Abstract: In this paper, we propose an optimized brain computer
interface (BCI) system for unspoken speech recognition, based on
the fact that the constructions of unspoken words rely strongly on the
Wernicke area, situated in the temporal lobe. Our BCI system has four
modules: (i) the EEG Acquisition module based on a non-invasive
headset with 14 electrodes; (ii) the Preprocessing module to remove
noise and artifacts, using the Common Average Reference method;
(iii) the Features Extraction module, using Wavelet Packet Transform
(WPT); (iv) the Classification module based on a one-hidden layer
artificial neural network. The present study consists of comparing
the recognition accuracy of 5 Arabic words, when using all the
headset electrodes or only the 4 electrodes situated near the Wernicke
area, as well as the selection effect of the subbands produced by
the WPT module. After applying the articial neural network on the
produced database, we obtain, on the test dataset, an accuracy of
83.4% with all the electrodes and all the subbands of 8 levels of the
WPT decomposition. However, by using only the 4 electrodes near
Wernicke Area and the 6 middle subbands of the WPT, we obtain
a high reduction of the dataset size, equal to approximately 19% of
the total dataset, with 67.5% of accuracy rate. This reduction appears
particularly important to improve the design of a low cost and simple
to use BCI, trained for several words.
Abstract: In this paper, we study the factors which determine the capacity of a Convolutional Neural Network (CNN) model and propose the ways to evaluate and adjust the capacity of a CNN model for best matching to a specific pattern recognition task. Firstly, a scheme is proposed to adjust the number of independent functional units within a CNN model to make it be better fitted to a task. Secondly, the number of independent functional units in the capsule network is adjusted to fit it to the training dataset. Thirdly, a method based on Bayesian GAN is proposed to enrich the variances in the current dataset to increase its complexity. Experimental results on the PASCAL VOC 2010 Person Part dataset and the MNIST dataset show that, in both conventional CNN models and capsule networks, the number of independent functional units is an important factor that determines the capacity of a network model. By adjusting the number of functional units, the capacity of a model can better match the complexity of a dataset.
Abstract: Motion recognition from videos is actually a very
complex task due to the high variability of motions. This paper
describes the challenges of human motion recognition, especially
motion representation step with relevant features. Our descriptor
vector is inspired from Laban Movement Analysis method. We
propose discriminative features using the Random Forest algorithm
in order to remove redundant features and make learning algorithms
operate faster and more effectively. We validate our method on
MSRC-12 and UTKinect datasets.
Abstract: Feature selection and attribute reduction are crucial
problems, and widely used techniques in the field of machine
learning, data mining and pattern recognition to overcome the
well-known phenomenon of the Curse of Dimensionality. This paper
presents a feature selection method that efficiently carries out attribute
reduction, thereby selecting the most informative features of a dataset.
It consists of two components: 1) a measure for feature subset
evaluation, and 2) a search strategy. For the evaluation measure,
we have employed the fuzzy-rough dependency degree (FRFDD)
of the lower approximation-based fuzzy-rough feature selection
(L-FRFS) due to its effectiveness in feature selection. As for the
search strategy, a modified version of a binary shuffled frog leaping
algorithm is proposed (B-SFLA). The proposed feature selection
method is obtained by hybridizing the B-SFLA with the FRDD. Nine
classifiers have been employed to compare the proposed approach
with several existing methods over twenty two datasets, including
nine high dimensional and large ones, from the UCI repository.
The experimental results demonstrate that the B-SFLA approach
significantly outperforms other metaheuristic methods in terms of the
number of selected features and the classification accuracy.
