![An edited volume, published by Cambridge University Press](header.png width="100%") # Bayesian Methods for Interaction and Design ## Editors * [John Williamson](http://johnhw.com) University of Glasgow * [Antti Oulasvirta](http://users.comnet.aalto.fi/oulasvir/) Aalto University * [Per Ola Kristensson](http://pokristensson.com/) Cambridge University * [Nikola Banovic](http://www.nikolabanovic.net) University of Michigan ## Summary This book will present for the first time a coherent foundation of Bayesian approaches to all aspects of how humans interact with computers. It takes a broad view and weaves threads of research in HCI involving Bayesian ideas into a consistent volume, covering probabilistic models in-the-loop, efficient Bayesian optimisation with users, principled Bayesian statistical analyses of empirical evaluation, and the wider implications of Bayesian philosophy and Bayesian models of human cognition on interaction. # Structure and theme ## Motivation Bayesian statistics offers a consistent and well-founded basis for understanding and designing all levels of human-computer interaction. A coherent core of mathematical concepts give rise to probabilistic ways of reasoning about the world and a powerful class of algorithms to implement them. While these have had substantial impact in computer science, particularly in machine learning, applications to human-computer interaction have been more patchy and the methods have remained relatively niche within HCI. Their full value both theoretically and application-wise has not been properly expounded. However, interest in these approaches is rapidly growing as it becomes clear that Bayesian methods have transformational potential for a huge range of HCI-related problems. The problems of HCI are particularly in need of robust, probabilistic methods that quantify and represent uncertainty; human-computer behaviour is too complex to be exhaustively modeled and is inherently noisy. Advances in computational power and software packages have rendered Bayesian techniques accessible to non-statisticians and these ideas can easily be applied to problems of interest. It is now feasible for researchers to implement and apply state-of-the-art algorithms for Bayesian inference in interactive applications. Beyond practical computational, the Bayesian perspective and distinctive world view offers original insights into HCI problems, new ways of thinking about designs and new ways of approaching gathering and interpreting empirical data. ## Broad aims The book will draw together research which involves Bayesian approaches across the whole spectrum of HCI and will, for the first time, present the state-of-the-art in these methods and consolidate them in a coherent structure and accessible style. Bayesian approaches give rise to a consistent basis of thinking about and talking about problems of interaction. We will address five core scientific problems in Bayesian HCI: * (i) how Bayesian models for richer and more robust analysis of empirical research; * (ii) how probabilistic computation can be used in-the-loop to robustly infer intention online; * (iii) how Bayesian optimisation can be used to optimise the structure and parameters of interfaces directly from noisy observations of users; * (iv) how to design the representation and visualisation of probabilistic models to communicate with appropriately quantified uncertainty; and * (v) how Bayesian predictive models of cognition might give rise to design insights in interaction. ## Structure The book is divided into six parts: ### Part I: Introduction and tutorial This part will motivate the use of Bayesian statistics in human-computer interaction, and illustrate the distinct aspects of Bayesian philosophy. It will guide the reader on how to rethink interaction problems probabilistically, and provide a concise tutorial on technical background and notation used for the remainder of the text. * Editors' Introduction * Summary of probability, Bayesian thinking, and notation * Tutorial on Bayesian modelling in interaction design ### Part II: Bayesian methods in empirical research This part will cover the use of Bayesian statistical models for analysis of empirical work in human-computer interaction. This will contrast Bayesian approaches with widespread frequentist approaches in the HCI literature, and identify benefits of a Bayesian approach. It will discuss HCI specific problems in Bayesian modelling, and case studies of Bayesian models in interaction evaluation. ### Part III: Probabilistic inference of intent and probabilistic interfaces This part will cover Bayesian methods to infer intention in the interaction loop. Bayesian methods offer robust ways to infer intentions and fuse together input channels, and can accurately represent and propagate uncertainty. It will discuss practical approaches, designs and algorithms to build user interfaces from probabilistic components. ### Part IV: Bayesian optimisation for elicitation and adaptation Bayesian optimisation offers ways of adapting interfaces to maximise objectives based on user observations. Gathering input from users is expensive and noisy. This part will cover processes to efficiently construct statistical surrogate functions to adapt interfaces or elicit preferences, from small-sample single user problems to large scale mass trial robust Bayesian A/B testing. ### Part V: Communication of Bayesian models This part will cover how Bayesian models can be communicated to users, including the problems of visualisation of uncertainty, interaction with probabilistic models, exploration of hypotheses and supporting reliable decision making under uncertainty. ### Part VI: Bayesian cognitive modeling Bayesian models of cognition propose that the human brain operates on approximately Bayesian principles. Theories of the “Bayesian brain” and the “predictive mind” offer novel perspectives on how humans make sense of the world and act upon it. This leads to consideration of how interfaces could be designed under these psychological models.