Chapter 15

Adaptive Interfaces and Agents

Anthony Jameson
International University in Germany

 

Outline

Introduction

Concepts

chapter Preview

Functions: Supporting System Use

Taking Over Parts of Routine Tasks

Adapting the Interface

Giving Advice About System Use

Conducting a Dialogue

Functions: Supporting Information Acquisition

Helping Users to Find Information

Support for Browsing

Support for Query-Based Search or Filtering

Spontaneous Provision of Information

Tailoring Information Presentation

Recommending Products

Supporting Collaboration

Supporting Learning

Usability Challenges

Predictability and Transparency

Controllability

Unobtrusiveness

Privacy

Breadth of Experience

Obtaining Information About Users

Explicit Self-Reports and Self-Assessments

Self-Reports About Objective Personal Characteristics

Self-Assessments With Respect to General Dimensions

Self-Reports on Specific Evaluations

Responses to Test Items

Nonexplicit Input

Naturally Occurring Actions

Previously Stored Information

Low-Level Indices of Psychological States

Signals Concerning the Current Surroundings

Learning, Inference, and Decision Making

Classification Learning

Example Systems

Requirements

Collaborative Filtering

Requirements

Combinations With Other Paradigms

Decision-Theoretic Methods

Basic Characteristics

Potential Advantages

Requirements

Other Approaches

Techniques for Plan Recognition

The Stereotype Approach

Empirical Methods

Wizard-of-Oz Studies

Assessing Accuracy

Assessing Usability

Controlled Studies

Assessing Accuracy

Assessing Usability

Studies of Actual System Use

Observation and Interviewing

Use of Questionnaires

Assessing Accuracy

Assessing Usability

The Future of User-Adaptive Systems

Growing Need for User-Adaptivity

Diversity of Users and Contexts of Use

Number and Complexity of Interactive Systems

Scope of Information to Be Dealt With

Increasing Feasibility of Successful Adaptation

Ways of Acquiring Information About Users

Advances in Techniques for Learning, Inference, and Decision

Attention to Empirical Methods

Concluding Remarks

Acknowledgements

References

 

Figures

Figure 15.1: Screen shot from SWIFTFILE, showing its three shortcut buttons for the ?ling of the current e-mail message. Note.F rom “Incremental Learning in SwiftFile,” by R. B. Segal and J.O.Kephart,2000,in P. Langley (Ed.), Machine Learning: Proceedings of the 2000 International Conference ,San Francisco: Morgan Kaufmann. Copyright 2000 by Morgan Kaufmann Publishers. Adapted with permission.

Figure 15.2: General schema for the processing in a user adaptive system. Ovals: input or output; rectangles: processing methods; cylinder: stored information; dotted arrows: use of information; solid arrows: production of results.

Figure 15.3: Application of the schema of Fig.15.2 to the example of SWIFTFILE.

Figure 15.4: Example of adaptation in SMART MENUS. U accesses the “Insert ”menu. Not finding the desired option, U clicks on the extension arrows and selects the “Field ”option. When U later accesses the same menu, “Field ”now appears in the main section.

Figure 15.5: Overview of adaptation in SMART MENUS.

Figure 15.6: Example of assistance offered by the LUMI`ERE prototype. U has just searched through several menus, selected the entire spreadsheet, and paused. Note.  From “The Lumi`ere Project: Bayesian User Modeling for Inferring the Goals and Needs of Software Users,” by E. Horvitz, J. Breese, D. Heckerman, D. Hovel, and K.Rommelse,1998, in G. F. Cooper and S. Moral (Eds.), Uncertainty in Artificial Intelligence: Proceedings of the Fourteenth Conference (pp.256 –265),San Francisco: Morgan Kaufmann.C opyright 1998 by Morgan Kaufmann Publishers. Adapted with permission.

Figure 15.7: Overview of adaptation in LUMI`ERE.

Figure 15.8: Excerpts from a dialog with TOOT. Note. From “Predicting and Adapting to Poor Speech Recognition in a Spoken Dialogue System,” by D. J. Litman and S.Pan,2000,in Proceedings of the Seventeenth National Conference on Artificial Intelligence , Austin, TX (pp.722 –728). Copyright 2000 by the American Association for Artificial Intelligence. Adapted with permission.

Figure 15.9: Overview of adaptation in TOOT.

