Transcript Learning from Imbalanced Data Sets with Boosting and Data

國立雲林科技大學
National Yunlin University of Science and Technology
Adaptive self-organized maps
based on bidirectional approximate reasoning and
its applications to information filtering
Yu Liu
Knowledge-Based Systems, Vol. 19, 2006, pp. 719–729.
Presenter : Wei-Shen Tai
Advisor : Professor Chung-Chian Hsu
2006/11/29
N.Y.U.S.T.
I. M.
Outline







Introduction
BAR inference network and fuzzy similarity
distance
ASOMBAR
ASOMBAR algorithm for information filtering
Information filtering example and results
Conclusion
Comments
N.Y.U.S.T.
I. M.
Motivation

Similarity measuring in SOM

The Euclidean distance supposes that all the
factors are in the same level of importance

The distortion of the large weighted factors
should have more influence on the similarity
measuring than the small weighted ones do.
N.Y.U.S.T.
I. M.
Objective

A novel fuzzy similarity distance

Replace the Euclidean distance to reflect the
real relation more precisely.

Improve the effectiveness of the competitive
process and the cooperative process.
BAR inference network and
fuzzy similarity distance

Similarity measure of BAR

~
~
If any variable X is similar to Xj, then the output Y of the
inference network is to be similar to Yj.


Solves the problem of different pairs input and output data
often include inconsistencies among them in real situations.
Ordered weighted averaging (OWA)

Combines fuzzy membership values in constraint satisfaction
problems, scheduling and group decision making.



Biggest factor
Smallest factor
where λ decides how much the big factor and other smaller factors
in one input vector influence the proposed fuzzy distance.
N.Y.U.S.T.
I. M.
N.Y.U.S.T.
I. M.
ASOMBAR (Adaptive SOMBAR)

Competition:


Cooperation:


The particular neuron with the minimum fuzzy similarity distance
is the declared winner of the competition.
The winning vector determines the spatial location of a topological
neighborhood of excited neurons.
Synaptic adaptation:

Enables the excited neurons to increase their individual values
through suitable adjustments applied to their synaptic weights.
ASOMBAR algorithm for
information filtering


N.Y.U.S.T.
I. M.
Information filtering
 Separating relevant documents and
removing irrelevant documents to meet
the user’s current interest.
ASOMBAR
 Filters available information through the
similarity degrees between the documents
and user profile.
N.Y.U.S.T.
I. M.
Information filtering example

User profile


Described by those high-frequency important words
related to the user’s interest.
Parameter λ


Decides how much the big factor and other factors in
one input vector impact.
Influences the convergence of the distortion errors.
N.Y.U.S.T.
I. M.
Results

ASOMBAR in λ ≧ 0.5


More effective than both GNG and basic SOM.
Topology neighborhoods


ASOMBAR and SOM seem very alike.
ASOMBAR based on fuzzy similarity distance changes
more smoothly than that of SOM does.
N.Y.U.S.T.
I. M.
Conclusions

A new fuzzy similarity distance



Replaces the Euclidean distance inspired by
BAR and OWA theory.
Has paid more attention on the large weighted
element and less emphasis on the small
weighted elements.
Information filtering

Applied to constructs user’s profile and
processes document filtering.
N.Y.U.S.T.
I. M.
Comments

Advantage



Drawback



The idea of non-Euclidean similarity measurement in SOM.
Ordered weighted averaging (OWA) in the similarity
measurement.
Effectiveness can not be improved significantly.
BAR concept is not applied in this study appropriately.
Application

Information filter and applications related to similarity
measurement.