Abstract: Human motion recognition has been extensively increased in recent years due to its importance in a wide range of applications, such as human-computer interaction, intelligent surveillance, augmented reality, content-based video compression and retrieval, etc. However, it is still regarded as a challenging task especially in realistic scenarios. It can be seen as a general machine learning problem which requires an effective human motion representation and an efficient learning method. In this work, we introduce a descriptor based on Laban Movement Analysis technique, a formal and universal language for human movement, to capture both quantitative and qualitative aspects of movement. We use Discrete Hidden Markov Model (DHMM) for training and classification motions. We improve the classification algorithm by proposing two DHMMs for each motion class to process the motion sequence in two different directions, forward and backward. Such modification allows avoiding the misclassification that can happen when recognizing similar motions. Two experiments are conducted. In the first one, we evaluate our method on a public dataset, the Microsoft Research Cambridge-12 Kinect gesture data set (MSRC-12) which is a widely used dataset for evaluating action/gesture recognition methods. In the second experiment, we build a dataset composed of 10 gestures(Introduce yourself, waving, Dance, move, turn left, turn right, stop, sit down, increase velocity, decrease velocity) performed by 20 persons. The evaluation of the system includes testing the efficiency of our descriptor vector based on LMA with basic DHMM method and comparing the recognition results of the modified DHMM with the original one. Experiment results demonstrate that our method outperforms most of existing methods that used the MSRC-12 dataset, and a near perfect classification rate in our dataset.
Abstract: The constant monitoring of blood glucose level is necessary for maintaining health of patients and to alert medical specialists to take preemptive measures before the onset of any complication as a result of diabetes. The current clinical monitoring of blood glucose uses invasive methods repeatedly which are uncomfortable and may result in infections in diabetic patients. Several attempts have been made to develop non-invasive techniques for blood glucose measurement. In this regard, the existing methods are not reliable and are less accurate. Other approaches claiming high accuracy have not been tested on extended dataset, and thus, results are not statistically significant. It is a well-known fact that acetone concentration in breath has a direct relation with blood glucose level. In this paper, we have developed the first of its kind, reliable and high accuracy breath analyzer for non-invasive blood glucose measurement. The acetone concentration in breath was measured using MQ 138 sensor in the samples collected from local hospitals in Pakistan involving one hundred patients. The blood glucose levels of these patients are determined using conventional invasive clinical method. We propose a linear regression classifier that is trained to map breath acetone level to the collected blood glucose level achieving high accuracy.
Abstract: Over the past decade, there has been a steep rise in
the data-driven analysis in major areas of medicine, such as clinical
decision support system, survival analysis, patient similarity analysis,
image analytics etc. Most of the data in the field are well-structured
and available in numerical or categorical formats which can be used
for experiments directly. But on the opposite end of the spectrum,
there exists a wide expanse of data that is intractable for direct
analysis owing to its unstructured nature which can be found in the
form of discharge summaries, clinical notes, procedural notes which
are in human written narrative format and neither have any relational
model nor any standard grammatical structure. An important step
in the utilization of these texts for such studies is to transform
and process the data to retrieve structured information from the
haystack of irrelevant data using information retrieval and data mining
techniques. To address this problem, the authors present Q-Map in
this paper, which is a simple yet robust system that can sift through
massive datasets with unregulated formats to retrieve structured
information aggressively and efficiently. It is backed by an effective
mining technique which is based on a string matching algorithm
that is indexed on curated knowledge sources, that is both fast
and configurable. The authors also briefly examine its comparative
performance with MetaMap, one of the most reputed tools for medical
concepts retrieval and present the advantages the former displays over
the latter.