Figure 15.10: Sequence of three screens presented by the ADAPTIVE NEWS SERVER. Note .From a slide supplied by Michael J. Pazzani. Copyright 2000 by Michael J.Pazzani. Adapted with permission.

Figure 15.11: Overview of adaptation in ADAPTIVE NEWS SERVER.

Figure 15.12: Part of a screen from the AVANTI tourist information system. Note. From “Adaptable and Adaptive Information Provision for All Users, Including Disabled and Elderly People,” by J. Fink, A. Kobsa, and A.Nill,1998,New Review of Hypermedia and Multimedia (Vol.4,pp.163 –188).Copyright 2000 by Taylor Graham Publishers. Adapted with permission.

Figure 15.13: Overview of adaptation in AVANTI.

Figure 15.14: Part of a screen from the MOVIECENTRAL ?lm recommendation Web site describing the movie 2001, A Space Odyssey. Screen shot was made from http://www.qrate.com/in January 2001 and edited for compactness. This Web site is no longer in operation.

Figure 15.15: Overview of adaptation in MOVIECENTRAL.

Figure 15.16: Screen shot from the PHELPS system and overview of adaptation. Left: U is having difficulty arranging an escorted temporary absence for a prisoner. S offers information on a number of possible helpers at various places in Canada. The window in the right-hand side of the screen shows a profile of a potential helper’s knowledge of the task in question. Note. From “Supporting Peer Help and Collaboration in Distributed Workplace Environments,” by J. E. Greer, G. I. McCalla, J.A.Collins, V. S. Kumar, P. Meagher, and J. Vassileva,1998,International Journal of AI and Education (Vol.9,pp.159 –177).Copyright 1998 by The International Artificial Intelligence in Education Society. Adapted with permission.

Figure 15.17: Overview of adaptation in PHELPS.

Figure 15.18: Example screen from ELM-ART showing the sys tem ’s assessment of the suitability of particular learning units for the current user. Screen shot made from http://www.psychologie.uni –trier.de:8000/elmart in December 2000.In the actual system, the different colors of the folder icons are clearly distinguishable. Adapted with the permission of Gerhard Weber.

Figure 15.19: Overview of adaptation in ELM-ART.

Figure 15.20: Overview of usability challenges for user-adaptive systems. Solid and dashed arrows denote positive and negative causal influences, respectively; further explanation is given in the text.

Figure 15.21: Example of a screen with which the LIFESTYLE FINDER elicits demographic information. Note. From “Lifestyle Finder: Intelligent User Pro ?ling Using Large-Scale Demographic Data,” by B.Krulwich,1997,AI Magazine,18 (2),pp.37-45. Copyright 1997 by the American Association for Artificial Intelligence. Adapted with permission.

Figure 15.22: Rating scale from the PERSONALOGIC decision guide for dog seekers. Part of a screen shot made from http://www.purina.personalogic.com in July 2001.Before mid-2001,many similar recommenders had been available via http://www.personalogic.com

Figure 15.23: Summary of a typical algorithm for generating recommendations through collaborative filtering.

Figure 15.24: Part of a Bayesian network used in the LUMI`ERE prototype for inferring the likelihood that U requires assistance. Note. From “The Lumi`ere Project: Bayesian User Modeling for Inferring the Goals and Needs of Software Users,” by E. Horvitz, J. Breese, D. Heckerman, D. Hovel, and K. Rommelse, 1998,in G. F. Cooper and S. Moral (Eds.), Uncertainty in Artificial Intelligence: Proceedings of the Fourteenth Conference (pp.256 –265), San Francisco: Morgan Kaufmann. Copyright 1998 by Morgan Kaufmann Publishers. Adapted with permission.

Figure 15.25: Summary of a Wizard-of-Oz study conducted in the LUMI`ERE research project.Summarized on the basis of p.258 of Horvitz et al.,1998.

Figure 15.26: Comparison of alternative variants of SWIFTFILE on the basis of data concerning the use of a nonadaptive system. Upper curves: Accuracy of SWIFTFILE ’s predictions with different numbers N of suggestion buttons, shown on the x-axis. Lower curves: Accuracy of the naive strategy that simply predicts the N most frequently used folders. Note. From “MailCat:An Intelligent Assistant for Organizing E-mail,” by R. B. Segal and J.O.Kephart,1999,in Proceedings of the Third International Conference on Autonomous Agents (pp.276 –282).Copyright 1998 by the Association for Computing Machinery, Inc. Adapted with permission.