Skip to main content
main-content

Über dieses Buch

This book constitutes refereed proceedings of the 12th International Conference on International Conference on Computational Collective Intelligence, ICCCI 2020, held in Da Nang, Vietnam, in November – December 2020. Due to the the COVID-19 pandemic the conference was held online.
The 68 papers were thoroughly reviewed and selected from 314 submissions. The papers are organized according to the following topical sections: ​data mining and machine learning; deep learning and applications for industry 4.0; recommender systems; computer vision techniques; decision support and control systems; intelligent management information systems; innovations in intelligent systems; intelligent modeling and simulation approaches for games and real world systems; experience enhanced intelligence to IoT; data driven IoT for smart society; applications of collective intelligence; natural language processing; low resource languages processing; computational collective intelligence and natural language processing.

Inhaltsverzeichnis

Frontmatter

Data Mining and Machine Learning

Frontmatter

Rule Induction of Automotive Historic Styles Using Decision Tree Classifier

For industrial designers, how to classify vehicle styles was a challenge, and at a large depends on the designer’s internal knowledge. Although data mining and machine learning technologies had become mature and affordable today, the applications using machine learning in car styling were few. This investigation focuses on using the decision tree method to discuss the relationship between automotive styles and the design features of 35 cars produced by the automotive manufacturer, Dodge between 1942 and 2017. The study summarized 8 design features from previous literature: the length, fender design, number of headlamps, rear form, the position of the quarter glass, engine hood scoop, rocket tail, and side decoration design while the styles are chosen were streamlined style, popular style, and modern style. The decision tree algorithm (C5.0) was employed to obtain the optimal rule of decision tree to compare the historic design styles from ten sets of decision tree rules. The result showed that there was a clear relationship between the key design features and the historic style of vehicles. The average accuracy of the ten sets of decision trees is 90.6%. The highest accuracy of the optimal model is 97%. However, the variation between the predicted accuracies of decision tree models calculated is high, ranging from 80% to 97%. Based on the decision tree and statistics method, the design features include the length, fender design, rocket tail design, rear form, and position of the quarter glass was more important than the others. This method had the potential to identify automotive historic styles based on key features.

Hung-Hsiang Wang, Chih-Ping Chen

Deep Learning for Multilingual POS Tagging

Various neural networks for sequence labeling tasks have been studied extensively in recent years. The main research focus on neural networks for the task are range from the feed-forward neural network to the long short term memory (LSTM) network with CRF layer. This paper summarizes the existing neural architectures and develop the most representative four neural networks for part-of-speech tagging and apply them on several typologically different languages. Experimental results show that the LSTM type of networks outperforms the feed-forward network in most cases and the character-level networks can learn the lexical features from characters within words, which makes the model achieve better results than no-character ones.

Alymzhan Toleu, Gulmira Tolegen, Rustam Mussabayev

Study of Machine Learning Techniques on Accident Data

Road crash is one of the major burning issues for Bangladesh. There are several factors that are responsible for occurring road crashes. If we can understand the causes and predict the severity level of a particular type of accident upfront, we can take necessary steps in the proper time to lessen the damages. In this study, we have built some predictive models of different homogeneous road crash groups of Bangladesh using machine learning methods that can predict that particular road crash severity level based on the environmental factors and road conditions. We have applied Agglomerative Hierarchical Clustering to find different clusters of road crashes and then applied Random Forest technique to extract the significant predictors of each cluster and then applied C5.0 to build predictive models of each cluster. Finally we have discussed the patterns of fatal and non-fatal accidents of Bangladesh through rule generation technique.

Zakaria Shams Siam, Rubyat Tasnuva Hasan, Soumik Sarker Anik, Ankit Dev, Sumaia Islam Alita, Mustafizur Rahaman, Rashedur M. Rahman

Soil Analysis and Unconfined Compression Test Study Using Data Mining Techniques

In this study, Random Forest Regressor, Linear Regression, Generalized Regression Neural Network (GRNN) and Fully connected Neural Network (FCNN) models are leveraged for predicting unconfined compression coefficient with respect to standard penetration test (N-value), depth and soil type. The study is focused on a particular correlation of undrained shear strength of clay (Cu) with the standard penetration strength. The data used is from 14 no. ward in Mymensingh and Rangamati districts which are situated in Bangladesh. By using this data, the study tries to solidify the correlation of SPT (N-value) with Cu. It also tries to check the goodness of the relationship by comparing it with unconfined compression strength values gained from the unconfined compression test calculated from the field by experts.

Abdullah Md. Sarwar, Sayeed Md. Shaiban, Suparna Biswas, Arshi Siddiqui Promiti, Tarek Ibne Faysal, Lubaba Bazlul, Md. Sazzad Hossain, Rashedur M. Rahman

Self-sorting of Solid Waste Using Machine Learning

In waste recycling, the source separation model, decentralises the sorting responsibility to the consumer when they dispose, resulting in lower cross contamination, significantly increased recycling yield, and superior recovery material quality. This recycling model is problematic however, as it is prone to human error and community-level participation is difficult to incentivise with the greater inconvenience being placed on consumers. This paper aims to conceptualise a solution by proposing a unique mechatronic system in the form of a self-sorting smart bin. It is hypothesised that in order to overcome the high variability innate to disposed waste, a robust supervised machine learning classification model supported by IoT integration needs to be utilised. A dataset comprising of 680 samples of plastic, metal and glass recyclables was manually collected from a custom-built identification chamber equipped with a suite of sensors. The dataset was then split and used to train a modular neural network comprising of three concurrent individual classifiers for images (CNN), sounds (MLP) and time series (KNN-DTW). The output class probabilities were then integrated by one combined classifier (MLP), resulting in a prediction time of 0.67 s per sample, a prediction accuracy of 100%, and an average confidence of 99.75% averaged over 10 runs of an 18% validation split.

