Abstract: In this paper, we propose an approach of unsupervised
segmentation with fuzzy connectedness. Valid seeds are first specified
by an unsupervised method based on scale space theory. A region is
then extracted for each seed with a relative object extraction method of
fuzzy connectedness. Afterwards, regions are merged according to the
values between them of an introduced measure. Some theorems and
propositions are also provided to show the reasonableness of the
measure for doing mergence. Experiment results on a synthetic image,
a color image and a large amount of MR images of our method are
reported.
Abstract: We report on the results of a pilot study in which a data-mining tool was developed for mining audiology records. The records were heterogeneous in that they contained numeric, category and textual data. The tools developed are designed to observe associations between any field in the records and any other field. The techniques employed were the statistical chi-squared test, and the use of self-organizing maps, an unsupervised neural learning approach.
Abstract: Named Entity Recognition (NER) aims to classify each word of a document into predefined target named entity classes and is now-a-days considered to be fundamental for many Natural Language Processing (NLP) tasks such as information retrieval, machine translation, information extraction, question answering systems and others. This paper reports about the development of a NER system for Bengali and Hindi using Support Vector Machine (SVM). Though this state of the art machine learning technique has been widely applied to NER in several well-studied languages, the use of this technique to Indian languages (ILs) is very new. The system makes use of the different contextual information of the words along with the variety of features that are helpful in predicting the four different named (NE) classes, such as Person name, Location name, Organization name and Miscellaneous name. We have used the annotated corpora of 122,467 tokens of Bengali and 502,974 tokens of Hindi tagged with the twelve different NE classes 1, defined as part of the IJCNLP-08 NER Shared Task for South and South East Asian Languages (SSEAL) 2. In addition, we have manually annotated 150K wordforms of the Bengali news corpus, developed from the web-archive of a leading Bengali newspaper. We have also developed an unsupervised algorithm in order to generate the lexical context patterns from a part of the unlabeled Bengali news corpus. Lexical patterns have been used as the features of SVM in order to improve the system performance. The NER system has been tested with the gold standard test sets of 35K, and 60K tokens for Bengali, and Hindi, respectively. Evaluation results have demonstrated the recall, precision, and f-score values of 88.61%, 80.12%, and 84.15%, respectively, for Bengali and 80.23%, 74.34%, and 77.17%, respectively, for Hindi. Results show the improvement in the f-score by 5.13% with the use of context patterns. Statistical analysis, ANOVA is also performed to compare the performance of the proposed NER system with that of the existing HMM based system for both the languages.
Abstract: Intelligent systems based on machine learning
techniques, such as classification, clustering, are gaining wide spread
popularity in real world applications. This paper presents work on
developing a software system for predicting crop yield, for example
oil-palm yield, from climate and plantation data. At the core of our
system is a method for unsupervised partitioning of data for finding
spatio-temporal patterns in climate data using kernel methods which
offer strength to deal with complex data. This work gets inspiration
from the notion that a non-linear data transformation into some high
dimensional feature space increases the possibility of linear
separability of the patterns in the transformed space. Therefore, it
simplifies exploration of the associated structure in the data. Kernel
methods implicitly perform a non-linear mapping of the input data
into a high dimensional feature space by replacing the inner products
with an appropriate positive definite function. In this paper we
present a robust weighted kernel k-means algorithm incorporating
spatial constraints for clustering the data. The proposed algorithm
can effectively handle noise, outliers and auto-correlation in the
spatial data, for effective and efficient data analysis by exploring
patterns and structures in the data, and thus can be used for
predicting oil-palm yield by analyzing various factors affecting the
yield.
Abstract: Data mining uses a variety of techniques each of which is useful for some particular task. It is important to have a deep understanding of each technique and be able to perform sophisticated analysis. In this article we describe a tool built to simulate a variation of the Kohonen network to perform unsupervised clustering and support the entire data mining process up to results visualization. A graphical representation helps the user to find out a strategy to optmize classification by adding, moving or delete a neuron in order to change the number of classes. The tool is also able to automatically suggest a strategy for number of classes optimization.The tool is used to classify macroeconomic data that report the most developed countries? import and export. It is possible to classify the countries based on their economic behaviour and use an ad hoc tool to characterize the commercial behaviour of a country in a selected class from the analysis of positive and negative features that contribute to classes formation.
