufacturing and electronics. The specifications are defined using a three-level hierarchy of process clusters, segments, and partner interface processes (PIPs) to enable firms to find a level of process abstraction that works best for them. See
For example, the introductory text for the Dewey Decimal Classification(DDC) is 38 pages long (
). A full set of online training modules “focused on the needs of experienced librarians needing Dewey application training” runs 30 hours (http://www.oclc.org/dewey/resources/teachingsite/courses/default.htm
(Taylor and Joudrey 2009, p. 392) define integrity as the stability of notations (class identifiers) in a classification so that resources are never given new notations when the category meaning changes. This is especially pertinent in a physical world where class notations are affixed to resources (books in a traditional library, for example) and where the changing of meaning would necessitate the changing of many numbers.
For example, the Universal Decimal Classification(UDC) intentionally left the main class 4 blank in order to have space for currently unknown subjects on the highest hierarchy level. (
). The Library of Congress Classification(LCC) also left space on the highest hierarchical level by not using all letters in the alphabet. Classifications also leave spaces in the enumeration of more specific classes.
(Rahm and Bernstein 2006) provide a crisp introduction to the challenges and approaches for changing deployed schemas in databases, conceptual models, ontologies, XML schemas, and software application interfaces. They operate an online bibliography on schema evolution that contains several hundred sources. See
See How the Government Measures Unemployment,
from the Department of Labor’s Bureau of Labor Statistics, and a critical commentary about the measurement scheme titled Making 9 Million Jobless Vanish: How the Government Manipulates Unemployment Statistics athttp://danielamerman.com/articles/2012/WorkC.html
(Monmonier 1996) is a highly-readable treatment of intentional and inadvertent bias in mapmaking. A web search for “lying with maps” yields a large number of examples. See also When Maps Lie” by Wiseman
See the Wikipedia entry Race and ethnicity in the United States census,
, and (Lee 1993) for arguments against any racial categorization because of the “political motivations and non-scientific character of the classifications.”