Tyson Chan, Jacky H. Cai, Francis Chen, Ka C. Chan

Clustering Algorithms in Mining Fans Operating Mode Identification Problem

Most of the machinery and equipment of the mine infrastructure is controlled by an industrial automation system. In practice, SCADA (Supervisory Control And Data Acquisition) systems very often acquire many operational parameters that have no further analytical use. The variability of recorded signals very often depends on the machine load, as well as organizational and technical aspects. Therefore, SCADA systems can be a practically free source of information used to determine KPI (e.g. performance, energy and diagnostic) for a single object as well as given mining process. For example, the ability to reliably identify different operational modes of the mining industrial fans based on data from SCADA gives a wide range of potential applications. Accurate information on this subject could be used - apart from basic monitoring and reporting needs (e.g. actual work vs. schedule comparison) – also in more complex problems, like power consumption predictions. Given the variety of industrial fans used in the mining industry and the different operational data collected, this is yet not a trivial task in a general case. The main aim of this article is to provide reliable algorithms solving fans operational mode identification issue, which will be possible to apply in a wide range of potential applications.

Bartosz Jachnik, Paweł Stefaniak, Natalia Duda, Paweł Śliwiński

K-Means Clustering for Features Arrangement in Metagenomic Data Visualization

Personalized medicine is one of the most concern of the scientists to propose successful treatments for diseases. This approach considers patients’ genetic make-up and attention to their preferences, beliefs, attitudes, knowledge and social context. Deep learning techniques hold important roles and obtain achievements in bioinformatics tasks. Metagenomic data analysis is very important to develop and evaluate methods and tools applying to Personalized medicine. Metagenomic data is usually characterized by high-dimensional spaces where humans meet difficulties to interpret data. Visualizing metagenomic data is crucial to provide insights in data which can help researchers to explore patterns in data. Moreover, these visualizations can be fetched into deep learning such as Convolutional Neural Networks to do prediction tasks. In this study, we propose a visualization method for metagenomic data where features are arranged in the visualization based on K-means clustering algorithms. We show by experiments on metagenomic datasets of three diseases (Colorectal Cancer, Obesity and Type 2 Diabetes) that the proposed approach not only provides a robust method for visualization where we can observe clusters in the images but also enables us to improve the performance in disease prediction with deep learning algorithms.

Hai Thanh Nguyen, Toan Bao Tran, Huong Hoang Luong, Trung Phuoc Le, Nghi C. Tran, Quoc-Dinh Truong

Small Samples of Multidimensional Feature Vectors

A small sample of multidimensional feature vectors appears when the number of features is much greater than the number of objects (feature vectors).For example, such circumstances appear typically in genetic data sets. In such cases, feature clustering can become a useful tool in classification or prognosis tasks. Feature clustering can be performed through the minimization of the convex and piecewise linear (CPL) criterion functions.

Leon Bobrowski

Using Fourier Series to Improve the Discrete Grey Model (1, 1)

Discrete grey model (1, 1) (abbreviates as DGM (1, 1)), is a version of grey forecasting model. Since appeared, its has been attracted by many scientists in dealing with the problem related to uncertainly information and small sample data. In recent years, this model has been improved the accuracy in forecast by scientifics. However, the existing DGM (1, 1) model cannot be used in some special scenarios such as the significant fluctuation or noise in data. Solving this issue, this paper propose a novel grey forecasting model named as Fourier Discrete Grey Model (1, 1) (abbreviated as F-DGM (1, 1)). This model was built by combined the Fourier series and DGM (1, 1) model. Through the example in Xie and Liu’s paper (Xie and Liu [28]) and practical application, these simulation outcomes demonstrated that the F-DGM (1, 1) model provided remarkable prediction performance compared with the other grey forecasting models. Future direction, the authors will use different equations or different methodologies to improve the DGM (1, 1) model. The other direction is applied the proposed model for dealing with the highly fluctuation data in different industries.

Van-Thanh Phan, Zbigniew Malara, Ngoc Thang Nguyen

Studying on the Accuracy Improvement of GM (1, 1) Model

In order to expand the application of GM (1, 1) in the condition of fluctuation data and incomplete information, this paper proposed the new systematic optimization based on three steps as follows. First step, we used parameters c1 to transform any sequence data into the non-negative sequence data. The second, we used moving average operation method on the new sequence data to smooth the sequence data aim to satisfy the quasi-exponential condition and quasi-smooth condition. The final, we adopt Fourier series to modify residual error of model a grey sequence. To demonstrate the superiority of the proposed model, the numerical example in the research of Wang and Hsu and the raw data sequence are used. The simulation outcomes show that the proposed approach provides a better forecast results than several different kinds of grey forecasting models with the lowest average of MAPE for in and out-of-samples in two cases. For future direction, the author will applied different methodologies to improve the performance of GM (1, 1) or use proposed model to analyse the issues with high fluctuation data.

Van Đat Nguyen, Van-Thanh Phan, Ngoc Thang Nguyen, Doan Nhan Dao, Le Thanh Ha

Deep Learning and Applications for Industry 4.0

Frontmatter

An Evaluation of Image-Based Malware Classification Using Machine Learning

This paper investigates the image-based malware classification using machine learning techniques. It is a recent approach for malware classification in which malware binaries are converted into images (i.e. malware images) prior to feeding machine learning models, i.e. k-nearest neighbour (k-NN), Naïve Bayes (NB), Support Vector Machine (SVM) or Convolution Neural Networks (CNN). This approach relies on image texture to classify a malware instead of signatures or behaviours of malware collected via malware analysis, thus it does not encounter a problem if the signatures of a new malware variant has not been collected or the behaviours of a new malware variant has not been updated.This paper evaluates classification performance of various machine learning classifiers (i.e. k-NN, NB, SVM, CNN) fed by malware images in various dimensions (i.e., 128 × 128, 64 × 64, 32 × 32, 16 × 16). The experiment results achieved on three different datasets including Malimg, Malheur and BIG2015 show that k-NN outperforms others on three datasets with high accuracy (i.e. 97.9%, 94.41% and 95.63% respectively). On the contrary, NB showed its weakness on image-based malware classification. Experiment results also indicate that the accuracy of the k-NN reaches the highest value at the input image size of 32 × 32 and tends to reduce if too many feature information provided by large input images, i.e. 64 × 64, 128 × 128.

