67 Evaluating Interactions
Managing the quality of an interaction with respect to its intent or goal is a crucial part of every step from design through implementation and especially during operation. Evaluating the quality of interactions at different times in the design process (design concept, prototype, implementation, and operation) reveals both strengths and weaknesses to the designers or operators of the organizing system.
During the design and implementation stages, interactions need to be tested against the original goals of the interaction and the constraints that are imposed by the organizing system, its resources and external conditions. It is very common for processes in interactions to be tweaked or tuned to better comply with the original goals and intentions for the interaction. Evaluation during these stages often attempts to provide a calculable way to measure this compliance and supports the fine-tuning process. It should be an integral part of an iterative design process.
During the later implementation and operation stages, interactions are evaluated with respect to the dynamically changing conditions of the organizing system and its environment. User expectations as well as environmental conditions or constraints can change and need to be checked periodically. A systematic evaluation of interactions ensures that changes that affect an interaction are observed early and can be integrated in order to adjust and even improve the interaction. At these stages, more subjective evaluation aspects like satisfaction, experience, reputation, or “feel” also play a role in fine-tuning the interactions. This subjective part of the evaluation process is as important as the quantitative, objective part. Many factors during the design and implementation processes need to be considered and made to work together. Ongoing quality evaluation and feedback ensures that interactions work as intended.
Evaluation aspects can be distinguished in numerous ways: by the effort and time to perform them (both data collection and analysis); by how quantifiable they are or how comparable they are with measures in other organization systems; by what component of the interaction or organizing system they focus on; or by the discipline, expertise, or methodologies that are used for the evaluation.
A common and important distinction is the difference between efficiency, effectiveness, and satisfaction. An interaction is efficient when it performs its actions in a timely and economical manner, effective when it performs its actions correctly and completely, satisfactory when it performs as expected. Satisfaction is the least quantifiable of the evaluation aspects because it is highly dependent on individual tastes and experiences.
Let us assume that Shopstyle.com develops a new interaction that lets you compare coat lengths from the offerings of their various retailers. Once the interaction is designed, an evaluation takes place in order to determine whether all coats and their lengths are integrated in the interaction and whether the coat lengths are measured and compared correctly. The designers would not only want to know whether the coat lengths are represented correctly but also whether the interaction performs efficiently. When the interaction is ready to be released (usually first in beta or test status), users and retailers will be asked whether the interaction improves their shopping experience, whether the comparison performs as they expected, and what they would change. These evaluation styles work hand in hand in order to improve the interaction.
When evaluating the efficiency of an organizing system, we focus on the time, energy and economic resources needed in order to achieve the interaction goals of the system. Commonly, the fewer resources are needed for achieving a successful interaction, the more efficient the interaction.
Efficiency measures are usually related to engineering aspects such as the time to perform an action, number of steps to perform an interaction, or amount of computing resources used. Efficiency with respect to the human costs of memory load, attention, and cognitive processing is also important if there is to be a seamless user experience where users can interact with the system in a timely manner.
For a lot of organizing system interactions, however, effectiveness is the more important aspect, particularly for those interactions that we have looked at so far. If search results are not correct, then users will not be satisfied by even the most usable interface. Many interactions are evaluated with respect to their ability to return relevant resources. Why and how this is evaluated is the focus of the remainder of this section.
Effectiveness evaluates the correct output or results of an interaction. An effective interaction achieves relevant, intended or expected results. The concept of relevance and its relationship to effectiveness is pivotal in information retrieval and machine learning interactions. (“Relevance”) Effectiveness measures are often developed in the fields that developed the algorithm for the interaction, information retrieval, or machine learning. Precision and recall are the fundamental measures of relevance or effectiveness in information retrieval or machine learning interactions. (“The Recall / Precision Tradeoff”)
Relevance is widely regarded as the fundamental concept of information retrieval, and by extension, all of information science. Despite being one of the more intuitive concepts in human communication, relevance is notoriously difficult to define and has been the subject of much debate over the past century.
Historically, relevance has been addressed in logic and philosophy since the notion of inference was codified (to infer B from A, A must be relevant to B). Other fields have attempted to deal with relevance as well: sociology, linguistics, and psychology in particular. The subject knowledge view, subject literature view, logical view, system’s view, destination’s view, pertinence view, pragmatic view and the utility-theoretic interpretation are different perspectives on the question of when something is relevant.
In 1997, Mizzaro surveyed 160 research articles on the topic of relevance and arrived at this definition: “relevance can be seen as a point in a four-dimensional space, the values of each of the four dimensions being: (i) Surrogate, document, information; (ii) query, request, information need, problem; (iii) topic, context, and each combination of them; and (iv) the various time instants from the arising of problem until its solution.”
This rather abstract definition points to the terminological ambiguity surrounding the concept.
