Explaining Preference LearningExplaining Preference LearningAlyssa Glass

CS229 Final ProjectComputer Science Department, Stanford University

Augment PLIANT to gather additional meta-information

about the SVM itself: Support vectors identified by SVM Support vectors nearest to the query point Margin to the query point Average margin over all data points Non-support vectors nearest to the query point Kernel transform used, if any

Represent SVM learning and meta-information as

justification in Proof Markup Language (PML), adding SVM

rules as needed.

Design abstraction strategies for presenting justification

to user as a similarity-based explanation.

(Work on PML representation and abstraction strategies is on-going;

details will be in final report.)

Active Preference Learning in PLIANTActive Preference Learning in PLIANT(Yorke-Smith et al. 2007)(Yorke-Smith et al. 2007)

MotivationMotivationStudies of users interacting with systems that learn

preferences show that, when the system behaves

incorrectly, users quickly lose patience and trust in the

system. Even when the system is correct, users view such

outcomes as “magical” in some way, but are unable to

understand why a particular suggestion is correct, or

whether the system is likely to be helpful in the future.

We describe the augmentation of a preference learner to

provide meaningful feedback to the user through

explanations. This work extends the PLIANT (Preference

Learning through Interactive Advisable Nonintrusive

Training) SVM-based preference learner, part of the PTIME

personalized scheduling assistant in the CALO project.

AcknowledgementsAcknowledgementsWe thank Melinda Gervasio, Pauline Berry, Neil Yorke-Smith, and Bart

Peintner for access to the PLIANT and PTIME systems, the above

architecture picture, and for helpful collaborations, partnerships, and

feedback on this work. We also thank Deborah McGuinness, Michael

Wolverton, and Paulo Pinheiro da Silva for the IW and PML systems, and

for related discussions and previous work that helped to lay the foundation

for this effort. We thank Mark Gondek for access to the CALO CLP data,

and Karen Myers for related discussions, support, and ideas.

Usability and Active LearningUsability and Active Learning

Providing Transparency into Providing Transparency into Preference LearningPreference Learning

Select ReferencesSelect References PLIANT:

Yorke-Smith, N., Peintner, B., Gervasio, M., and Berry, P. M. Balancing the Needs of Personalization and Reasoning in a User-Centric Scheduling Assistant. Conference on Intelligent User Interfaces 2007 (IUI-07) (to appear).

PTIME:Berry, P., Gervasio, M., Uribe, T., Pollack, M., and Moffitt, M. A Personalized Time Management Assistant. AAAI Spring Symposium Series, Stanford, CA, March 2005.

Partial preference updates:Joachims, T. Optimizing Search Engines using Clickthrough Data. Proceedings of the ACM Conference on Knowledge Discovery and Data Mining (KDD), ACM, 2002.

User study on explaining statistical machine learning methods:Stumpf, S., Rajaram, V., Li, L., Burnett, M., Dietterich, T., Sullivan, E., Drummond, R., and Herlocker, J. Towards Harnessing User Feedback for Machine Learning. Conference on Intelligent User Interfaces 2007 (IUI-07) (to appear).

System WorkflowSystem Workflow1. Elicit initial preferences from user (A vector from above)

2. User specifies new meeting parameters

3. Constraint solver generates candidate schedules (Z’s)

4. Candidate schedules ranked using evaluation function, F`(Z)

5. Candidate schedules presented to user in (roughly) the

calculated preference order, with explanations for each one

6. User can ask questions, then chooses a schedule (Z)

7. Preferences (ai and aij weights) are updated based on choice

Features:

1. Scheduling windows for requested meeting

2. Duration of meeting

3. Overlaps and conflicts

4. Location of meeting

5. Participants in meeting

6. Preferences of other meeting participants

Model of preferences: aggregation function, a 2-order

Choquet integral over partial utility functions based on the

above features learning 21 coefficient weights:

F(z1, …, zn) = i ai zi + i,j aij (zi zj)

where each zi = ui(xi), the utility for criterion i based on value xi

Evaluation function: combine learned weights with initial

elicited preferences:

F`(Z) = AZ + (1-) WZEach schedule chosen by the user provides information

about a partial preference ordering, as in (Joachims 2002).

ArchitectureArchitecture

Several user studies show that transparency is key to

trusting learning systems:

Our trust study

Lack of understanding of update gives appearance that

preferences are ignored seems untrustworthy

Typical user reaction: “I trust [the system’s] accuracy,

but not its judgment.”

PTIME user study (Yorke-Smith et al. 2007)

“The preference model must be explainable to the user

… in terms of familiar, domain-relevant concepts.”

Explaining statistical ML methods (Stumpf et al. 2007)

Looked at explaining naïve Bayes learner and rule-

learning system (classification), not SVMs

Rule-based explanations most easily understood, but

similarity-based explanations found to be more natural

and easily trusted

Our approach: extend similarity-based explanations to SVM

learning

explanation

SVM meta-information

selected schedule

solution set

Calendar Manager

Constraint Reasoner

preference profile

PLIANT

scheduling request

presentation set

1

2

3

5

6

74

SVM Explainer

current profile4

5