Tran The Son, Chando Lee, Hoa Le-Minh, Nauman Aslam, Moshin Raza, Nguyen Quoc Long

Automatic Container Code Recognition Using MultiDeep Pipeline

Identification of license plates on intermodal containers (or containers) while entering and departing from the yard provides a wide range of practical benefits, such as organizing automatic opening of the rising arm barrier at the entrance and exit to and from the site. In addition, automatic container code recognition can also assist in thwarting the entrance of unauthorized vehicles into the territory. With the recent development of AI, this process is preferably automatic. However, the poor quality of images obtained from surveillance cameras might have detrimental effects on AI models. To deal with this problem, we present a pipeline dubbed as MultiDeep system, which combines several state-of-the-art deep learning models for character recognition and computer vision processes to solve problems of real camera data. We have also compared our results with other pipeline models on real data and accomplished fairly positive results. In this paper, without further references, we will only consider intermodal containers when referring to them as containers.

Duy Nguyen, Duc Nguyen, Thong Nguyen, Khoi Ngo, Hung Cao, Thinh Vuong, Tho Quan

An Efficient Solution for People Tracking and Profiling from Video Streams Using Low-Power Compute

Balancing between performance and speed is vital for real-time applications. Given some of the latest edge devices, such as Raspberry Pi 4, Intel Neural Compute Stick 2, or Nvidia Jetson series, edge processing can become a valid choice for deploying computer vision algorithms in real-time scenarios. Object detection and tracking are two common problems that can be solved using such algorithms, which can be deployed with reasonable performance and speed on edge devices. In this paper, we show that the YOLO architecture can be successfully used for object detection and DeepSORT for object tracking on edge devices. The objects of interest in our scenario are persons, thus indicating face detection and tracking as another problem that is solved in the scope of the paper. Using Raspberry Pi 4 and Intel Neural Compute Stick 2, object detection and tracking models can be run on edge devices, though at around half the performance and more than 10 times slower than on a GPU server.

Marius Eduard Cojocea, Traian Rebedea

Simple Pose Network with Skip-Connections for Single Human Pose Estimation

Recently, following the success of deep convolutional neural networks, human pose estimation problem has been largely improved. This paper introduces an improved version of the Simple Pose network for single human pose estimation. It adds the skip-connections between the same-resolution layers of the backbone and up-sampling stream to fuse low-level and high-level features. To make the depth of features from low-level and high-level are same, this paper uses $$1\,\times \,1$$ 1 × 1 convolutional layer. The experiments show that this naive technique makes the new networks better over 1% mAP scores with just a small increment in model size.

Van-Thanh Hoang, Kang-Hyun Jo

Simple Fine-Tuning Attention Modules for Human Pose Estimation

The convolution neural networks (CNNs) have achieved the best performance not only for human pose estimation but also for other computer vision tasks (e.g., object detection, semantic segmentation, image classification). Then this paper focuses on a useful attention module (AM) for feed-forward CNNs. Firstly, feed the feature map after a block in the backbone network into the attention module, split into two separate dimensions, channel and spatial. After that, the AM combines these two feature maps by multiplication and gives it to the next block in the backbone. The network can capture the information in the long-range dependencies (channel) and the spatial data, which can gain better performance in accuracy. Therefore, our experimental results will illustrate how different between when using the attention module and the existing methods. As a result, the predicted joint heatmap maintains the accuracy and spatially better with the simple baseline. Besides, the proposed architecture gains 1.0 points in AP higher than the baseline. Moreover, the proposed network trained on COCO 2017 benchmarks, which is an accessible dataset nowadays.

Tien-Dat Tran, Xuan-Thuy Vo, Moahamammad-Ashraf Russo, Kang-Hyun Jo

Human Eye Detector with Light-Weight and Efficient Convolutional Neural Network

The human eye detection plays an important role in computer vision. Along with face detection, it is widely applied in practical security, surveillance, and warning systems such as eye tracking system, eye disease detection, gaze detection, eye blink, and drowsiness detection system. There have been many studies to detect eyes from applying traditional methods to using modern methods based on machine learning and deep learning. This network is deployed with two main blocks, namely the feature extraction block and the detection block. The feature extraction block starts with the use of the convolution layers, C.ReLU (Concatenated Rectified Linear Unit) module, and max pooling layers alternately, followed by the last six inception modules and four convolution layers. The detection block is constructed by two sibling convolution layers using for classification and regression. The experiment was trained and tested on CEW (Closed Eyes In The Wild), BioID Face and GI4E (Gaze Interaction for Everybody) dataset with the results achieved 96.48%, 99.58%, and 75.52% of AP (Average Precision), respectively. The speed was tested in real-time by 37.65 fps (frames per second) on Intel Core I7-4770 CPU @ 3.40 GHz.

