Toolkit to Support Intelligibility in Context Aware Applications - - PowerPoint PPT Presentation

Toolkit to Support Intelligibility in Context Aware Applications Context-Aware Applications P t d b M S li Presented by Mary Salinas

What problem are they trying l ? to solve?

• Context-aware applications are great for

Context aware applications are great for helping users to automatically serve users better but the complexity of models can make it difficult for users to understand them.

• Users can become frustrated and lose trust in

the applications. Applications must be intelligible and provide explanations of context awareness context-awareness

2 Worcester Polytechnic Institute 2

What did they do? What did they do?

• Created an architecture for generating a wide

Created an architecture for generating a wide range of explanations including Why, Why Not, How To, What, What If, Inputs, Outputs and Certainty.

• Create a library of reference implementations
• f explanation algorithms for 4 different

models. P id d f h

• Provide automated support for the common

explanations to promote good design practices

3

practices.

Worcester Polytechnic Institute 3

Surveyed existing use of d l model types

4 Worcester Polytechnic Institute 4

Rule-based Decision Model Rule based Decision Model

• Rules are usually

Rules are usually if/else logic or simple mapping

• Most popular

decision model.

• Used in

personalization, i i i i activity recognition, monitoring, location guides

5

guides

Worcester Polytechnic Institute 5

Decision Tree Model Decision Tree Model

• Similar to rules, but

Similar to rules, but decisions are made from top down rather than bottom up.

• Decision trees are

built from statistical data so they can data so they can model certainty from probability of leaves

6

probability of leaves.

Worcester Polytechnic Institute 6

Bayes Decision Model Bayes Decision Model

• Naïve Bayes is sum

Naïve Bayes is sum

• f evidence

assumes features are independent of each other.

• Includes prior

probabilities of selected class value selected class value and from each feature value

7

feature value

Worcester Polytechnic Institute 7

Hidden Markov Model Hidden Markov Model

• Apply weights of

Apply weights of evidence as similar to Bayes and include temporal factors as well

8 Worcester Polytechnic Institute 8

Implementation Implementation

• Implemented in Java based on Enactor and

Implemented in Java based on Enactor and Context Toolkit created by Anind.

• Explainer to generate explanations for model-

p g p independent types and one Explainer for each of the 4 decision model types.

• Reducer to remove explanations that include

too many reasons or each reason is too long.

• Presenter renders the explanation in form

suitable for users. Developers can build l P t if d d

9

several Presenters if needed.

Worcester Polytechnic Institute 9

Architecture for Rules and Cl ifi Classifiers

10 Worcester Polytechnic Institute 10

Claimed Results Claimed Results

• Allows for fast prototyping of context-aware

Allows for fast prototyping of context aware applications

• Provides lower barrier to providing

p g explanations

• Provides flexibility of using explanations

y g p

• Facilitate appropriate explanations

automatically.

11 Worcester Polytechnic Institute 11

What has happened since h ? then?

• Active development is continuing by Lim

Active development is continuing by Lim at http://www.contexttoolkit.org/

• Lim’s home page:

htt // b i li t/ http://www.brianlim.net/

12 Worcester Polytechnic Institute 12

Any Questions? Any Questions?

13 Worcester Polytechnic Institute 13