Abstract: Bidding is a very important business function to find
latent contractors of construction projects. Moreover, bid markup is
one of the most important decisions for a bidder to gain a reasonable
profit. Since the bidding system is a complex adaptive system, bidding
agent need a learning process to get more valuable knowledge for a bid,
especially from past public bidding information. In this paper, we
proposed an iterative agent leaning model for bidders to make markup
decisions. A classifier for public bidding information named PIBS is
developed to make full use of history data for classifying new bidding
information. The simulation and experimental study is performed to
show the validity of the proposed classifier. Some factors that affect
the validity of PIBS are also analyzed at the end of this work.
Abstract: Understanding proteins functions is a major goal in
the post-genomic era. Proteins usually work in context of other
proteins and rarely function alone. Therefore, it is highly relevant to
study the interaction partners of a protein in order to understand its
function. Machine learning techniques have been widely applied to
predict protein-protein interactions. Kernel functions play an
important role for a successful machine learning technique. Choosing
the appropriate kernel function can lead to a better accuracy in a
binary classifier such as the support vector machines. In this paper,
we describe a Bayesian kernel for the support vector machine to
predict protein-protein interactions. The use of Bayesian kernel can
improve the classifier performance by incorporating the probability
characteristic of the available experimental protein-protein
interactions data that were compiled from different sources. In
addition, the probabilistic output from the Bayesian kernel can assist
biologists to conduct more research on the highly predicted
interactions. The results show that the accuracy of the classifier has
been improved using the Bayesian kernel compared to the standard
SVM kernels. These results imply that protein-protein interaction can
be predicted using Bayesian kernel with better accuracy compared to
the standard SVM kernels.
Abstract: This paper describes a new supervised fusion (hybrid)
electrocardiogram (ECG) classification solution consisting of a new
QRS complex geometrical feature extraction as well as a new version
of the learning vector quantization (LVQ) classification algorithm
aimed for overcoming the stability-plasticity dilemma. Toward this
objective, after detection and delineation of the major events of ECG
signal via an appropriate algorithm, each QRS region and also its
corresponding discrete wavelet transform (DWT) are supposed as
virtual images and each of them is divided into eight polar sectors.
Then, the curve length of each excerpted segment is calculated
and is used as the element of the feature space. To increase the
robustness of the proposed classification algorithm versus noise,
artifacts and arrhythmic outliers, a fusion structure consisting of
five different classifiers namely as Support Vector Machine (SVM),
Modified Learning Vector Quantization (MLVQ) and three Multi
Layer Perceptron-Back Propagation (MLP–BP) neural networks with
different topologies were designed and implemented. The new proposed
algorithm was applied to all 48 MIT–BIH Arrhythmia Database
records (within–record analysis) and the discrimination power of the
classifier in isolation of different beat types of each record was
assessed and as the result, the average accuracy value Acc=98.51%
was obtained. Also, the proposed method was applied to 6 number
of arrhythmias (Normal, LBBB, RBBB, PVC, APB, PB) belonging
to 20 different records of the aforementioned database (between–
record analysis) and the average value of Acc=95.6% was achieved.
To evaluate performance quality of the new proposed hybrid learning
machine, the obtained results were compared with similar peer–
reviewed studies in this area.
Abstract: Prediction of bacterial virulent protein sequences can
give assistance to identification and characterization of novel
virulence-associated factors and discover drug/vaccine targets against
proteins indispensable to pathogenicity. Gene Ontology (GO)
annotation which describes functions of genes and gene products as a
controlled vocabulary of terms has been shown effectively for a
variety of tasks such as gene expression study, GO annotation
prediction, protein subcellular localization, etc. In this study, we
propose a sequence-based method Virulent-GO by mining informative
GO terms as features for predicting bacterial virulent proteins.
Each protein in the datasets used by the existing method
VirulentPred is annotated by using BLAST to obtain its homologies
with known accession numbers for retrieving GO terms. After
investigating various popular classifiers using the same five-fold
cross-validation scheme, Virulent-GO using the single kind of GO
term features with an accuracy of 82.5% is slightly better than
VirulentPred with 81.8% using five kinds of sequence-based features.