Duy-Linh Nguyen, Muhamad Dwisnanto Putro, Kang-Hyun Jo

Recommender Systems

Frontmatter

Robust Content-Based Recommendation Distribution System with Gaussian Mixture Model

Recommendation systems play an very important role in boosting purchasing consumption for many manufacturers by helping consumers find the most appropriate items. Furthermore, there is quite a range of recommendation algorithms that can be efficient; however, a content-based algorithm is always the most popular, powerful, and productive method taken at the begin time of any project. In the negative aspect, somehow content-based algorithm results accuracy is still a concern that correlates to probabilistic similarity. In addition, the similarity calculation method is another crucial that affect the accuracy of content-based recommendation in probabilistic problems. Face with these problems, we propose a new content-based recommendation based on the Gaussian mixture model to improve the accuracy with more sensitive results for probabilistic recommendation problems. Our proposed method experimented in a liquor dataset including six main flavor taste, liquor main taste tags, and some other criteria. The method clusters n liquor records relied on n vectors of six dimensions into k group ( $$k < n$$ k < n ) before applying a formula to sort the results. Compared our proposed algorithm with two other popular models on the above dataset, the accuracy of the experimental results not only outweighs the comparison to those of two other models but also attain a very speedy response time in real-life applications.

Dat Nguyen Van, Van Toan Pham, Ta Minh Thanh

Incremental SVD-Based Collaborative Filtering Enhanced with Diversity for Personalized Recommendation

Along with the rapid rise of the internet, an e-commerce website brings enormous benefits for both customers and vendors. However, many choices are given at the same time makes customers have difficulty in choosing the most suitable products. A rising star solution for this is the recommender system which helps to narrow down the amount of suitable and relevant products for each customer. Matrix factorization is one of the most popular techniques used in recommender systems because of its effectiveness and simplicity. In this paper, we introduce a matrix factorization-based recommender system using Singular Value Decomposition (SVD) with some improvements in collaborative filtering and incremental learning. The SVD-based collaborative filtering methods can help generate personalized recommendations by combining user profiles. Moreover, the recommendation lists generated by the system are enhanced with diversity, which might attract more customer interests. Amazon’s Electronic data set is used to evaluate our proposed framework of the SVD-based recommender system. The experimental results show that our framework is promising.

Minh Quang Pham, Thi Thanh Sang Nguyen, Pham Minh Thu Do, Adrianna Kozierkiewicz

Collaborative Filtering Recommendation Based on Statistical Implicative Analysis

In recent studies on recommender models, association rules have been applied in many studies to improve the effectiveness of recommender models. However, these studies also reveal some drawbacks, such as the models take a considerable amount of time to generate association rules for large datasets; generation algorithms can ignore rules with the significant implication that affect the quality of recommender models. This study proposes collaborative filtering recommender models (CF models) based on association rules following an asymmetric approach of the statistical implicative analysis method to enhance the precision of recommender models. Through experiments on standard datasets and quality comparison with other CF models, we conclude that the proposed models based on the asymmetric relationship achieve better accuracy on the experimental datasets.

Hiep Xuan Huynh, Nghia Quoc Phan, Nghia Duong-Trung, Ha Thu Thi Nguyen

Computer Vision Techniques

Frontmatter

Object Searching on Video Using ORB Descriptor and Support Vector Machine

One of the main stages in object searching on video is extracting object regions from video. Template matching is popular technique for performing a such task. However, the use of template matching has a limitation that requires a large object as a template. If the template size is too small, it would obtain few features. On the other hand, ORB descriptors are often used for representing the object with a good accuracy and fast processing time. Therefore, this research proposed to use machine learning method combining with ORB descriptor for object searching on video data. Processing video in all frames is inefficient. Thus, frames are selected into keyframes using mutual information entropy. The ORB descriptors are then extracted from selected frame in order to find candidate region of objects. To verify and classify the object regions, multiclass support vector machine was used to train ORB descriptor of regions. For evaluation, the use of ORB would be compared with other descriptor, such as SIFT and SURF for showing its superiority in both accuracy and processing time. In experiment, it is found that object searching with ORB descriptor performs faster processing time, which is 0.219 s, while SIFT 1.011 s and SURF 0.503 s. Meanwhile, it also achieves the best F1 value, which is 0.9 compared to SIFT 0.63 and SURF 0.65.

Faisal Dharma Adhinata, Agus Harjoko, Wahyono

An Improved of Joint Reversible Data Hiding Methods in Encrypted Remote Sensing Satellite Images

Data protection security is very necessary when distributing high resolution remote sensing satellite images from LAPAN to users via electronic media. Reversible data hiding and encryption are two very useful methods for protecting privacy and data security. This paper proposes an increase in the method of joint reversible data hiding on remote sensing satellite images based on the algorithm of Zhang’s work, Hong et al.’s work, and Fatema et al.’s work. To evaluate the smoothness of the blocks, a modification of the fluctuation calculation function is presented. The experimental results show that the modified calculation function gives better estimation results. Then, the proposed method gives a lower extracted bit error rate and the visual quality of the image from the proposed method is better than the three references. For example, when the block size is 8 × 8, the extracted-bit error rate (EER) of the SPOT-6 test image of the proposed modified function was 8.40%, which is quite lower than the 14.14% EER of Zhang’s function, 9.62% EER of Hong et al.’s function and 11.87% EER of Fatema’s et al.’s method. Likewise, the quality of SPOT-6 image recovery represented by the peak signal-to-noise ratio (PSNR) of proposed modified function is 50.52 dB, which is slightly higher than the 48.23 dB PSNR of Zhang’s function, 49.93 dB PSNR of Hong et al.’s function and 49.00 dB PSNR of Fatema’s et al.’s function.

Ali Syahputra Nasution, Gunawan Wibisono

3D Kinematics of Upper Limb Functional Assessment Using HTC Vive in Unreal Engine 4

The purpose of research in this paper is to quantify the accuracy and precision of HTC Vive by making upper limb assessment measurements and performing functional tasks in the Unreal Engine 4. Thirty healthy males performed daily aim functional tasks, and arm length measurement and assessment were made. Each participant attended two testing sessions and one arm length measurement session. The upper limb length was measured using HTC Vive after making three types of hand posture exercises. The arm assessment included the minimum and maximum angle of shoulder adduction, abduction, flexion and extension. The experiment showed all the upper limb measurements collected from the functional tasks as well as the position and rotation of the upper limb could be estimated correctly.

