A Preference Evolution Perspective on Lifelong User Modeling

This workshop paper briefly presents a theoretical framework for explaining and predicting differences and changes in users’ preferences concerning the systems they use—in particular, user-adaptive systems. The ultimate purpose of this framework is to increase our ability to understand and predict users’ preferences and to design (user-adaptive) systems accordingly. In our workshop presentation, we will apply this framework to the other contributions to the lifelong user modeling workshop. 1 The Nature of This Submission Since this workshop submission does not fall into either of the two normal submission categories, we will begin by explaining its intent. In the context of the targeted research unit PREVOLUTION, the authors have been developing a theoretical framework for understanding and predicting users’ preferences with regard to interactive systems, in particular those that involve some sort of intelligent processing. This framework has been applied to a number of specific systems and problems, and it formed the basis for the full-day IUI 2009 workshop on Users’ Preferences Regarding Intelligent User Interfaces: Differences Among Users and Changes Over Time ([1]; http://prevolution.fbk.eu). Some of the participants in that workshop are currently preparing articles for a special journal issue on this topic. Because of its attention to the evolution of preferences over time, this framework has a good deal of relevance to the topic of lifelong user modeling, as will be explained below. In the Lifelong User Modeling workshop, we will (a) explain the framework on a general level and (b) apply it to the particular systems and studies that are described in the papers of the other workshop participants. In this way, we hope to be able to encourage cross-fertilization among the workshop participants, while at the same time expanding and enriching our own theoretical framework. In the following sections, we will first give a brief discussion of our evolving framework and then (in Section 5) discuss its relevance to the topic of lifelong user modeling. ⋆ The research described in this paper is being conducted in the targeted research unit Prevolution (code PsychMM), which is funded by the Autonomous Province of Trento. 2 Introduction: Choices People Make Regarding User-Adaptive Systems Because building user-adaptive systems that accurately model their users is in itself such a challenging problem, it is easy to forget that the users of such systems are not just passive entities waiting to be modeled; they regularly make choices about how they are going to deal with the modeling and adaptation that is offered by the system. But ignoring these choices can be a big mistake. As anyone who looks at the way in which users respond to user-adaptive systems will have noticed, people often exhibit striking differences in the individual choices that they make as well as in the patterns of choice that they exhibit when they are faced with multiple choices of the same type (cf. [2]). These differences can arise at different levels: – When faced with a choice between an adaptive and a nonadaptive version of the system, typically some people prefer the former and some prefer the latter (see, e.g., [3]). – When a decision about adaptivity can be made on a case-by-case basis, some users may choose adaptivity more often than others, and their criteria for doing so may be different. – Where an adaptive system offers some flexibility in the way in which the adaptivity can be used (for example, allowing the user to choose between automatic adaptation and adaptation on demand, as in [4]), users may again make different choices, for different reasons. When we observe people using a user-adaptive system over an extended period of time, in addition to the differences just mentioned we typically also see changes in users’ patterns of choice over time. People may come to appreciate the adaptivity more or less, or (on a more detailed level) they may use the adaptive functionality in different ways. There are several reasons why it is important to be able to understand and anticipate differences and changes in users’ patterns of choice at least to some extent: – We can then better make sense of the sometimes puzzling results that emerge from studies of the use of user-adaptive systems. – By being aware of typical changes over time, we are in a better position to extrapolate from the results of such studies, which usually are restricted to a limited period of time. – When designing a user-adaptive system, we can anticipate likely differences and changes in patterns of choice and try to design the system accordingly. – The specific ways in which patterns of choice differ and change often turn out to be surprising and not necessarily in the best interests of the users in question; we may say that a user’s patterns of choice are not necessarily well-aligned with their actual needs (though defining and determining “actual needs” is admittedly a tricky matter). It is therefore not enough simply to provide a good user-adaptive system, enable users to learn how to operate it correctly, and assume that users will arrive at patterns of choice that are in their own best interests. Fig. 1. Screenshot of the experimental, profile-based version of Google’s personalized search that was used in Google’s “Lab” in 2005. (By moving the slider in the upper left, the user could change the degree to which the searc results were reranked the on the basis of a previously specified interest profile. The reordering was visualized with an animation.) Since speaking of “choices or patterns of choice” can be awkward, in the rest of this paper we will often use the broader term preferences instead; but the discussion above about what is more specifically meant should be borne in mind. 3 A Framework for Understanding and Prediction When looking at any particular system or study, we may seem to be able to explain the observed preference differences and changes in terms of one or two fairly obvious factors, such as the user’s level of experience or their personality traits. But when we look at a broader range of experience with user-adaptive systems, we see that there is a great variety of factors that can influence user’s preferences; a focus on any subset is likely to lead to inaccurate conclusions. Therefore, in this brief paper we offer a high-level view of users’ preferences that aims to capture in a structured way all of the factors that can lead to differences among users and changes over time. This theoretical framework has been continually evolving and expanding as we have applied it to one case after the other, and we expect it to evolve further on the basis of feedback at the Lifelong User Modeling Workshop, as we engage in dialog with participants about how the theory can be applied to their specific systems. The framework is presented in Tables 1, 2, and 3 (which are separated only because of the limited page size). On the highest level, three categories of factors are distinguished that can help to explain differences in users’ preferences: Users’ needs, The system’s properties, and Aspects of process of preference formation and evolution. For each individual factor, it is usually possible to offer some hypothesis about how the factor changes over time (see the second column); these hypotheses are relevant Table 1. Overview of the factors that can lead to differences and changes in users’ preferences concerning user-adaptive systems (continued in the next two tables). (Symbols like “+++” indicate roughly, on a scale from “–” to “+++”, the amount of attention that the factor has attracted so far in the literature.) Factor Systematic evolution Reason(s) for evolution