For the evaluation of independent test, Virulent-GO also yields better
results (82.0%) than VirulentPred (80.7%). When evaluating single
kind of feature with SVM, the GO term feature performs much well,
compared with each of the five kinds of features.
Abstract: A state of the art Speaker Identification (SI) system requires a robust feature extraction unit followed by a speaker modeling scheme for generalized representation of these features. Over the years, Mel-Frequency Cepstral Coefficients (MFCC) modeled on the human auditory system has been used as a standard acoustic feature set for SI applications. However, due to the structure of its filter bank, it captures vocal tract characteristics more effectively in the lower frequency regions. This paper proposes a new set of features using a complementary filter bank structure which improves distinguishability of speaker specific cues present in the higher frequency zone. Unlike high level features that are difficult to extract, the proposed feature set involves little computational burden during the extraction process. When combined with MFCC via a parallel implementation of speaker models, the proposed feature set outperforms baseline MFCC significantly. This proposition is validated by experiments conducted on two different kinds of public databases namely YOHO (microphone speech) and POLYCOST (telephone speech) with Gaussian Mixture Models (GMM) as a Classifier for various model orders.
Abstract: Corporate credit rating prediction using statistical and
artificial intelligence (AI) techniques has been one of the attractive
research topics in the literature. In recent years, multiclass
classification models such as artificial neural network (ANN) or
multiclass support vector machine (MSVM) have become a very
appealing machine learning approaches due to their good
performance. However, most of them have only focused on classifying
samples into nominal categories, thus the unique characteristic of the
credit rating - ordinality - has been seldom considered in their
approaches. This study proposes new types of ANN and MSVM
classifiers, which are named OMANN and OMSVM respectively.
OMANN and OMSVM are designed to extend binary ANN or SVM
classifiers by applying ordinal pairwise partitioning (OPP) strategy.
These models can handle ordinal multiple classes efficiently and
effectively. To validate the usefulness of these two models, we applied
them to the real-world bond rating case. We compared the results of
our models to those of conventional approaches. The experimental
results showed that our proposed models improve classification
accuracy in comparison to typical multiclass classification techniques
with the reduced computation resource.
Abstract: Text categorization is the problem of classifying text documents into a set of predefined classes. After a preprocessing step the documents are typically represented as large sparse vectors. When training classifiers on large collections of documents, both the time and memory restrictions can be quite prohibitive. This justifies the application of features selection methods to reduce the dimensionality of the document-representation vector. Four feature selection methods are evaluated: Random Selection, Information Gain (IG), Support Vector Machine (called SVM_FS) and Genetic Algorithm with SVM (GA_FS). We showed that the best results were obtained with SVM_FS and GA_FS methods for a relatively small dimension of the features vector comparative with the IG method that involves longer vectors, for quite similar classification accuracies. Also we present a novel method to better correlate SVM kernel-s parameters (Polynomial or Gaussian kernel).
Abstract: As emails communications have no consistent
authentication procedure to ensure the authenticity, we present an
investigation analysis approach for detecting forged emails based on
Random Forests and Naïve Bays classifiers. Instead of investigating
the email headers, we use the body content to extract a unique writing
style for all the possible suspects. Our approach consists of four main
steps: (1) The cybercrime investigator extract different effective
features including structural, lexical, linguistic, and syntactic
evidence from previous emails for all the possible suspects, (2) The
extracted features vectors are normalized to increase the accuracy
rate. (3) The normalized features are then used to train the learning
engine, (4) upon receiving the anonymous email (M); we apply the
feature extraction process to produce a feature vector. Finally, using
the machine learning classifiers the email is assigned to one of the
suspects- whose writing style closely matches M. Experimental
results on real data sets show the improved performance of the
proposed method and the ability of identifying the authors with a
very limited number of features.