Kai Liang Lew, Kok Swee Sim, Shing Chiang Tan, Fazly Salleh Abas

2D-CNN Based Segmentation of Ischemic Stroke Lesions in MRI Scans

Stroke is the second overall driving reason for human death and disability. Strokes are categorized into Ischemic and Hemorrhagic strokes. Ischemic stroke is 85% of strokes while hemorrhagic is 15%. An exact automatic lesion segmentation of ischemic stroke remains a test to date. A few machine learning techniques are applied previously to beat manual human observers yet slacks to survive. In this paper, we propose a completely automatic lesion segmentation of ischemic stroke in view of the Convolutional Neural Network (CNN). The dataset used as a part of this study is obtained from ISLES 2015 challenge, included four MRI modalities DWI, T1, T1c, and FLAIR of 28 patients. The CNN model is trained on 25 patient’s data while tested on the remaining 3 patients. As CNN is only used for classification, we convert segmentation to the pixel-by-pixel classification tasks. Dice Coefficient (DC) is used as a performance evaluation metric for assessing the performance of the model. The experimental results show that the proposed model achieves a comparatively higher DC rate from 4–5% than the considered state-of-the-art machine learning techniques.

Pir Masoom Shah, Hikmat Khan, Uferah Shafi, Saif ul Islam, Mohsin Raza, Tran The Son, Hoa Le-Minh

Melanoma Skin Cancer Classification Using Transfer Learning

Melanoma is one of the most aggressive types of skin cancer as it rapidly spreads to various areas of the body. With the increase and fatal nature of melanoma, it is of utmost importance to establish computer assisted diagnostic support systems to aid physicians in diagnosing skin cancer. In this paper, we make use of deep learning and transfer learning by testing 14 pre-trained models for the classification and detection of skin cancer. Historically, the data in which Deep Convolutional Neural Networks are fed and trained on comes predominantly from European datasets resulting in biased data. To overcome this issue, we first determine the differences of melanoma that lie within people of different skin tones. Thereafter, we make use of the GrabCut segmentation technique to accurately segment the lesion from the surrounding skin tone in order to solely focus on the lesion. The pre-trained CNN, Squeezenet1-1, achieved the best experimental results with an accuracy rate of 93.42%, sensitivity of 92.11% and specificity of 94.74%. The experimental results achieved indicate that there is a possible solution to the underrepresented data of dark-skinned people.

Verosha Pillay, Divyan Hirasen, Serestina Viriri, Mandlenkosi Gwetu

Decision Support and Control Systems

Frontmatter

Design a Neural Controller to Control Rescue Quadcopter in Hang Status

Quadcopters can be used for various applications in many fields including service, life, security, military… This is a research of oriented application development prospects in the near future because of the quadcopters ability to move flexibly regardless of the terrain and can support human issues that quadcopter on the ground do not. However, due to the motion characteristics in the air, so the quadcopter has certain limitations, including moving the static problem, the suspend state of the quadcopter. This is a complex subject and many scientists are interested in studying the development of large size quadcopters that can carry both people and heavy equipment. This paper analyzes some problems associated with quadcopter motion in a static state and designs a neural controller as the basis for developing more advanced applications in practice to manufacture the big quadcopter for load. The simulation results illustrate the problem and explain the relevance of the theory.

Nguyen Hoang Mai, Le Quoc Huy, Tran The Son

Multidimensional Analysis of SCADA Stream Data for Estimating the Energy Efficiency of Mining Transport

This paper outlines the recommendation of analytical tools likely to be derived from the data recorded within the industrial automation system. The means might facilitate optimization of process efficiency, especially in terms of energy efficiency. Basically, each electromechanical device is electrically charged and controlled by the industrial automation system. A kind of the signal usually depends on various operational modes of the given device which are classified by its load. Available signal segmentation and statistical methods lead to the automatic identification of these modes and working patterns or abnormal performances caused by poor technical condition. Therefore, simple electrical signal allows to count the real device performance time and utilities usage, to identify its working modes, to recognize process losses, to specify KPI factors and to develop diagnostics. This paper describes multidimensional processing of conveyor stream data along with their exemplary use in real-time data. The algorithm of identifying operational regimes is characterized based on machine learning and further in-context analyses paired with visualisations.

Paweł Stefaniak, Paweł Śliwiński, Natalia Duda, Bartosz Jachnik

A Simple Method of the Haulage Cycles Detection for LHD Machine

In underground mining of metal ores, horizontal transport of material is performed using self-propelled machines, especially Load-Haul-Dump machines. For example, in KGHM underground mines, where room-and-pillar system is used to deposit exploitation, the haulage process is provided by wheel loaders and haul trucks with suitably adjusted operation configuration. In case of shorter haulage routes, only wheel loaders take part in haulage process. Currently, there is observed a global tendency reliant on develop predictive maintenance as well as navigation or production optimization using Industrial Internet of Things (IIOT). Unfortunately, analytics development in this domain requires full insight into machine’s workflow in mining excavations and multivariate analysis in order widely understanding of machine operating contexts. In this article, a quick method to haulage cycle identification on example of wheel loader has been proposed. Developed algorithm is based on hydraulic pressure signal segmentation which provides to recognize loading operation, haulage and return of machine to mining face after unloading material in dumping point. The method is based on smooth hydraulic pressure signal in order to reduce signal interference but introduce to apply a convolution of smoothed signal with inverted step function.