Abstract: Poor air quality is one of the main environmental causes of premature deaths worldwide, and mainly in cities, where the majority of the population lives. It is a consequence of successive land cover (LC) and use changes, as a result of the intensification of human activities. Knowing these landscape modifications in a comprehensive spatiotemporal dimension is, therefore, essential for understanding variations in air pollutant concentrations. In this sense, the use of air quality models is very useful to simulate the physical and chemical processes that affect the dispersion and reaction of chemical species into the atmosphere. However, the modelling performance should always be evaluated since the resolution of the input datasets largely dictates the reliability of the air quality outcomes. Among these data, the updated LC is an important parameter to be considered in atmospheric models, since it takes into account the Earth’s surface changes due to natural and anthropic actions, and regulates the exchanges of fluxes (emissions, heat, moisture, etc.) between the soil and the air. This work aims to evaluate the performance of the Weather Research and Forecasting model coupled with Chemistry (WRF-Chem), when different LC classifications are used as an input. The influence of two LC classifications was tested: i) the 24-classes USGS (United States Geological Survey) LC database included by default in the model, and the ii) CLC (Corine Land Cover) and specific high-resolution LC data for Portugal, reclassified according to the new USGS nomenclature (33-classes). Two distinct WRF-Chem simulations were carried out to assess the influence of the LC on air quality over Europe and Portugal, as a case study, for the year 2015, using the nesting technique over three simulation domains (25 km2, 5 km2 and 1 km2 horizontal resolution). Based on the 33-classes LC approach, particular emphasis was attributed to Portugal, given the detail and higher LC spatial resolution (100 m x 100 m) than the CLC data (5000 m x 5000 m). As regards to the air quality, only the LC impacts on tropospheric ozone concentrations were evaluated, because ozone pollution episodes typically occur in Portugal, in particular during the spring/summer, and there are few research works relating to this pollutant with LC changes. The WRF-Chem results were validated by season and station typology using background measurements from the Portuguese air quality monitoring network. As expected, a better model performance was achieved in rural stations: moderate correlation (0.4 – 0.7), BIAS (10 – 21µg.m-3) and RMSE (20 – 30 µg.m-3), and where higher average ozone concentrations were estimated. Comparing both simulations, small differences grounded on the Leaf Area Index and air temperature values were found, although the high-resolution LC approach shows a slight enhancement in the model evaluation. This highlights the role of the LC on the exchange of atmospheric fluxes, and stresses the need to consider a high-resolution LC characterization combined with other detailed model inputs, such as the emission inventory, to improve air quality assessment.
Abstract: This paper has critically examined the use of Machine Learning procedures in curbing unauthorized access into valuable areas of an organization. The use of passwords, pin codes, user’s identification in recent times has been partially successful in curbing crimes involving identities, hence the need for the design of a system which incorporates biometric characteristics such as DNA and pattern recognition of variations in facial expressions. The facial model used is the OpenCV library which is based on the use of certain physiological features, the Raspberry Pi 3 module is used to compile the OpenCV library, which extracts and stores the detected faces into the datasets directory through the use of camera. The model is trained with 50 epoch run in the database and recognized by the Local Binary Pattern Histogram (LBPH) recognizer contained in the OpenCV. The training algorithm used by the neural network is back propagation coded using python algorithmic language with 200 epoch runs to identify specific resemblance in the exclusive OR (XOR) output neurons. The research however confirmed that physiological parameters are better effective measures to curb crimes relating to identities.
Abstract: Network security engineers work to keep services available all the time by handling intruder attacks. Intrusion Detection System (IDS) is one of the obtainable mechanisms that is used to sense and classify any abnormal actions. Therefore, the IDS must be always up to date with the latest intruder attacks signatures to preserve confidentiality, integrity, and availability of the services. The speed of the IDS is a very important issue as well learning the new attacks. This research work illustrates how the Knowledge Discovery and Data Mining (or Knowledge Discovery in Databases) KDD dataset is very handy for testing and evaluating different Machine Learning Techniques. It mainly focuses on the KDD preprocess part in order to prepare a decent and fair experimental data set. The J48, MLP, and Bayes Network classifiers have been chosen for this study. It has been proven that the J48 classifier has achieved the highest accuracy rate for detecting and classifying all KDD dataset attacks, which are of type DOS, R2L, U2R, and PROBE.
