Many are convinced that the field of Human Computer Interaction (HCI or human-computer interaction) is reduced only to the design of sites or applications, and the main task of a specialist is to satisfy users, increasing the like button by a few pixels. In the post we want to show that this is not at all, and to tell what is happening in HCI at the junction with the studies of machine learning and artificial intelligence. Perhaps this will allow readers to look at this area from a new perspective.

For review, we took the proceedings of the conference CHI: Conference on Human Factors in Computing Systems for 10 years, and using NLP and network analysis societies looked at topics and areas at the intersection of disciplines.

ITKarma picture

The focus on applied UX design problems is especially strong in Russia. Many of the events that helped the growth of HCI abroad did not occur: we didn’t have iSchools , many specialists involved in science left related aspects of engineering psychology, etc. As a result, the profession arose again, starting from applied tasks and research. One of the results of this is visible even now - this is an extremely low representation of Russian HCI papers at key conferences.

But outside of Russia, HCI has evolved very differently, focusing on many topics and directions. At the master's program " Information systems and human-computer interaction " in St. Petersburg HSE, we, among other things, are discussing with students, colleagues, graduates of similar specialties of European universities, partners who help to develop the program - what applies to the field of human-computer interaction. And these discussions show the heterogeneity of the direction in which each specialist has his own, incomplete picture of the area.

From time to time, we hear questions about how this area is related (and whether at all) to machine learning and data analysis. To answer them, we turned to recent studies presented at the CHI conference.

First of all, we will tell you what happens in areas such as xAI and iML (eXplainable Artificial Intelligence and Interpretable Machine Learning) on the part of interfaces and users, as well as how HCI studies the cognitive aspects of the work of data science specialists, and give examples of interesting works of recent years in each area.

xAI and iML

Machine learning methods are being developed intensively and - more importantly from the point of view of the area under discussion - are actively being introduced into automated decision making. Therefore, researchers are increasingly discussing questions: how do non-machine learning users interact with systems where similar algorithms are used? One of the important issues of this interaction: how to make users trust model-based decisions? Therefore, every year the topic of interpreted machine learning (Interpretable Machine Learning - iML) and explainable artificial intelligence (eXplainable Artificial Intelligence - XAI) becomes more and more hot.

At the same time, if at conferences such as NeurIPS, ICML, IJCAI, KDD, the algorithms and means of iML and XAI are discussed, several topics related to the design features and experience of using these systems will be in focus on CHI. For example, at CHI-2020 several sections were devoted to this subject at once, including "AI/ML & amp; seeing through the black box ”and“ Coping with AI: not agAIn! ”. But even before the appearance of individual sections, there were quite a lot of such works. We identified four areas in them.

Design of interpreting systems for solving applied problems

The first direction is the design of systems based on interpretability algorithms in various applied problems: medical, social, etc. Such works arise in very different fields.For example, working at CHI-2020 CheXplain: Enabling Physicians to Explore and Understand Data-Driven, AI-Enabled Medical Imaging Analysis describes a system that helps doctors examine and explain chest x-ray results. It offers additional textual and visual explanations, as well as pictures with the same and opposite result (supporting and contradictory examples). If the system predicts that a disease is visible on radiography, then it will show two examples. The first, supporting, example is a snapshot of the lungs of another patient who has the same disease. The second, contradictory, example is a picture in which there is no disease, that is, a picture of the lungs of a healthy person. The main idea is to reduce obvious errors and reduce the number of calls to third-party specialists in simple cases in order to make a diagnosis faster.

ITKarma picture

CheXpert: automated selection of areas + examples (unlikely vs definitely)

Development of systems for the study of machine learning models

The second direction is the development of systems that help interactively compare or combine several methods and algorithms. For example, in Silva: Interactively Assessing Machine Learning Fairness Using Causality on CHI -2020, a system was introduced that builds several machine learning models on user data and provides the possibility of their subsequent analysis. The analysis includes constructing a causal graph between the variables and calculating a series of metrics that evaluate not only the accuracy but also the fairness of the model (Statistical Parity Difference, Equal Opportunity Difference, Average Odds Difference, Disparate Impact, Theil Index), which helps to find distortions in predictions.

ITKarma picture

Silva : graph of relationships between variables + graphs for comparing honesty metrics + color highlighting of the influential variables in each group

General Model Interpretability Issues

The third area is a discussion of approaches to the problem of interpretability of models in general. Most often these are reviews, criticisms of approaches and open-ended questions: for example, what is meant by “interpretability”. Here I would like to note a review at CHI-2018 Trends and Trajectories for Explainable, Accountable and Intelligible Systems: An HCI Research Agenda , in which the authors examined 289 major works devoted to explanations in artificial intelligence, and 12,412 publications citing them. Using network analysis and thematic modeling, they identified four key areas of research: 1) Intelligent and Ambient (I & amp; A) Systems, 2) Explainable AI: Fair, Accountable, and Transparent (FAT) algorithms and Interpretable Machine Learning (iML), 3) Theories of Explanations: Causality & amp; Cognitive Psychology, 4) Interactivity and Learnability. In addition, the authors described the main research trends: interactive learning and interaction with the system.