Koperska Wioletta, Skoczylas Artur, Stefaniak Paweł

Haul Truck Cycle Identification Using Support Vector Machine and DBSCAN Models

The haul trucks are one of the most often used assets in horizontal transport in underground copper ore mining. This haulage process has a cyclic form and in simple terms the machine drives from point A to point B, where its cargo box is respectively loaded and dumped. What is most important in its basic performance assessment is to identify each cycle and its parametrization in terms of total duration, idling, fuel consumption, and driving speed. In the literature, we can find a few similar works but the majority of them is based on a poorly available hydraulic pressure signal of the actuator in cargo box unloading system or braking system pressure signal. Unfortunately, all of them are not robust and unreliable in real, noisy signals. For this reason, searching for an innovative new concept of the solving problem seems right in this state. This paper describes the new method of the operation cycles identification for underground haul trucks, which is based on multidimensional techniques of operational data analysis using machine learning. The leading idea consists of three parts: searching characteristic non-hydraulic values in signals which correspond to cycles, identifying distinctive periods in haulage process and splitting signal into cycles accordingly. In the first step, the data mining techniques are used to find significant variables, afterwards SVM classifying model identifies unloadings, which are then applied to cluster data by DBSCAN algorithm. The whole process is presented on haul trucks real data from KGHM Lubin mine.

Dawid Gawelski, Bartosz Jachnik, Pawel Stefaniak, Artur Skoczylas

Intelligent Management Information Systems

Frontmatter

Data Quality Management in ERP Systems – Accounting Case

ERP systems process data obtained from heterogeneous sources and therefore the data are characterized by different quality. In order to effectively support management, ERP systems must be based on high-quality data. This is a prerequisite for making decisions within the company. The aim of this paper is to analyse the problems of data quality management in ERP systems and its main contribution is to develop procedures for data quality management in data processing, especially accounting data. The basic outcome of the conducted research is the systematization of the knowledge concerning the data allowing for making quick and effective economic decisions. Based on the results of the research it was concluded that a large number of data control procedures embedded in the ERP allows for ensuring appropriate data quality.

Marcin Hernes, Andrzej Bytniewski, Karolina Mateńczuk, Artur Rot, Szymon Dziuba, Marcin Fojcik, Tran Luong Nguyet, Paweł Golec, Agata Kozina

A Model of Enterprise Analytical Platform for Supply Chain Management

The article discusses an analytical platform for executives in the context of the needs for integrated supply chain management in large multi-site production companies. The first part discusses the development of methods, techniques, and technologies to improve supply chain management, while the second part presents an enterprise analytical platform model along with its basic assumptions and requirements. The basic elements of the platform and their functional description were discussed and groups of users who apply them to perform particular business roles in the company were identified. The third part presents a case study of selected business processes performed using on the enterprise analytical platform as part of integrated supply chain management. The contribution attempted with this article is to propose a model of an enterprise analytical platform taking into account an integrated supply chain for large multi-site production companies. Conclusions presented in this article may be significant in the following areas: the development of an enterprise analytical platform in terms of selecting individual components, understanding the process of developing executive information, making key decisions related to the development of an analytical platform, as well as developing the company’s strategy.

Paweł Pyda, Helena Dudycz, Paweł Stefaniak

Identification the Determinants of Pre-revenue Young Enterprises Value

Valuation of a companies at the start-up stage is an important factor in its development. Like determining the risk of its success. Valuation process support by machine learning systems can significantly simplify investment decision making, identify risks and start-up value. Hence the problem of how to estimate the value of an enterprise and what factors determine it. The purpose of the article is to identify independent variables (attribute) for the enterprise value in the pre-revenue phase. The practical result of the research is the possibility of using machine learning solutions defining the dependence of values on independent variables in internal, spatial and temporal systems. For the identification of variables’, a methodical review of the start-up’s valuation was used. An analysis of the innovation evaluation proposals and the significant concept of the Business Model were carried. The analysis of independent variables’ in various methods was performed by means of comparisons method in the form of a table. As a result of research, main independent variables were identified. Variables are both internal and external to the organization. Limitations: access to data to build a model with the right amount of data.

Robert Golej

Blockchain Platform Taxonomy

Now blockchain technology continues to develop – lots of new projects and platforms are being created. Some of these technological solutions are based on approved platforms, but other developers use their own new model types. The statistics show that blockchain technology is one of the future-proof technologies like artificial intelligence, Internet of Things and nanotechnology. However, the lack of standardisation in the field of the modern technologies can cause negative consequences for developers and financial institutes. The purpose of this study is to create the common open classification of existing blockchain platforms. Classification is a hierarchical structure consisting of six key characteristics: token, transaction, block, framework, network and communication. Each characteristic consists of groups (qualitative features), including sets. Such approaches allow to standardise the process of forming the blockchain platform architecture. Moreover, it is possible to create various scenarios of the network changes.

Andrew A. Varnavskiy, Ulia M. Gruzina, Anastasiya O. Buryakova

Brain Tumor Medical Diagnosis. How to Assess the Quality of Projection Model?

One of the critical problems of decision models used for medical purposes is to assess its projection quality. The paper refers to the most popular approaches (e.g., SKALE, GSS, CRGS) used to estimate postoperative outcomes in neurosurgery i.e., before the surgery begins. The methods were examined to select the best one in terms of model performance i.e., predictive power.The analysis revealed that the PS SCORE approach gave the most reliable forecasts. For the purpose of model performance assessment, the authors used AUROC statistics. It was shown that AUROC could be interpreted as the probability of correct classification. As the clinical data are usually scarce, the authors decided to enrich the model performance analysis by incorporating the confidence interval approach. This gave additional information regarding model stability.