Abstract: Missing values in real-world datasets are a common
problem. Many algorithms were developed to deal with this
problem, most of them replace the missing values with a fixed
value that was computed based on the observed values. In
our work, we used a distance function based on Bhattacharyya
distance to measure the distance between objects with missing
values. Bhattacharyya distance, which measures the similarity of
two probability distributions. The proposed distance distinguishes
between known and unknown values. Where the distance between
two known values is the Mahalanobis distance. When, on the other
hand, one of them is missing the distance is computed based on the
distribution of the known values, for the coordinate that contains
the missing value. This method was integrated with Wikaya, a
digital health company developing a platform that helps to improve
prevention of chronic diseases such as diabetes and cancer. In order
for Wikaya’s recommendation system to work distance between users
need to be measured. Since there are missing values in the collected
data, there is a need to develop a distance function distances between
incomplete users profiles. To evaluate the accuracy of the proposed
distance function in reflecting the actual similarity between different
objects, when some of them contain missing values, we integrated it
within the framework of k nearest neighbors (kNN) classifier, since
its computation is based only on the similarity between objects. To
validate this, we ran the algorithm over diabetes and breast cancer
datasets, standard benchmark datasets from the UCI repository. Our
experiments show that kNN classifier using our proposed distance
function outperforms the kNN using other existing methods.
Abstract: Hierarchical classification is a special type of classification task where the class labels are organised into a hierarchy, with more generic class labels being ancestors of more specific ones. Meta-learning for classification-algorithm recommendation consists of recommending to the user a classification algorithm, from a pool of candidate algorithms, for a dataset, based on the past performance of the candidate algorithms in other datasets. Meta-learning is normally used in conventional, non-hierarchical classification. By contrast, this paper proposes a meta-learning approach for more challenging task of hierarchical classification, and evaluates it in a large number of bioinformatics datasets. Hierarchical classification is especially relevant for bioinformatics problems, as protein and gene functions tend to be organised into a hierarchy of class labels. This work proposes meta-learning approach for
recommending the best hierarchical classification algorithm to a
hierarchical classification dataset. This work’s contributions are: 1)
proposing an algorithm for splitting hierarchical datasets into
new datasets to increase the number of meta-instances, 2) proposing
meta-features for hierarchical classification, and 3) interpreting
decision-tree meta-models for hierarchical classification algorithm
recommendation.
Abstract: This paper presents a method for improving object search accuracy using a deep learning model. A major limitation to provide accurate similarity with deep learning is the requirement of huge amount of data for training pairwise similarity scores (metrics), which is impractical to collect. Thus, similarity scores are usually trained with a relatively small dataset, which comes from a different domain, causing limited accuracy on measuring similarity. For this reason, this paper proposes a deep learning model that can be trained with a significantly small amount of data, a clustered data which of each cluster contains a set of visually similar images. In order to measure similarity distance with the proposed method, visual features of two images are extracted from intermediate layers of a convolutional neural network with various pooling methods, and the network is trained with pairwise similarity scores which is defined zero for images in identical cluster. The proposed method outperforms the state-of-the-art object similarity scoring techniques on evaluation for finding exact items. The proposed method achieves 86.5% of accuracy compared to the accuracy of the state-of-the-art technique, which is 59.9%. That is, an exact item can be found among four retrieved images with an accuracy of 86.5%, and the rest can possibly be similar products more than the accuracy. Therefore, the proposed method can greatly reduce the amount of training data with an order of magnitude as well as providing a reliable similarity metric.
Abstract: Data mining technique used in the field of clustering is a subject of active research and assists in biological pattern recognition and extraction of new knowledge from raw data. Clustering means the act of partitioning an unlabeled dataset into groups of similar objects. Each group, called a cluster, consists of objects that are similar between themselves and dissimilar to objects of other groups. Several clustering methods are based on partitional clustering. This category attempts to directly decompose the dataset into a set of disjoint clusters leading to an integer number of clusters that optimizes a given criterion function. The criterion function may emphasize a local or a global structure of the data, and its optimization is an iterative relocation procedure. The K-Means algorithm is one of the most widely used partitional clustering techniques. Since K-Means is extremely sensitive to the initial choice of centers and a poor choice of centers may lead to a local optimum that is quite inferior to the global optimum, we propose a strategy to initiate K-Means centers. The improved K-Means algorithm is compared with the original K-Means, and the results prove how the efficiency has been significantly improved.
