Adaptive local dissimilarity measures for discriminative dimension of labeled data

Intelligent Database Systems Lab

Presenter : Kung, Chien-Hao

Authors : Kerstin Bunte, Barbara Hammer, Axel Wismuller,

Michael Biehl

2010,NC

Adaptive local dissimilarity measures for discriminative dimension of labeled data


OutlinesMotivationObjectivesMethodologyExperimentsConclusionsComments


Motivation• Dimension reduction embedding in lower

dimensions necessarily includes a loss of

information.

• To explicitly control the information kept by a

specific dimension reduction technique are

highly desirable.


Objectives• The aim of this paper is to combine an adaptive metric

and recent visualization techniques towards a

discriminative approach.


Methodology

LiRaM LVQ

Stochastic neighbor embedding

(SNE)

Exploration observation machine

(XOM)

Maximum variance unfolding

(MVU)

Charting

Locally linear embedding

(LLE)

Isomap


LiRaM LVQ– Prototype based classifier, extension of LVQ– Modified Euclidean distance:

– Adapt local matrices during training(minimize a cost function by gradient descent)

Methodology


Combination of local linear patches by charting• The charting technique can decompose the sample data

into locally linear patches and combine them into a single low-dimensional coordinate system.

Methodology


Locally linear embedding (LLE)• Locally linear embedding (LLE) uses the criterion of

topology preservation for dimension reduction.

Methodology


Isomap• Isomap is an extension of metric Multi-Dimensional

Scaling(MDS) which uses distance preservation as criterion

Methodology


Stochastic neighbor embedding (SNE)• Stochastic neighbor embedding (SNE) constitutes an

unsupervised projection which follows a probability based approach.

Methodology


Exploration observation machine(XOM)• The exploratory observation machine (XOM) has

recently been introduced as a novel computational framework for structure-preserving dimension reduction.

Methodology


Maximum variance unfolding(MVU)• Maximum variance unfolding(MVU) is a dimension

reduction technique which aims at preservation if local distances.

Methodology


Experiments


ExperimentsThree tip star data set


ExperimentsWine data set


ExperimentsSegmentationdata set


ExperimentsUSPSdata set


Experiments


Conclusions• The results are quite diverse and no single method

which is optimum for every case an be identified.

• In general, discriminative visualization as introduced

in this paper improves all the corresponding

unsupervised methods.


Comments• Advantages– This paper is easy to read.

• Applications– Dimension reduction

Adaptive local dissimilarity measures for discriminative dimension of labeled data

Documents