SLIDE 1
www.vistrails.org
Outline
- Provenance reuse
- We have all this rich metadata - let’s use it
- Query-by-example
- Visualization by Analogy
- (VisTrails intro)
- Transparent provenance tracking
Querying and Creating Visualizations by Analogy Carlos E. - - PowerPoint PPT Presentation
Querying and Creating Visualizations by Analogy Carlos E. - - PowerPoint PPT Presentation
Querying and Creating Visualizations by Analogy Carlos E. Scheidegger, Huy T. Vo, David Koop, Juliana Freire, Cludio T. Silva SCI Institute, School of Computing University of Utah Outline Provenance reuse We have all this rich
SLIDE 2
SLIDE 3
www.vistrails.org
Related Work
- Visualization Systems and Libraries
- AVS, DX, SCIRun, VTK
- History tracking and formalisms
- Jankun-Kelly et al’s pset-calculus
- Kreuseler et al, VDM history
- Brodlie’s et al’s GRASPARC
- VisTrails
SLIDE 4
www.vistrails.org
Provenance
- The “pedigree” of an artifact
- Where did it come from? Who held it?
SLIDE 5
www.vistrails.org
Provenance in VisTrails
- Process provenance
- How was this visualization created?
SLIDE 6
www.vistrails.org
Version Tree
- Persistent
- Transparent
- Reuse
- Can we do
better than just presenting?
SLIDE 7
www.vistrails.org
Why not query languages?
SLIDE 8
www.vistrails.org
Why not query languages?
wf{*}: upstream(x) union x where x.module = “SoftMean” and executed (x) and y in upstream(x) and y.module = “AlignWarp” and y.parameter(“model”) = “12”
SLIDE 9
www.vistrails.org
Why not query languages?
wf{*}: upstream(x) union x where x.module = “SoftMean” and executed (x) and y in upstream(x) and y.module = “AlignWarp” and y.parameter(“model”) = “12”
This is still only mildly better than straight SQL... Does not expose mapping to relational schema
SLIDE 10
www.vistrails.org
Query-by-Example
- Do not teach the user new forms of interaction!
SLIDE 11
www.vistrails.org
Visualization by Analogy
- Create new visualizations by saying “do as they
did”
- Specify what, not how
SLIDE 12
www.vistrails.org
Query-by-Example
- Trivially reducible from MAX-CLIQUE
- ... and MAX-CLIQUE is NP-Complete
- ... and MAX-CLIQUE is fundamentally hard to
approximate
- Solution: algorithm tailored to problem domain
SLIDE 13
www.vistrails.org
1 2 3
Query-by-Example
- Split every subgraph in topologically sorted layers
- Ok, since all pipelines are DAGs in VisTrails
SLIDE 14
www.vistrails.org
Query-by-Example
- Now search for layers that are connected in the
same way in the database
1 2 3
Query Database
SLIDE 15
www.vistrails.org
4 5 1 2 3
Match
Query-by-Example
- Now search for layers that are connected in the
same way in the database
1 2 3
Query Database
SLIDE 16
www.vistrails.org
1 4 2 3
No match
4 5 1 2 3
Match
Query-by-Example
- Now search for layers that are connected in the
same way in the database
1 2 3
Query Database
SLIDE 17
www.vistrails.org
1 2 3
No match
1 4 2 3
No match
4 5 1 2 3
Match
Query-by-Example
- Now search for layers that are connected in the
same way in the database
1 2 3
Query Database
SLIDE 18
www.vistrails.org
Query-by-Example
- Might return false positives - it ignores the
particular connectivity between topological layers
1 2 3
Query Database
- Not too harmful - most modules cannot connect to
- ne another
SLIDE 19
www.vistrails.org
Query-by-Example
- Might return false positives - it ignores the
particular connectivity between topological layers
4 5 1 2 3
1 2 3
Query Database
- Not too harmful - most modules cannot connect to
- ne another
SLIDE 20
www.vistrails.org
Query-by-Example
- Might return false positives - it ignores the
particular connectivity between topological layers
4 5 1 2 3
1 2 3
Query
4 5 1 2 3
Database
- Not too harmful - most modules cannot connect to
- ne another
SLIDE 21
www.vistrails.org
4 5 1 2 3
Query-by-Example
- Might return false positives - it ignores the
particular connectivity between topological layers
4 5 1 2 3
1 2 3
Query
4 5 1 2 3
Database
- Not too harmful - most modules cannot connect to
- ne another
SLIDE 22
www.vistrails.org
QBE Demo
SLIDE 23
www.vistrails.org
Vistrail diffs
- A version tree stores a set of actions
- Each action is a function on the set of all
possible visualizations:
- We can use those to determine the difference
between visualizations
- Moving up, then down the version tree
V → V an ◦ an−1 ◦ an−2 · · · ◦ a0
SLIDE 24
www.vistrails.org
- Action to go from A to B is
Vistrail diffs
A B
a0 a1 a2 a3
a3 ◦ a2 ◦ a−1
- a−1
1
SLIDE 25
www.vistrails.org
Visualization by Analogy
- A diff is a template: reapply it elsewhere
- How do we match two pipelines?