User Studies

Finally, the fourth direction is user research of algorithms and systems that interpret machine learning models. In other words, this is a study of whether new systems become more understandable and transparent in practice, what difficulties users experience when working with interpretive rather than original models, how to determine whether the system is used as planned (or found a new application for it - maybe incorrect), what are the needs of users and do developers offer them what they really need.

There are a lot of tools and interpretation algorithms, so the question arises: how to understand which algorithm to choose? In Questioning the AI: Informing Design Practices for Explainable AI User Experiences how once the issues of motivation for using explanatory algorithms are discussed and problems are highlighted that, with all the variety of methods, have not yet been sufficiently resolved. The authors come to an unexpected conclusion: most of the existing methods are designed so that they answer the question “why” (“why I have such a result”), while users need to answer the question “why not” (“why not another "), and sometimes -" what to do so that the result changes. "

The work also says that users need to understand what are the limits of applicability of methods, what limitations they have - and this must be explicitly implemented in the proposed tools. This issue is illustrated more clearly in the article Interpreting Interpretability: Understanding Data Scientists' Use of Interpretability Tools for Machine Learning . The authors conducted a small experiment with specialists in the field of machine learning: they showed them the results of several popular tools for interpreting machine learning models and offered to answer questions related to making decisions based on these results. It turned out that even experts too trust such models and do not take results critically. Like any tool, explanatory models can be used incorrectly. When developing tools, it is important to take this into account when attracting accumulated knowledge (or specialists) in the field of human-computer interaction in order to take into account the features and needs of potential users.

Data Science, Notebooks, Visualization

Another interesting area of ​​HCI is the analysis of the cognitive aspects of working with data. Recently, a question has been raised in science about how the “degrees of freedom” of a researcher — features of data collection, design of experiments, and choice of analysis methods — affect research results and their reproducibility. Although the bulk of the discussion and criticism is related to psychology and the social sciences, many problems relate to the reliability of findings in the work of data analysts in general, as well as the difficulties in delivering these findings to consumers of analysis.

Therefore, the subject of this HCI area is the development of new methods for visualizing uncertainty in model predictions, creating systems for comparing analysis performed in different ways, and also analyzing the work of analysts with tools, for example, Jupyter notebooks.

Visualization of Uncertainty

Visualization of uncertainty is one of the features that distinguish scientific graphics from presentation and business visualization. For quite some time, the principle of minimalism and focus on the main trends was considered key in the latter. However, this leads to an overconfidence of users in a point estimate of the value or forecast, which can be critical, especially if we must compare forecasts with varying degrees of uncertainty. The work Uncertainty Displays Using Quantile Dotplots or CDFs Improve Transit Decision-Making analyzes how visualization of prediction uncertainty for scatter plots and cumulative distribution functions help users make more rational decisions using the example of the task of estimating the bus arrival time according to the mobile application. Which is especially nice, one of the authors supports the ggdist package for R with various options for visualizing uncertainty.

ITKarma picture

Examples of visualization of uncertainty ( )

However, the tasks of visualizing possible alternatives are often encountered, for example, for sequences of user actions in web analytics or application analytics.The work Visualizing Uncertainty and Alternatives in Event Sequence Predictions analyzes how graphically the alternatives are based on Time-Aware Recurrent Neural Network (TRNN) helps experts make decisions and trust them.

Comparison of models

No less important than the visualization of uncertainty, an aspect of the work of analysts - a comparison of how - often hidden - the researcher's choice of different approaches to modeling at all its stages can lead to different analysis results. In psychology and the social sciences, pre-registration of research design and a clear separation of exploratory and confirmatory studies are gaining popularity. However, in tasks where the study is more based on data, an alternative may be tools to assess the hidden risks of analysis by comparing models. Increasing the Transparency of Research Papers with Explorable Multiverse Analyzes suggests using interactive visualization several approaches to analysis in articles. In essence, the article turns into an interactive application where the reader can evaluate what will change in the results and conclusions if a different approach is applied. This seems like a good idea for practical analytics as well.

Working with tools for organizing and analyzing data

The final block of work is about researching how analysts work with systems like Jupyter Notebooks, which have become a popular tool for organizing data analysis. The article Exploration and Explanation in Computational Notebooks analyzes the contradictions between research and explanatory goals, by looking at interactive documents found on Github, and Managing Messes in Computational Notebooks , authors analyze how notes evolve, parts of the code and visualization in the iterative process of analysts, and offer possible additions to the tools to support this process. Finally, already at CHI 2020, the main problems of analysts at all stages of work, from loading data to transferring the model to production, as well as ideas for improving tools are summarized in the article What’s Wrong with Computational Notebooks? Pain Points, Needs, and Design Opportunities .

ITKarma picture

Transformation of the report structure based on the execution logs ( )

To summarize

Concluding part of the discussion “What is HCI doing” and “why does a HCI specialist know machine learning”, I would like to once again note the general conclusion from the motivation and results of these studies. As soon as a person appears in the system, this immediately leads to a number of additional questions: how to simplify interaction with the system and avoid errors, how the user changes the system, does the actual use differ from the planned one. As a result, we need those who understand how the process of designing systems with artificial intelligence works and know how to take into account the human factor.

We teach all this at the master's program " Information systems and human-computer interaction ". If you are interested in research in the field of HCI, look at the light (now the admission campaign just started ). Or follow our blog: we’ll talk about projects that students worked on this year.