Pawel Siarka

Meta-learning Process Analytics for Adaptive Tutoring Systems

The effectiveness of the learning process depends to a large extent on the provision of an adequate instruction program. Individual approach to learners’ needs posits the necessity to comply with the characteristics of their preferences and predispositions for learning, which ought to be considered in tutoring content delivery. Adapting the instruction material to a particular learner requires intelligent tutoring system (ITS). In turn, the adaptation mechanism itself is based on comprehensive data analytics derived from the learning environment and the learner’s behaviour as well. The paper presents a new approach in which intelligent tutor can deliver a support on meta level of learning process through analytics and adaptability. Most commonly, the analytics and adaptation in ITS are targeted mainly at the instruction program, but since the individual approach assumes that each learner should be treated individually, we must consider replacing active role of a human teacher, by the functionality of the intelligent tutor. Therefore, the role of supervising the learning process depends on the user themselves. It means that the meta-learning process has to be the subject of analysis and adjustment. Users usually do not have enough knowledge about how to learn, and for that reason the task of monitoring learner’s performance and adjusting all activities should be ensured by the system.

Gracja Niesler, Andrzej Niesler

Innovations in Intelligent Systems

Frontmatter

Visualization of Structural Dependencies Hidden in a Large Data Set

Conceptual models are very often used for visualization of complex domains. Such models help, for example, in understanding the relationships between entities. The paper presents valuable extensions to a method proposed by the author that creates a domain model in the form of a UML class diagram from raw data included in a data set. The formally defined extensions address the main limitations of previously presented algorithms, significantly increasing the readability of generated diagram. The new elements include inferring data types and class names from additional information in data.

Bogumila Hnatkowska

Internet Advertising Strategy Based on Information Growth in the Zettabyte Era

This research introduces a new method to evaluate and make most practical use of the growth of information on the Internet. The method is based on the “Internet in Real Time” statistics delivered by WebFX and Worldometer tools, and develops a combination of these two options. A method-based application is deployed with the following five live counters: Internet traffic, Tweets sent, Google searches, Emails sent and Tumblr posts, to measure the dynamics of the overall trend in user activity. Four parts of a day and three categories of days are examined as corresponding discrete inputs. In the search for the most effective web advertisement, the displaying strategy will vary directly with the level of users’ activity in the form of live counters over 12 months of 2018. Two competent surveys consolidate the outcome of our work and demonstrate how the company can identify the best web advertising strategy closer to his business needs and interests. We conclude with the discussion whether the considered method can lead to superior efficiency level of all other communication activities.

Amadeusz Lisiecki, Dariusz Król

An Approach Using Linked Data for Open Tourist Data Exploration of Thua Thien Hue Province

In five domains of Digital Transformation, data plays an important role as a key tangible asset for value-creation. Nowadays, data is continuously generated everywhere, even the unstructured data is increasingly becoming usable and valuable. The challenge of data is turning it into valuable information and unleash its potentials for Digital Transformation. Motivated by Open Data initiatives in over the world which took the lead for publishing government data for public use, Thua Thien Hue Open Data Initiative (TTHODI) (Thua Thien Hue Open Data Portal: http://opendata.thuathienhue.gov.vn/ ) took first steps in public its local government data to create data-driven benefits of community and enterprise in the region. On this trend, Linked Data is naturally coming at the right time for us as an ideal standard for representing our open datasets thanks to its expressiveness in form of triples. Furthermore, with huge potentials for linking our open datasets to existing datasets in Linked Open Data Cloud for enriching our data cloud for an added value to TTHODI. This paper presents our approach using Linked Data for Tourist Open Data as a case study and also shows the first results of using Linked Open Data for TTHODI in Tourism.

Hoang Bao Hung, Hanh Huu Hoang, Le Vinh Chien

A Literature Review on Dynamic Pricing - State of Current Research and New Directions

In recent years, the topic of dynamic pricing has drawn a significant attention. Many different scientific communities have studied demand estimation techniques and price policies, committing greatly to its development. This paper surveys latest literature streams, providing a brief overview of research background and state of current works, identifying most active sub-fields and emerging directions. It focuses mostly on computer science and operations literature, however it also discusses contributions from management, economics and marketing fields. This study goal is to provide scientific community, as well as business sector with brief summary of latest research efforts.

Karol Stasinski

Intelligent Modeling and Simulation Approaches for Games and Real World Systems

Frontmatter

Sentiment Analysis by Using Supervised Machine Learning and Deep Learning Approaches

With the growth of online review sites, we can use the opportunity to find out what other people think. Sentiment analysis is a popular text classification task in the data mining domain. In this research, we develop a text classification machine learning solution that can classify customer reviews into positive or negative class by predicting the overall sentiment of the review text on a document level. For this purpose, we carry out the experiment with two approaches, i.e. traditional machine learning approach and deep learning. In the first approach, we utilize four traditional machine learning algorithms with TF-IDF model using n-gram approach. These classifiers are multinomial naive Bayes, logistic regression, k-nearest neighbour and random forest. Out of these four classifiers, logistic regression achieves the highest accuracy. The second approach is to utilize deep learning methodologies with the word2vec approach, for which we develop a sequential deep learning neural network. The accuracy we achieve with deep learning is much lower than our traditional machine learning approach. After finding out the best performing approach and the classifier, the next step of the work is to build our final model with logistic regression using some advanced machine learning methodologies, i.e. Synthetic Minority Over-sampling for data balancing issues, Shuffle Split cross-validation. The accuracy of the final logistic regression model is approximately 87% which is 3% higher from the initial experimentation. Our finding in this research work is that, in smaller dataset scenarios, traditional machine learning would outperform deep learning models in terms of accuracy and other evaluation metrics. Another finding in this work is, by addressing data balancing issues in the dataset the accuracy of the model can be improved.