SLIDE 26
www.vistrails.org
- Compute the difference
- Compute the map
- Apply to
- Compute the new pipeline
Algorithm Overview
mapac = map(pa, pc) δab = ∆(pa, pb) pd = δ
cb(pc)
δ
cb = mapac(δab)
δab
mapac
SLIDE 27
www.vistrails.org
Visualization by Analogy
- Simplest version is again reducible from MAX-
CLIQUE
- We will now use a probabilistic argument to create
a Markov chain
SLIDE 28
www.vistrails.org
How does it work?
- Module compatibility: prior
- Independent of graph topology
- Probability of match between a pair
- Dependent of graph topology
- Linear combination of probability of match in
the neighborhood pairs and data
- This is a Markov chain!
f : M2 → [0,1]
SLIDE 29
www.vistrails.org
How does it work?
- Graph product G of the two input graphs
- each vertex in G represents a possible match
- similarity is then defined as
- is an eigenvector of
- It is the limit distribution of the transition matrix
π = αA(G)π+(1−α)c(G) = MGπ
π MG
SLIDE 30
www.vistrails.org
GA GB GA ×GB How does it work?
SLIDE 31
www.vistrails.org
GA GB GA ×GB How does it work?
SLIDE 32
www.vistrails.org
GA GB GA ×GB How does it work?
SLIDE 33
www.vistrails.org
GA GB GA ×GB How does it work?
SLIDE 34
www.vistrails.org
GA GB GA ×GB How does it work?
SLIDE 35
www.vistrails.org
How does it work?
Each node is assigned some initial value. (It doesn’t matter which, as long as the values sum to one!)
SLIDE 36
www.vistrails.org
How does it work?
pk(a0 → b0) pk(a0 → b1) pk(a0 → b2) pk(a0 → b3) pk(a1 → b0)
SLIDE 37
www.vistrails.org
How does it work?
pk(a0 → b0) pk(a0 → b1) pk(a0 → b2) pk(a0 → b3) pk(a1 → b0) pk+1(a0 → b0) = (1 − α)c(a0, b0) + α/3 (pk(a0 → b3)+ pk(a0 → b1)+ pk(a1 → b0))
SLIDE 38
www.vistrails.org
How does it work?
pk+1(a0 → b0) = (1 − α)c(a0, b0) + α/3 (pk(a0 → b3)+ pk(a0 → b1)+ pk(a1 → b0))
SLIDE 39
www.vistrails.org
How does it work?
pk+1(a0 → b0) = (1 − α)c(a0, b0) + α/3 (pk(a0 → b3)+ pk(a0 → b1)+ pk(a1 → b0)) c(a0, b0)
SLIDE 40
www.vistrails.org
How does it work?
pk+1(a0 → b0) = (1 − α)c(a0, b0) + α/3 (pk(a0 → b3)+ pk(a0 → b1)+ pk(a1 → b0)) c(a0, b0) Do it for all nodes, until convergence
SLIDE 41
www.vistrails.org
How does it work?
- is defined over graph product
- For each module in the second pipeline, pick
maximal value of on first pipeline: this is the match
- Many others possible
π π
SLIDE 42
www.vistrails.org
The matching algorithm
SLIDE 43
www.vistrails.org
The matching algorithm
SLIDE 44
www.vistrails.org
The matching algorithm
SLIDE 45
www.vistrails.org
The matching algorithm
SLIDE 46
www.vistrails.org
Failure Modes
- Analogies are not fool-proof
SLIDE 47
www.vistrails.org
Case study
- Creating a complex visualization out of simple
- nes
- (demo)
SLIDE 48
www.vistrails.org
Discussion
- If your system can encode actions as functions on
the space of objects of interest, store these explicitly
- That will be your “version tree” - everything else is
just the same
- Easy to incorporate domain-specific knowledge in
analogies: change and
A(G) c(G)
SLIDE 49
www.vistrails.org
Acknowledgments
- Sarang Joshi, Suresh Venkatasubramanian, Erik
Anderson, João Comba
- VisTrails dev team
- Many open source packages and devs: VTK, SciPy,
teem, matplotlib
- VisTrails is open source! http://www.vistrails.org
- Shameless plug: Visit the SCI booth!
- NSF, DOE, IBM Faculty Award
SLIDE 50
www.vistrails.org
Thank you!
- Questions?
SLIDE 51
www.vistrails.org
Too much data
- We are better off with visualization systems than
without - but it’s still pretty messy
SLIDE 52
www.vistrails.org