Upload
villerd
View
625
Download
1
Tags:
Embed Size (px)
Citation preview
Using Concept Lattices for Visual Navigation Assistance in Large
Databases
Application to a Patent Database
Jean Villerd, Sylvie Ranwez, Michel Crampes, David Carteret
LGI2P – École des Mines d’Alès, France I-Nova Company, Villeurbanne – Lyon, France
2
Outline
1. Context and Problem Setting
2. State of the Art
3. Objective : Intended Visualization
4. Proposal for a Visual Browsing Method
5. Discussion and Perspectives
3
1. Context and Problem Setting
Scalability issuesstorage of new information + 30 % each year [Lyman & Varivan 2003]
Visual scalability“capability of visualization representations and visualization tools to display massive data sets effectively, in terms of either the number or the dimension of individual data elements” [Eick & Karr 2002]
Binary-tree visualization of the Yahoo search engine bot crawling the experimental website
Microsoft Research's Netscan project
1. Context and Problem Setting | 2. State of the Art | 3. Expected Visualization | 4. Proposal for a Visual Browsing Method | 5. Perspectives
4
2. State of the Art
Information Visualization“focus + context” paradigm [Schneiderman 1996]
Collection Visualization collection’s structure
Grokker, Kartoo, TreeMaps semantic distance (PCA, MDS)
Molage
Molage
Grokker
1. Context and Problem Setting | 2. State of the Art | 3. Expected Visualization | 4. Proposal for a Visual Browsing Method | 5. Perspectives
our goal: merging both
5
FCA for Information Retrieval classification querying and browsing
Credo, MailSleuth, ImageSleuth
ImageSleuth [Eklund, Ducrou et al. 2006]
Can information visualization techniques improvethis lattice-based navigation process?
Credo [Carpineto et al. 2004]
1. Context and Problem Setting | 2. State of the Art | 3. Expected Visualization | 4. Proposal for a Visual Browsing Method | 5. Perspectives
6
3. Objective: Expected Visualization
extract the structure overview the structure focus on a structure element content (local
view) propose navigation paths through the structure
1. Context and Problem Setting | 2. State of the Art | 3. Expected Visualization | 4. Proposal for a Visual Browsing Method | 5. Perspectives
structure
semantic distance
7
overview : collection of formal concepts
local view : focus on a particular concept :intent = { plasma display, display panel }
1. Context and Problem Setting | 2. State of the Art | 3. Expected Visualization | 4. Proposal for a Visual Browsing Method | 5. Perspectives
3. Objective: Expected Visualization
8
4. Proposal for a Visual Browsing Method
Boolean features
numerical features
plasma_display
display_panel
plasma
p1 = x1
p2 = x2
p0 = x0
document indexation vector in raw data
document x =
1. Context and Problem Setting | 2. State of the Art | 3. Expected Visualization | 4. Proposal for a Visual Browsing Method | 5. Perspectives
9
document / term matrixresulting lattice
overview : collection of clustersi.e. the concept lattice
local view : one cluster’s content(i.e. one formal concept’s extent)
selectionby user
document / documentdistance matrix
raw data
Boolean featuresnumerical features
4. Proposal for a Visual Browsing Method
1. Context and Problem Setting | 2. State of the Art | 3. Expected Visualization | 4. Proposal for a Visual Browsing Method | 5. Perspectives
10
1. Context and Problem Setting | 2. State of the Art | 3. Expected Visualization | 4. Proposal for a Visual Browsing Method | 5. Perspectives
4.1 Lattice spatialization
Computation of a distance matrix thanks to the distance described in [Ranwez 2006]
Projection using Force Direct Placement method with Molage
11
4.1 Lattice spatialization
Neighbours emphasized when selecting a concept Suggesting further navigation paths
Intent and simplified extent cardinal displayed
1. Context and Problem Setting | 2. State of the Art | 3. Expected Visualization | 4. Proposal for a Visual Browsing Method | 5. Perspectives
12
4.2 Document spatialization
FDP using a single distance matrix for all local views
Distance independent from intent’s features
Documents appear and disappear at the same position during navigation
d(o1,o2)
document / documentdistance matrix
1. Context and Problem Setting | 2. State of the Art | 3. Expected Visualization | 4. Proposal for a Visual Browsing Method | 5. Perspectives
13
329 patents indexed by 10 terms
0.8 average term per patent
14
plasma_display
15
plasma_display
plasma_displaydisplay_panel
16
plasma_displaydisplay_panel
17
18
plasma_displaydisplay_paneldisplay_device
19
plasma_displaydisplay_device
20
Goals Solutions
extract the structureformal concept extraction from indexed documents
overview the structuredisplay the lattice with a semantic distance on edges
focus on a structure element content (local view)
display documents in a concept intent with respect to a semantic distance on their numerical features
propose navigation paths through the structure
emphasize current concept’s neighbours on the overview
4.3 Synthesis
1. Context and Problem Setting | 2. State of the Art | 3. Expected Visualization | 4. Proposal for a Visual Browsing Method | 5. Perspectives
21
5. Discussion and perspectives
Prototype and test on greater data Improve lattice distance Visual Assistance for Feature Selection
1. Context and Problem Setting | 2. State of the Art | 3. Expected Visualization | 4. Proposal for a Visual Browsing Method | 5. Perspectives
Using Concept Lattices for Visual Navigation Assistance in Large
Databases
Application to a Patent Database
[email protected]@ema.fr
[email protected]@i-nova.fr
École des Mines d’Alès – http://www.ema.frI-Nova Company, Villeurbanne – http://www.i-nova.fr