Abstract: In the recent past Learning Classifier Systems have
been successfully used for data mining. Learning Classifier System
(LCS) is basically a machine learning technique which combines
evolutionary computing, reinforcement learning, supervised or
unsupervised learning and heuristics to produce adaptive systems. A
LCS learns by interacting with an environment from which it
receives feedback in the form of numerical reward. Learning is
achieved by trying to maximize the amount of reward received. All
LCSs models more or less, comprise four main components; a finite
population of condition–action rules, called classifiers; the
performance component, which governs the interaction with the
environment; the credit assignment component, which distributes the
reward received from the environment to the classifiers accountable
for the rewards obtained; the discovery component, which is
responsible for discovering better rules and improving existing ones
through a genetic algorithm. The concatenate of the production rules
in the LCS form the genotype, and therefore the GA should operate
on a population of classifier systems. This approach is known as the
'Pittsburgh' Classifier Systems. Other LCS that perform their GA at
the rule level within a population are known as 'Mitchigan' Classifier
Systems. The most predominant representation of the discovered
knowledge is the standard production rules (PRs) in the form of IF P
THEN D. The PRs, however, are unable to handle exceptions and do
not exhibit variable precision. The Censored Production Rules
(CPRs), an extension of PRs, were proposed by Michalski and
Winston that exhibit variable precision and supports an efficient
mechanism for handling exceptions. A CPR is an augmented
production rule of the form: IF P THEN D UNLESS C, where
Censor C is an exception to the rule. Such rules are employed in
situations, in which conditional statement IF P THEN D holds
frequently and the assertion C holds rarely. By using a rule of this
type we are free to ignore the exception conditions, when the
resources needed to establish its presence are tight or there is simply
no information available as to whether it holds or not. Thus, the IF P
THEN D part of CPR expresses important information, while the
UNLESS C part acts only as a switch and changes the polarity of D
to ~D. In this paper Pittsburgh style LCSs approach is used for
automated discovery of CPRs. An appropriate encoding scheme is
suggested to represent a chromosome consisting of fixed size set of
CPRs. Suitable genetic operators are designed for the set of CPRs
and individual CPRs and also appropriate fitness function is proposed
that incorporates basic constraints on CPR. Experimental results are
presented to demonstrate the performance of the proposed learning
classifier system.
Abstract: It is important problems to increase the detection rates
and reduce false positive rates in Intrusion Detection System (IDS).
Although preventative techniques such as access control and
authentication attempt to prevent intruders, these can fail, and as a
second line of defence, intrusion detection has been introduced. Rare
events are events that occur very infrequently, detection of rare
events is a common problem in many domains. In this paper we
propose an intrusion detection method that combines Rough set and
Fuzzy Clustering. Rough set has to decrease the amount of data and
get rid of redundancy. Fuzzy c-means clustering allow objects to
belong to several clusters simultaneously, with different degrees of
membership. Our approach allows us to recognize not only known
attacks but also to detect suspicious activity that may be the result of
a new, unknown attack. The experimental results on Knowledge
Discovery and Data Mining-(KDDCup 1999) Dataset show that the
method is efficient and practical for intrusion detection systems.
Abstract: Emerging Bio-engineering fields such as Brain
Computer Interfaces, neuroprothesis devices and modeling and
simulation of neural networks have led to increased research activity
in algorithms for the detection, isolation and classification of Action
Potentials (AP) from noisy data trains. Current techniques in the field
of 'unsupervised no-prior knowledge' biosignal processing include
energy operators, wavelet detection and adaptive thresholding. These
tend to bias towards larger AP waveforms, AP may be missed due to
deviations in spike shape and frequency and correlated noise
spectrums can cause false detection. Also, such algorithms tend to
suffer from large computational expense.
A new signal detection technique based upon the ideas of phasespace
diagrams and trajectories is proposed based upon the use of a
delayed copy of the AP to highlight discontinuities relative to
background noise. This idea has been used to create algorithms that
are computationally inexpensive and address the above problems.
Distinct AP have been picked out and manually classified from
real physiological data recorded from a cockroach. To facilitate
testing of the new technique, an Auto Regressive Moving Average
(ARMA) noise model has been constructed bases upon background
noise of the recordings. Along with the AP classification means this
model enables generation of realistic neuronal data sets at arbitrary
signal to noise ratio (SNR).
