graph based pattern recognition
DESCRIPTION
Brief introduction to graph based pattern recognition. It shows advantages and disantavantages of using graphs and how existing pattern recognition techniques are adapted to graph space.TRANSCRIPT
page.1Introduction Graph-based Methods Conclusions
Graph-based Pattern Recognition
Nicola Strisciuglio
University of Groningen
28/10/2013
page.2Introduction Graph-based Methods Conclusions
Statistical vs. Graph-based PR
Statistical vs. Graph-based Pattern Recognition
Statistical PR
I =(x1, x2, . . . , xn
) Graph-based PR
page.3Introduction Graph-based Methods Conclusions
Statistical vs. Graph-based PR
Statistical vs. Graph-based Pattern Recognition
Statistical PR
I =(y1, y2, . . . , yn
) Graph-based PR
page.4Introduction Graph-based Methods Conclusions
Pure Impure and Extreme Methods
Graph-based Methods
Pure MethodsLearning and classification problems are faced directly in the graphspace.
Impure Methods
Transposition of the methods of Statistical Pattern Recognition tograph space.
Extreme MethodsTransformation of graphs into vectors. Use of the well estabilishedlearning and classification techniques.
page.5Introduction Graph-based Methods Conclusions
Pure Impure and Extreme Methods
Pure Methods: Graph Matching
I Exact Matching: find an exact correspondence betweengraphs (or sub-graphs)
I Inexact Matching: deals with distortions in findinf acorrespondence between graphs
I It needs a metric to define dissimilarities
page.6Introduction Graph-based Methods Conclusions
Pure Impure and Extreme Methods
Graph edit distance
We need a distance metric between graphs: we have Graph editdistance. Cheapest sequence of operations to trasform one graphin another graph.
page.7Introduction Graph-based Methods Conclusions
Pure Impure and Extreme Methods
Impure Methods: k-NN
I The reference set is made by a collection of graphs, instead ofpoints in a m-dimensional space
I For a graph to classify, the graph edit distance is computedwith respect to each of the graphs in the reference set
I The decision is taken as majority voting on the K nearestgraphs
I Generally, the time needed for a classification is very high
page.8Introduction Graph-based Methods Conclusions
Pure Impure and Extreme Methods
Impure Methods: k-Means
For the classical k-Means algorithm a centroid is computed as theaverage of the vectors around it, while for the graph a mediangraph should be computed.
Median Graph
S = arg minGi∈S
∑Gj∈S
Dg (Gi ,Gj)
Generalized Median Graph
S = arg minGi∈U
∑Gj∈S
Dg (Gi ,Gj)
page.9Introduction Graph-based Methods Conclusions
Pure Impure and Extreme Methods
Impure Methods: k-Means
For the classical k-Means algorithm a centroid is computed as theaverage of the vectors around it, while for the graph a mediangraph should be computed.
Median Graph
S = arg minGi∈S
∑Gj∈S
Dg (Gi ,Gj)
Generalized Median Graph
S = arg minGi∈U
∑Gj∈S
Dg (Gi ,Gj)
page.10Introduction Graph-based Methods Conclusions
Pure Impure and Extreme Methods
Impure Methods: k-Means
For the classical k-Means algorithm a centroid is computed as theaverage of the vectors around it, while for the graph a mediangraph should be computed.
Median Graph
S = arg minGi∈S
∑Gj∈S
Dg (Gi ,Gj)
Generalized Median Graph
S = arg minGi∈U
∑Gj∈S
Dg (Gi ,Gj)
page.11Introduction Graph-based Methods Conclusions
Pure Impure and Extreme Methods
Impure Methods: LVQ
I Given a pattern x , the winner prototype Pk moves towards xby ∆ = α(x − Pk) if the class of x is the same of Pk (by∆ = −α(x − Pk) otherwise)
I Updating a prototype requires its transformation in anothergraph more similar to the input pattern x by a fraction α ofthe distance
I We need to compute a fraction of edit distance!!!I Pk → Gx needs D = 7 operations on graphI If α = 0.3, we do bD ∗ αc = 2 operations to move Pk to an
intermediate graph
page.12Introduction Graph-based Methods Conclusions
Pure Impure and Extreme Methods
Impure Methods: LVQ
I Given a pattern x , the winner prototype Pk moves towards xby ∆ = α(x − Pk) if the class of x is the same of Pk (by∆ = −α(x − Pk) otherwise)
I Updating a prototype requires its transformation in anothergraph more similar to the input pattern x by a fraction α ofthe distance
I We need to compute a fraction of edit distance!!!I Pk → Gx needs D = 7 operations on graph
I If α = 0.3, we do bD ∗ αc = 2 operations to move Pk to anintermediate graph
page.13Introduction Graph-based Methods Conclusions
Pure Impure and Extreme Methods
Impure Methods: LVQ
I Given a pattern x , the winner prototype Pk moves towards xby ∆ = α(x − Pk) if the class of x is the same of Pk (by∆ = −α(x − Pk) otherwise)
I Updating a prototype requires its transformation in anothergraph more similar to the input pattern x by a fraction α ofthe distance
I We need to compute a fraction of edit distance!!!I Pk → Gx needs D = 7 operations on graphI If α = 0.3, we do bD ∗ αc = 2 operations to move Pk to an
intermediate graph
page.14Introduction Graph-based Methods Conclusions
Pure Impure and Extreme Methods
Extreme methods: Graph Embedding
I Represent a graph as a point in a suitable feature space
I Use of the classical statistical pattern recognition tools
I The similarity of graph in graph space should be preserved inthe vector space
I The translation of a graph (including all the attributes andrelations) into a fixed-lenght vector is difficult
page.15Introduction Graph-based Methods Conclusions
Conclusions
Some Applications
I Structural description and matching of moleculesI Segmentation of shapes (e.g. letters)I Stereo vision (e.g. for robot navigation)I Learning and recognising objects in scenesI Data mining
page.16Introduction Graph-based Methods Conclusions
Conclusions
Some Applications
I Structural description and matching of moleculesI Segmentation of shapes (e.g. letters)I Stereo vision (e.g. for robot navigation)I Learning and recognising objects in scenesI Data mining
page.17Introduction Graph-based Methods Conclusions
Conclusions
Statistical vs. Graph-based Pattern Recognition
Statistical PRAdvantages
I Theoretically estabilished
I Many powerful algorithms
Disadvantages
I Size of the feature vectorfixed
I Unary features: no relations
Graph-based PRAdvantages
I Variable representation size
I More description power(relationships)
Disadvantages
I Lack of algorithms
I Less mathematicalfoundations
page.18Introduction Graph-based Methods Conclusions
Conclusions
References
Mario Vento (2013)
A long trip in the charming world of graphs for Pattern Recognition
Pattern Recognition
page.19Introduction Graph-based Methods Conclusions
Conclusions
Thank you!!!