Abstract: One of the most important problems to solve is eye
location for a driver fatigue monitoring system. This paper presents an
efficient method to achieve fast and accurate eye location in grey level
images obtained in the real-word driving conditions. The structure of
eye region is used as a robust cue to find possible eye pairs. Candidates
of eye pair at different scales are selected by finding regions which
roughly match with the binary eye pair template. To obtain real one,
all the eye pair candidates are then verified by using support vector
machines. Finally, eyes are precisely located by using binary vertical
projection and eye classifier in eye pair images. The proposed method
is robust to deal with illumination changes, moderate rotations, glasses
wearing and different eye states. Experimental results demonstrate its
effectiveness.
Abstract: It is sometimes difficult to differentiate between
innocent murmurs and pathological murmurs during auscultation. In
these difficult cases, an intelligent stethoscope with decision support
abilities would be of great value. In this study, using a dog model,
phonocardiographic recordings were obtained from 27 boxer dogs
with various degrees of aortic stenosis (AS) severity. As a reference
for severity assessment, continuous wave Doppler was used. The data
were analyzed with recurrence quantification analysis (RQA) with
the aim to find features able to distinguish innocent murmurs from
murmurs caused by AS. Four out of eight investigated RQA features
showed significant differences between innocent murmurs and
pathological murmurs. Using a plain linear discriminant analysis
classifier, the best pair of features (recurrence rate and entropy)
resulted in a sensitivity of 90% and a specificity of 88%. In
conclusion, RQA provide valid features which can be used for
differentiation between innocent murmurs and murmurs caused by
AS.
Abstract: In the past years a lot of effort has been made in the
field of face detection. The human face contains important features
that can be used by vision-based automated systems in order to
identify and recognize individuals. Face location, the primary step of
the vision-based automated systems, finds the face area in the input
image. An accurate location of the face is still a challenging task.
Viola-Jones framework has been widely used by researchers in order
to detect the location of faces and objects in a given image. Face
detection classifiers are shared by public communities, such as
OpenCV. An evaluation of these classifiers will help researchers to
choose the best classifier for their particular need. This work focuses
of the evaluation of face detection classifiers minding facial
landmarks.
Abstract: This paper describes a method to improve the robustness of a face recognition system based on the combination of two compensating classifiers. The face images are preprocessed by the appearance-based statistical approaches such as Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA). LDA features of the face image are taken as the input of the Radial Basis Function Network (RBFN). The proposed approach has been tested on the ORL database. The experimental results show that the LDA+RBFN algorithm has achieved a recognition rate of 93.5%
Abstract: In this paper, a wavelet-based neural network (WNN) classifier for recognizing EEG signals is implemented and tested under three sets EEG signals (healthy subjects, patients with epilepsy and patients with epileptic syndrome during the seizure). First, the Discrete Wavelet Transform (DWT) with the Multi-Resolution Analysis (MRA) is applied to decompose EEG signal at resolution levels of the components of the EEG signal (δ, θ, α, β and γ) and the Parseval-s theorem are employed to extract the percentage distribution of energy features of the EEG signal at different resolution levels. Second, the neural network (NN) classifies these extracted features to identify the EEGs type according to the percentage distribution of energy features. The performance of the proposed algorithm has been evaluated using in total 300 EEG signals. The results showed that the proposed classifier has the ability of recognizing and classifying EEG signals efficiently.
Abstract: The use of High Order Statistics (HOS) analysis is
expected to provide so many candidates of features that can be selected for pattern recognition. More candidates of the feature can
be extracted using simple manipulation through a specific mathematical function prior to the HOS analysis. Feature extraction
method using HOS analysis combined with Difference to the Nth-Power manipulation has been examined in application for Automatic
Modulation Recognition (AMR) to perform scheme recognition of three digital modulation signal, i.e. QPSK-16QAM-64QAM in the
AWGN transmission channel. The simulation results is reported
when the analysis of HOS up to order-12 and the manipulation of Difference to the Nth-Power up to N = 4. The obtained accuracy rate
of AMR using the method of Simple Decision obtained 90% in SNR > 10 dB in its classifier, while using the method of Voted Decision is
96% in SNR > 2 dB.