Abstract: In unsupervised segmentation context, we propose a bi-dimensional hidden Markov chain model (X,Y) that we adapt to the image segmentation problem. The bi-dimensional observed process Y = (Y 1, Y 2) is such that Y 1 represents the noisy image and Y 2 represents a noisy supplementary information on the image, for example a noisy proportion of pixels of the same type in a neighborhood of the current pixel. The proposed model can be seen as a competitive alternative to the Hilbert-Peano scan. We propose a bayesian algorithm to estimate parameters of the considered model. The performance of this algorithm is globally favorable, compared to the bi-dimensional EM algorithm through numerical and visual data.
Abstract: Most paddy rice fields in East Asia are small parcels,
and the weather conditions during the growing season are usually
cloudy. FORMOSAT-2 multi-spectral images have an 8-meter
resolution and one-day recurrence, ideal for mapping paddy rice fields
in East Asia. To map rice fields, this study first determined the
transplanting and the most active tillering stages of paddy rice and
then used multi-temporal images to distinguish different growing
characteristics between paddy rice and other ground covers. The
unsupervised ISODATA (iterative self-organizing data analysis
techniques) and supervised maximum likelihood were both used to
discriminate paddy rice fields, with training areas automatically
derived from ten-year cultivation parcels in Taiwan. Besides original
bands in multi-spectral images, we also generated normalized
difference vegetation index and experimented with object-based
pre-classification and post-classification. This paper discusses results
of different image classification methods in an attempt to find a
precise and automatic solution to mapping paddy rice in Taiwan.
Abstract: The clustering ensembles combine multiple partitions
generated by different clustering algorithms into a single clustering
solution. Clustering ensembles have emerged as a prominent method
for improving robustness, stability and accuracy of unsupervised
classification solutions. So far, many contributions have been done to
find consensus clustering. One of the major problems in clustering
ensembles is the consensus function. In this paper, firstly, we
introduce clustering ensembles, representation of multiple partitions,
its challenges and present taxonomy of combination algorithms.
Secondly, we describe consensus functions in clustering ensembles
including Hypergraph partitioning, Voting approach, Mutual
information, Co-association based functions and Finite mixture
model, and next explain their advantages, disadvantages and
computational complexity. Finally, we compare the characteristics of
clustering ensembles algorithms such as computational complexity,
robustness, simplicity and accuracy on different datasets in previous
techniques.
Abstract: This paper introduces new algorithms (Fuzzy relative
of the CLARANS algorithm FCLARANS and Fuzzy c Medoids
based on randomized search FCMRANS) for fuzzy clustering of
relational data. Unlike existing fuzzy c-medoids algorithm (FCMdd)
in which the within cluster dissimilarity of each cluster is minimized
in each iteration by recomputing new medoids given current
memberships, FCLARANS minimizes the same objective function
minimized by FCMdd by changing current medoids in such away
that that the sum of the within cluster dissimilarities is minimized.
Computing new medoids may be effected by noise because outliers
may join the computation of medoids while the choice of medoids in
FCLARANS is dictated by the location of a predominant fraction of
points inside a cluster and, therefore, it is less sensitive to the
presence of outliers. In FCMRANS the step of computing new
medoids in FCMdd is modified to be based on randomized search.
Furthermore, a new initialization procedure is developed that add
randomness to the initialization procedure used with FCMdd. Both
FCLARANS and FCMRANS are compared with the robust and
linearized version of fuzzy c-medoids (RFCMdd). Experimental
results with different samples of the Reuter-21578, Newsgroups
(20NG) and generated datasets with noise show that FCLARANS is
more robust than both RFCMdd and FCMRANS. Finally, both
FCMRANS and FCLARANS are more efficient and their outputs
are almost the same as that of RFCMdd in terms of classification
rate.
Abstract: Face Recognition has always been a fascinating research area. It has drawn the attention of many researchers because of its various potential applications such as security systems, entertainment, criminal identification etc. Many supervised and unsupervised learning techniques have been reported so far. Principal Component Analysis (PCA), Self Organizing Maps (SOM) and Independent Component Analysis (ICA) are the three techniques among many others as proposed by different researchers for Face Recognition, known as the unsupervised techniques. This paper proposes integration of the two techniques, SOM and PCA, for dimensionality reduction and feature selection. Simulation results show that, though, the individual techniques SOM and PCA itself give excellent performance but the combination of these two can also be utilized for face recognition. Experimental results also indicate that for the given face database and the classifier used, SOM performs better as compared to other unsupervised learning techniques. A comparison of two proposed methodologies of SOM, Local and Global processing, shows the superiority of the later but at the cost of more computational time.