Abstract: Datasets or collections are becoming important assets by themselves and now they can be accepted as a primary intellectual output of a research. The quality and usage of the datasets depend mainly on the context under which they have been collected, processed, analyzed, validated, and interpreted. This paper aims to present a collection of program educational objectives mapped to student’s outcomes collected from self-study reports prepared by 32 engineering programs accredited by ABET. The manual mapping (classification) of this data is a notoriously tedious, time consuming process. In addition, it requires experts in the area, which are mostly not available. It has been shown the operational settings under which the collection has been produced. The collection has been cleansed, preprocessed, some features have been selected and preliminary exploratory data analysis has been performed so as to illustrate the properties and usefulness of the collection. At the end, the collection has been benchmarked using nine of the most widely used supervised multiclass classification techniques (Binary Relevance, Label Powerset, Classifier Chains, Pruned Sets, Random k-label sets, Ensemble of Classifier Chains, Ensemble of Pruned Sets, Multi-Label k-Nearest Neighbors and Back-Propagation Multi-Label Learning). The techniques have been compared to each other using five well-known measurements (Accuracy, Hamming Loss, Micro-F, Macro-F, and Macro-F). The Ensemble of Classifier Chains and Ensemble of Pruned Sets have achieved encouraging performance compared to other experimented multi-label classification methods. The Classifier Chains method has shown the worst performance. To recap, the benchmark has achieved promising results by utilizing preliminary exploratory data analysis performed on the collection, proposing new trends for research and providing a baseline for future studies.
Abstract: Intrusion detection systems (IDS) are the main components of network security. These systems analyze the network events for intrusion detection. The design of an IDS is through the training of normal traffic data or attack. The methods of machine learning are the best ways to design IDSs. In the method presented in this article, the pruning algorithm of C5.0 decision tree is being used to reduce the features of traffic data used and training IDS by the least square vector algorithm (LS-SVM). Then, the remaining features are arranged according to the predictor importance criterion. The least important features are eliminated in the order. The remaining features of this stage, which have created the highest level of accuracy in LS-SVM, are selected as the final features. The features obtained, compared to other similar articles which have examined the selected features in the least squared support vector machine model, are better in the accuracy, true positive rate, and false positive. The results are tested by the UNSW-NB15 dataset.
Abstract: Load forecasting has become crucial in recent years
and become popular in forecasting area. Many different power
forecasting models have been tried out for this purpose. Electricity
load forecasting is necessary for energy policies, healthy and reliable
grid systems. Effective power forecasting of renewable energy load
leads the decision makers to minimize the costs of electric utilities
and power plants. Forecasting tools are required that can be used
to predict how much renewable energy can be utilized. The purpose
of this study is to explore the effectiveness of LSTM-based neural
networks for estimating renewable energy loads. In this study, we
present models for predicting renewable energy loads based on
deep neural networks, especially the Long Term Memory (LSTM)
algorithms. Deep learning allows multiple layers of models to learn
representation of data. LSTM algorithms are able to store information
for long periods of time. Deep learning models have recently been
used to forecast the renewable energy sources such as predicting
wind and solar energy power. Historical load and weather information
represent the most important variables for the inputs within the
power forecasting models. The dataset contained power consumption
measurements are gathered between January 2016 and December
2017 with one-hour resolution. Models use publicly available data
from the Turkish Renewable Energy Resources Support Mechanism.
Forecasting studies have been carried out with these data via deep
neural networks approach including LSTM technique for Turkish
electricity markets. 432 different models are created by changing
layers cell count and dropout. The adaptive moment estimation
(ADAM) algorithm is used for training as a gradient-based optimizer
instead of SGD (stochastic gradient). ADAM performed better than
SGD in terms of faster convergence and lower error rates. Models
performance is compared according to MAE (Mean Absolute Error)
and MSE (Mean Squared Error). Best five MAE results out of
432 tested models are 0.66, 0.74, 0.85 and 1.09. The forecasting
performance of the proposed LSTM models gives successful results
compared to literature searches.
