19
Visual Cues for the Interactive Learning of Bayesian Networks Lisa Li Omar Ramadan Phoebe Schmidt CS 294-10: Visualization Fall 2014 UC Berkeley Computer Science

Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Visual Cues for the Interactive Learning of Bayesian NetworksLisa LiOmar RamadanPhoebe Schmidt CS 294-10: Visualization

Fall 2014UC Berkeley Computer Science

Page 2: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

What is a Bayes’ Net?

Page 3: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Bayes’ Nets

Slide: Dan Klein CS188 Fa13 “Bayes’ Nets Representation”

Page 4: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Slide: Dan Klein CS188 Fa13 “Bayes’ Nets Representation”

Page 5: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Why are they useful?

● causal relationships help us in exploratory data analysis, make predictions

● combine domain knowledge with data

(Heckerman, 1995)

Page 6: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

ProblemGiven a set of data, it is NP-hard to find an

optimal Bayes’ Net structure to represent their conditional relationships.

(existing interactive tools do not visually encode heuristics, poor UI experience)

Page 7: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

X4X3

X4X2

Finding the optimal structure can be modeled as a state space search

ADD(X1→ X3)

Score Δ: +436

X1

X4

X2X1

X3

X2X1

X3

X4

X2X1

X3 X4

X2X1

X3

ADD(X1→ X4)

Score Δ: +218 ADD(X1→ X2)

Score Δ: -28

ADD(X1→ X4)

Score Δ: +13

Problem: to find the most probable Bayes-network given data

Page 8: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Approximate algorithms for learning the structure of Bayes Nets

Network scoring methods- BDeu score

[Buntine, 1991]

- K2 score[Cooper and Herskovits, 1992]

- Cross entropy[Chickering, Geiger, and Heckerman, 1995]

Heuristic search methods- Local search with restarts

[Johnson, 1985]

- Simulated annealing[Chickering, Geiger, and Heckerman, 1996]

- Max min hill climbing[Tsamardinos, Brown, and Aliferis, 2005]

Page 9: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Methods for interactively learning Bayes net structures

[Myllymäki et. al, 2002][Bermejo et. al, 2012]

Tools for stepping through learning algorithms

Tools for visualizing relationship strength [Sucar and Arroyo, 1998][Ebert-Uphof, 2006]

More BN software:http://www.cs.berkeley.edu/~murphyk/Bayes/bnsoft.html

Page 10: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Motivation● Effectively building a Bayes net requires integrating

domain knowledge with insights from dataa. Inferring BN structure from data alone is difficult and

leads to overfittingb. Domain expert has beliefs but would like to be able

to validate with datac. No way to visualize data insights while manually

building BNs

Page 11: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Description

Create an interactive tool that aids domain experts to learn the structure of Bayes Net with insights from local scoring heuristics

Page 12: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Smoker

LungCancer

X-Ray

Bronchitis

Dyspnea

Tuberclosis

VisitToAsia

TuberclosisOrCancer

Tool Box

Look Ahead

Add an edge

Remove an edge

Invert an edge

Storyboard: Identifying variable to edit

Page 13: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Smoker

LungCancer

X-Ray

Bronchitis

Dyspnea

Tuberclosis

VisitToAsia

TuberclosisOrCancer

Tool Box

Look Ahead

Add an edge

Remove an edge

Invert an edge

AddRemoveInvert

2251.11

983.41869.32

1001.43

1679.02

Storyboard: Identifying proposed changes to node

877.6

Page 14: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Look Ahead

What: Show preview of the next K steps from learning algorithm

Why: Learning algorithms are derived from heuristics and changes are not always optimal. Expert can nevertheless gain insights from heuristics while using domain expertise to prevent overfitting to training data

Page 15: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Smoker

LungCancer

X-Ray

Bronchitis

Dyspnea

Tuberclosis

VisitToAsia

TuberclosisOrCancer

Tool Box

Look Ahead

Add an edge

Remove an edge

Invert an edge

1 105

Storyboard: Look ahead into search algorithms

Page 16: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

1. Graph building toola. add, remove, reorient edgesb. hard-coded heuristic data

Milestones

2. Interactive heuristicsa. K2 algorithmb. input: (graph, data) output: (score for each node)c. parameters: number of samples to compute

heuristic with (for speed)

Page 17: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Milestones, ctd

3. Look-ahead Grapha. implement greedy algorithm, k look-ahead stepsb. display algorithm’s “next k best modifications”c. resulting graph

Page 18: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Proposed Timeline

Code Paper Date

Interactive BN builder w/ hard-coded heuristic

Previous Work, Methods

11.17

Heuristics module finish Methods 11.24

Interactive look-ahead graph

References 11.28

usability testing extra time to for bugs and setbacks

Results, Discussion, Future Work

12.5

Page 19: Bayesian Networks Omar Ramadan Visual Cues for the ...vis.berkeley.edu/courses/cs294-10-fa14/wiki/images/d/da/...domain knowledge with insights from data a. Inferring BN structure

Questions