For the purpose of organizing systems, relevance is a concept for evaluating effectiveness that describes whether a stated or implicit information need is satisfied in a particular user context and at a particular time. One of the challenges for the evaluation of relevance in organizing systems is the gap between a user’s information need (often not directly stated), and an expression of that information need (a query). This gap might result in ambiguous results in the interaction. For example, suppose somebody speaks the word “Paris” (query) into a smart phone application seeking advice on how to travel to Paris, France. The response includes offers for the Paris Hotel in Las Vegas. Does the result satisfy the information need? What if the searcher receives advice on Paris but has already seen every one of the resources the organizing system offers? What is the correct decision on relevance here?
The key to calculating effectiveness is to be aware of what is being measured. If the information need as expressed in the query is measured, the topical relevance or topicality—a system-side perspective is analyzed. If the information need as in a person’s mind is measured, the pertinence, utility, or situational relevance—a subjective, personal perspective is analyzed. This juxtaposition is the point of much research and contention in the field of information retrieval, because topical relevance is objectively measurable, but subjective relevance is the real goal. In order to evaluate relevance in any interaction, an essential prerequisite is deciding which of these notions of relevance to consider.
The Recall / Precision Tradeoff
Precision measures the accuracy of a result set, that is, how many of the retrieved resources for a query are relevant. Recall measures the completeness of the result set, that is, how many of the relevant resources in a collection were retrieved. Let us assume that a collection contains 20 relevant resources for a query. A retrieval interaction retrieves 10 resources in a result set, 5 of the retrieved resources are relevant. The precision of this interaction is 50% (5 out of 10 retrieved resources are relevant); the recall is 25% (5 out of 20 relevant resources were retrieved).
It is in the nature of information retrieval interactions that recall and precision trade off with each other. To find all relevant resources in a collection, the interaction has to cast a wide net and will not be very precise. In order to be very precise and return only relevant resources to the searcher, an interaction has to be very discriminating and will probably not find all relevant resources. When a collection is very large and contains many relevant resources for any given query, the priority is usually to increase precision. However, when a collection is small or the information need also requires finding all relevant documents (e.g., in case law, patent searches, or medical diagnosis support), then the priority is put on increasing recall.
The completeness and granularity of the organizing principles in an organizing system have a large impact on the trade-off between recall and precision. (See Resources in Organizing Systems.) When resources are organized in fine-grained category systems and many different resource properties are described, high-precision searches are possible because a desired resource can be searched as precisely as the description or organization of the system allows. However, very specialized description and organization may preclude certain resources from being found; consequently, recall might be sacrificed. If the organization is superficial—like your sock drawer, for example—you can find all the socks you want (high recall) but you have to sort through a lot of socks to find the right pair (low precision). The trade-off between recall and precision is closely associated with the extent of the organization.
Satisfaction evaluates the opinion, experience or attitude of a user towards an interaction. Because satisfaction depends on individual user opinions, it is difficult to quantify. Satisfaction measures arise out of the user’s experience with the interaction—they are mostly aspects of user interfaces, usability, or subjective and aesthetic impressions.
Usability measures whether the interaction and the user interface designed for it correspond with the user’s expectations of how they should function. It particularly focuses on the usefulness of the interaction. Usability analyzes ease-of-use, learnability, and cognitive effort to measure how well users can use an interaction to achieve their task.
Although efficiency, effectiveness, and satisfaction are measured differently and affect different components of the interaction, they are equally important for the success of an interaction. Even if an interaction is fast, it is not very useful if it arrives at incorrect results. Even if an interaction works correctly, user satisfaction is not guaranteed. One of the challenges in designing interactions is that these factors invariably involve tradeoffs. A fast system cannot be as precise as one that prioritizes the use of contextual information. An effective interaction might require a lot of effort from the user, which does not make it very easy to use, so the user satisfaction might decrease. The priorities of the organizing system and its designers will determine which properties to optimize.
Let us continue our Shopstyle coat-length comparison interaction example. When the coat length calculation is performed in an acceptable amount of time and does not consume a lot of the organizing systems resources, the interaction is efficient. When all coat lengths are correctly measured and compared, the interaction is effective. When the interaction is seamlessly integrated into the shopping process, visually supported in the interface, and not cognitively exhausting, is it probably satisfactory for a user, as it provides a useful service (especially for someone with irregular body dimensions). What aspect should Shopstyle prioritize? It will probably weigh the consequences of effectiveness versus efficiency and satisfaction. For a retail- and consumer-oriented organizing system, satisfaction is probably one of the more important aspects, so it is highly likely that efficiency and effectiveness might be sacrificed (in moderation) in favor of satisfaction.
Recall and precision are only the foundation of measures that have been developed in information retrieval to evaluate the effectiveness of search algorithms. See (Baeza-Yates and Ribeiro 2011), (Manning et al. 2008) Ch. 8; (Demartini and Mizzaro 2006).