Abstract: With the Internet becoming the dominant channel for business and life, many IPs are increasingly masked using web proxies for illegal purposes such as propagating malware, impersonate phishing pages to steal sensitive data or redirect victims to other malicious targets. Moreover, as Internet traffic continues to grow in size and complexity, it has become an increasingly challenging task to detect the proxy service due to their dynamic update and high anonymity. In this paper, we present an approach based on behavioral graph analysis to study the behavior similarity of web proxy users. Specifically, we use bipartite graphs to model host communications from network traffic and build one-mode projections of bipartite graphs for discovering social-behavior similarity of web proxy users. Based on the similarity matrices of end-users from the derived one-mode projection graphs, we apply a simple yet effective spectral clustering algorithm to discover the inherent web proxy users behavior clusters. The web proxy URL may vary from time to time. Still, the inherent interest would not. So, based on the intuition, by dint of our private tools implemented by WebDriver, we examine whether the top URLs visited by the web proxy users are web proxies. Our experiment results based on real datasets show that the behavior clusters not only reduce the number of URLs analysis but also provide an effective way to detect the web proxies, especially for the unknown web proxies.
Abstract: Plant ecological stoichiometry, which is one of the most important tools to connect the components among different levels of ecosystem, has obtained increasingly extensive concern, especially on its responses to the environmental gradients. Based on the published literatures and datasets, this article focused on reviewing the variable responses of plant foliar ecological stoichiometry to the climatic factors, such as temperature, water, elevated CO2, and found that foliar ecological stoichiometry responded dynamically to climatic variations among different regions and different growth forms. Then, research status and deficiency were summarized and the expectation on studying the relationships between plant C, N and P ecological stoichiometry and environmental variations which can provide a reference to understand how plants will respond to global change in the future was pointed out.
Abstract: Crop yield prediction is a paramount issue in
agriculture. The main idea of this paper is to find out efficient
way to predict the yield of corn based meteorological records.
The prediction models used in this paper can be classified into
model-driven approaches and data-driven approaches, according to
the different modeling methodologies. The model-driven approaches are based on crop mechanistic
modeling. They describe crop growth in interaction with their
environment as dynamical systems. But the calibration process of
the dynamic system comes up with much difficulty, because it
turns out to be a multidimensional non-convex optimization problem.
An original contribution of this paper is to propose a statistical
methodology, Multi-Scenarios Parameters Estimation (MSPE), for the
parametrization of potentially complex mechanistic models from a
new type of datasets (climatic data, final yield in many situations).
It is tested with CORNFLO, a crop model for maize growth. On the other hand, the data-driven approach for yield prediction
is free of the complex biophysical process. But it has some strict
requirements about the dataset.
A second contribution of the paper is the comparison of these
model-driven methods with classical data-driven methods. For this
purpose, we consider two classes of regression methods, methods
derived from linear regression (Ridge and Lasso Regression, Principal
Components Regression or Partial Least Squares Regression) and
machine learning methods (Random Forest, k-Nearest Neighbor,
Artificial Neural Network and SVM regression).
The dataset consists of 720 records of corn yield at county scale
provided by the United States Department of Agriculture (USDA) and
the associated climatic data. A 5-folds cross-validation process and
two accuracy metrics: root mean square error of prediction(RMSEP),
mean absolute error of prediction(MAEP) were used to evaluate the
crop prediction capacity.
The results show that among the data-driven approaches, Random
Forest is the most robust and generally achieves the best prediction
error (MAEP 4.27%). It also outperforms our model-driven approach
(MAEP 6.11%). However, the method to calibrate the mechanistic
model from dataset easy to access offers several side-perspectives.
The mechanistic model can potentially help to underline the stresses
suffered by the crop or to identify the biological parameters of interest
for breeding purposes. For this reason, an interesting perspective is
to combine these two types of approaches.