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: There are several approaches for handling multiclass classification. Aside from one-against-one (OAO) and one-against-all (OAA), hierarchical classification technique is also commonly used. A binary classification tree is a hierarchical classification structure that breaks down a k-class problem into binary sub-problems, each solved by a binary classifier. In each node, a set of classes is divided into two subsets. A good class partition should be able to group similar classes together. Many algorithms measure similarity in term of distance between class centroids. Classes are grouped together by a clustering algorithm when distances between their centroids are small. In this paper, we present a binary classification tree with tuned observation-based clustering (BCT-TOB) that finds a class partition by performing clustering on observations instead of class centroids. A merging step is introduced to merge any insignificant class split. The experiment shows that performance of BCT-TOB is comparable to other algorithms.
Abstract: This paper describes an optimal approach for feature
subset selection to classify the leaves based on Genetic Algorithm
(GA) and Kernel Based Principle Component Analysis (KPCA). Due
to high complexity in the selection of the optimal features, the
classification has become a critical task to analyse the leaf image
data. Initially the shape, texture and colour features are extracted
from the leaf images. These extracted features are optimized through
the separate functioning of GA and KPCA. This approach performs
an intersection operation over the subsets obtained from the
optimization process. Finally, the most common matching subset is
forwarded to train the Support Vector Machine (SVM). Our
experimental results successfully prove that the application of GA
and KPCA for feature subset selection using SVM as a classifier is
computationally effective and improves the accuracy of the classifier.
Abstract: The purpose of this paper is to demonstrate the ability
of a genetic programming (GP) algorithm to evolve a team of data
classification models. The GP algorithm used in this work is
“multigene" in nature, i.e. there are multiple tree structures (genes)
that are used to represent team members. Each team member assigns
a data sample to one of a fixed set of output classes. A majority vote,
determined using the mode (highest occurrence) of classes predicted
by the individual genes, is used to determine the final class
prediction. The algorithm is tested on a binary classification problem.
For the case study investigated, compact classification models are
obtained with comparable accuracy to alternative approaches.
Abstract: Automatic reading of handwritten cheque is a computationally
complex process and it plays an important role in financial
risk management. Machine vision and learning provide a viable
solution to this problem. Research effort has mostly been focused
on recognizing diverse pitches of cheques and demand drafts with an
identical outline. However most of these methods employ templatematching
to localize the pitches and such schemes could potentially
fail when applied to different types of outline maintained by the
bank. In this paper, the so-called outline problem is resolved by
a cheque information tree (CIT), which generalizes the localizing
method to extract active-region-of-entities. In addition, the weight
based density plot (WBDP) is performed to isolate text entities and
read complete pitches. Recognition is based on texture features using
neural classifiers. Legal amount is subsequently recognized by both
texture and perceptual features. A post-processing phase is invoked
to detect the incorrect readings by Type-2 grammar using the Turing
machine. The performance of the proposed system was evaluated
using cheque and demand drafts of 22 different banks. The test data
consists of a collection of 1540 leafs obtained from 10 different
account holders from each bank. Results show that this approach
can easily be deployed without significant design amendments.
Abstract: This paper presents a new classification algorithm using colour and texture for obstacle detection. Colour information is computationally cheap to learn and process. However in many cases, colour alone does not provide enough information for classification. Texture information can improve classification performance but usually comes at an expensive cost. Our algorithm uses both colour and texture features but texture is only needed when colour is unreliable. During the training stage, texture features are learned specifically to improve the performance of a colour classifier. The algorithm learns a set of simple texture features and only the most effective features are used in the classification stage. Therefore our algorithm has a very good classification rate while is still fast enough to run on a limited computer platform. The proposed algorithm was tested with a challenging outdoor image set. Test result shows the algorithm achieves a much better trade-off between classification performance and efficiency than a typical colour classifier.