Ontology Based Dimensionality Reduction

SS
by Shiladitya Sen4 views·Updated May 7, 2015
Powered by AI

Need to present on this topic?

Create your own polished presentation in minutes. Just describe what you need.

Free to try, no credit card required

Transcript

Hughes effect in Machine Learning

Fixed examples, increasing features

Leads to poor learning

Model too complex to be

learnt with fewer features

Ontology?

Graph structure defining particular hierarchy among different “concepts” in a domain.

nodes : concepts.

Edges exist between two nodes if subconcept relation exists

3rd generation ML

Incorporating knowledge into black box

Idea: Use Ontology to guide Dimensionality

Reduction

Future Work

Hotho's Method

Natural Language Domain

#Concepts < #Terms

Map Terms to Concepts in Ontology

Autoencoders

Replace term vector by concept vector

Restricted Boltzmann Machines

Ontology Based

Dimensionality Reduction

Dimensionality Reduction achieved

Use other two ontologies

Form ensemble for

Dimensionality Reduction

Ontologies share knowledge

Complement each other

One-time reduction

Dependent on Ontology

Curse of

Dimensionality

Semantic

Analysis

Hierarchical

Feature

Reduction

Under the guidance of:

Prof. Jaya Sil

Department of CST

IIEST, Shibpur

Shiladitya Sen

Department of CST

IIEST, Shibpur.

ID No.: 110511036

The Proposed Approach

Information : Entropy

K-Means Clustering

Dimensionality Reduction Algorithm

Cluster Validation

Non-homogeneity -> more information

Silhouette coefficient

Principle Components Analysis

s(i) = [b(i) - a(i)]/max{b(i),a(i)}

Correlation between data should be less

PCA(X) returns first k eigenvectors of cov(X)

First component max variance

-1 <= avg(s(i)) <= 1

Eigenvectors orthogonal

{7,8,9} : C

{6,7} : B

{1,2,3,4,5,6} : A

{1,2,3,4,5,6,7,8,9} - {A,B,C}

First gen features - 3775

Siblings are more similar than cousins!

1st parent gen features - 2493

32% reduction

1st parent gen: 0.25

2nd parent gen

features - 1182

69% reduction

tf-idf : 0.09

NMF : 0.19

2nd parent gen : 0.33

Gene Ontology -

Cellular Components

Molecular Function

Bio Process

UniProt - Bacteria proteome data

3,830 proteomes

Mapped proteins to properties

or concepts in ontology

Counted frequency of occurrence for concept in proteome

Resembles docs - tf*idf

NMF: P = QxR

P - n x m : Q - n x k : R - k x m

k latent features

data_C : 3830 x 243

data_F : 3830 x 3775

data_P : 3830 x 2543

Questions?

Thank You!

1

Describe the subject's uses.