Visualizing and Understanding Large-Scale Bayesian Networks

placecornersdeceitΤεχνίτη Νοημοσύνη και Ρομποτική

7 Νοε 2013 (πριν από 3 χρόνια και 9 μήνες)

76 εμφανίσεις

Visualizing and Understanding Large-Scale
Bayesian Networks
Michele Cossalter, Ole Mengshoel, Ted Selker
Problem Domain
Problem Domain
Current Tools
Current Tools
LARGE HETEROGENEOUS DATASETS


Some information can be encoded in a node-edge structure,

other dimensions of data need a different representation


Bayesian networks
[1]:

A directed acyclic
graph
where nodes represent variables and
edges represent dependencies

Conditional probability tables (
CPTs
)
describing the probability
for the value of each node given the value of its parents


Other examples: e-mail collections, electrical networks, oil and
gas plants, manufacturing units
VISUALIZATION GOALS


Improved interactive and visual support for debugging the
network structure and tuning CPT parameters


Structure, states, and CPTs are all important


Scalability as the size and connectivity of the network increase
CPT BUBBLE ANCHORS


Cartoon-inspired bubble lines (B)
facilitate association
of CPTs
(A) with respective nodes
INTEGRATED TIME SERIES DISPLAY


Integration of time series
sensor readings together with the
CPTs (C), enriched with horizontal lines showing the thresholds
used to discretize the states
CPT MERGING


Facilitated comparison
of different nodes by allowing data from
two nodes to be merged into a single box by drag and drop
NETWORK OVERVIEW


Overview+detail
[5]: simultaneously displaying both an overview
(D) and detailed view (E) of the network helps orienting
Our NetEx Tool
Our NetEx Tool
References
References
1. Pearl, J. 1988. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible

Inference. San Mateo, CA: Morgan Kaufmann.
2. Andersen, S. K.; Olesen, K. G.; Jensen, F. V.; and Jensen, F. 1989. HUGIN - A Shell for
Building Bayesian Belief Universes for Expert Systems. In Proc. of IJCAI’89, 1080– 1085.
3. Druzdzel, M. J. 1999. SMILE: Structural Modeling, Inference, and Learning Engine and
GeNIe: A Development Environment for Graphical Decision-Theoretic Models. In Proc. of
AAAI’99, 902–903.
4. Wang, H., and Druzdzel, M. J. 2000. User Interface Tools for Navigation in Conditional
Probability Tables and Elicitation of Probabilities in Bayesian Networks. In Proc. of UAI’00,
617–625.
5. Cockburn, A.; Karlson, A.; and Bederson, B. B. 2008. A Review of Overview+Detail,
Zooming, and Focus+Context Interfaces. ACM Computing
Surveys 14(1):1–31.
Sponsor: NSF, Cylab

Acknowledgements:

This material is based, in part, upon work by Michele Cossalter, Ole
J. Mengshoel
and Ted Selker supported by NSF grants CCF-0937044 and ECCS-0931978
CHARACTERISTICS


Focus on editing, learning, and inference (e.g. Hugin [2])


Some tools support navigation in large CPTs (e.g. bar charts and
pie charts in GeNIe&SMILE [3], CPTree and shrinkable CPT [4])
LIMITATIONS OF CURRENT TOOLS


Little support for effective user-driven exploration of large-scale
networks


Traditional zooming techniques implemented in current tools
allow the user to see details, but at the price of loosing the global
structure of the network


No visual link between the CPTs and the corresponding nodes:

User need to match CPTs with nodes based only on the labels

Users need to visually inspect all the nodes of the network


Lack of support for easily comparing underlying data (e.g. time
series) which might provide a richer understanding of the model
Related NetEx Results
Related NetEx Results
USER STUDY


25 students with some knowledge of electrical circuits tried to
detect faults in the ADAPT network


Subjects viewed plots to find which and when components failed
in a simple and complex task


12 subjects used NetEx for Task 1 (simple) and a baseline tool
(Intelliviz) for Task 2 (complex), 13 subjects did the opposite


For both tasks,
NetEx users

were

more accurate
than Intelliviz
users, with ANOVA p=.061 for Task 1 and p=.075 for Task 2


On the complex task Intelliviz is faster (p=.018), but this is not
the case for the simple task


NetEx users perceived the difficult task to be harder

(p=.060) but
were more confident
about their performance
(p=.053) and
were more satisfied
with the tool (p=.019)