Saud Naeem, Doina Logofătu, Fitore Muharemi

EEG Based Source Localization and Functional Connectivity Analysis

Nowadays Electroencephalography (EEG) is one of the most attractive Brain-Computer Interfaces (BCI) models to analyze brain signals source localization and connectivity estimation. Unlike other neuroimaging modalities such as fMRI, MEG, and PET; EEG has its higher temporal resolution that senses EEG is currently interesting area for many BCI researchers. However, the precise results of source localization and connectivity are challenging problems that mostly depend on the head models and inverse modeling methods. This paper focused on source localization and functional connectivity analysis using EEG signal over single-trial movement imaginary (MI) tasks by using brainstorm toolbox. Data obtained from the nature dataset that was recorded from 12 subjects, 29 recordings with 19 channels EEG device and MATLAB software utilized for the task.

Soe Myat Thu, Khin Pa Pa Aung

Fitness Function Design for Neuroevolution in Goal-Finding Game Environments

Recently, games like Pac-Man have been hotbeds for neuroevolution research and NEAT has emerged as one of the leading techniques in the game playing domain [14] . In the context of the snake game, the goal of this paper is to enhance neuroevolution strategies with better fitness functions for effective goal finding. We develop greedy and non-greedy fitness functions, and demonstrate the effectiveness of these functions in both environments with and without dynamic obstacles. We then present an alternate implementation using the NEAT algorithm combined with Novelty Search to increase the genetic diversity of the agent population and explore the problem space without specifying direct objectives. These conclusions suggest that even with a low number of simple inputs, and simple fitness functions, agents are quickly able to achieve a novice amount of expertise in the Snake game using NEAT.

K. Vignesh Kumar, R. Sourav, C. Shunmuga Velayutham, Vidhya Balasubramanian

An Application of Machine Learning and Image Processing to Automatically Detect Teachers’ Gestures

Providing teachers with detailed feedback about their gesticulation in class requires either one-on-one expert coaching, or highly trained observers to hand code classroom recordings. These methods are time consuming, expensive and require considerable human expertise, making them very difficult to scale to large numbers of teachers. Applying Machine Learning and Image processing we develop a non-invasive detector of teachers’ gestures. We use a multi-stage approach for the spotting task. Lessons recorded with a standard camera are processed offline with the OpenPose software. Next, using a gesture classifier trained on a previous training set with Machine Learning, we found that on new lessons the precision rate is between 54 and 78%. The accuracy depends on the training and testing datasets that are used. Thus, we found that using an accessible, non-invasive and inexpensive automatic gesture recognition methodology, an automatic lesson observation tool can be implemented that will detect possible teachers’ gestures. Combined with other technologies, like speech recognition and text mining of the teacher discourse, a powerful and practical tool can be offered to provide private and timely feedback to teachers about communication features of their teaching practices.

Josefina Hernández Correa, Danyal Farsani, Roberto Araya

The Effect of Teacher Unconscious Behaviors on the Collective Unconscious Behavior of the Classroom

Normally teachers can consciously control to a great extent the behaviors of their students in the classroom. But additionally, there are unconscious teacher behaviors that also impact the collective behavior of their students. To study this phenomenon, we gather data obtained from mini video cameras mounted on eyeglasses worn by fourth graders. We found that the proportion of scenes where the teacher is pointing his body toward the student is higher than the proportion of scenes when there is mutual gaze, and that this effect is slight pronounced in STEM classes. We also found that this effect is greater among boys than girls, and that is particularly evident at certain distances. More precisely, we found that in STEM classes when a male student is observing the teacher, the teacher is generally pointing their body toward the student (67% of cases). However, with female students, this number is just 46%. However, there is no such difference in non-STEM classes. Moreover, the distance between the student and the teacher also has a significant effect. This is a powerful tool for teachers as it can help them reflect on their strategies, as well as the impact of their unconscious nonverbal behavior in classroom behavior.

Roberto Araya, Danyal Farsani

Experience Enhanced Intelligence to IoT

Frontmatter

Situational Awareness Model of IoV Based on Fuzzy Evaluation and Markov Chain

With the rapid development of the automobile industry, and the continuous advancement of the Internet of Things technology, the emergence of the Internet of Vehicles (IoV) has also led to the security of the IoV. However, scholars have focused on the security research of the components of the IoV, or Research on the safety analysis of the vehicle interior network, or research on vehicle behavior safety. There is almost no research on the overall situation of the entire vehicle at home and abroad. This thesis first analyzes the classic situational awareness model, puts forward the ideal function that the IoV situational awareness should have, and proposes an IoV situational awareness model based on the fuzzy evaluation, AHP hierarchical analysis and Markov chain according to the current situation of the Internet of Vehicles. And design experiments to verify its feasibility.

Pengfei Zhang, Li Fei, Zuqi Liao, Jiayan Zhang, Ding Chen

A Framework for Enhancing Supplier Selection Process by Using SOEKS and Decisional DNA

Supplier selection process is one of the significant stages in supply chain management for industrial manufactured products. It plays an integral role in the success of any manufacturing organization and is an important part starting right from selecting raw material to dispatch of finished products. This paper contributes to enhance the supplier selection process by proposing a multi-criteria decision making framework for industrial manufactured products. The proposed framework is based on smart knowledge management technique called Set of experience knowledge structure (SOEKS) and Decisional DNA, which makes the proposed approach dynamic in nature as it updates itself every time a decision is taken.

Muhammad Bilal Ahmed, Cesar Sanin, Edward Szczerbicki

An Efficient Approach for Improving Recursive Joins Based on Three-Way Joins in Spark

In the evolution of Big Data, efficiently processing large datasets is always a top concern for researchers. A join operation is one of such processing, a common operation appearing in many data queries. This operation generates plenty of intermediate data and data transmission over the network, especially a recursive join operation. Although extremely expensive, a recursive join has a wide variety of domains as database, social network and computer network analyses, compiler, data integration and graph mining. Therefore, this study was carried out to optimize recursive joins based on some solutions in a Spark environment. The solutions leverage the advantages of three-way join operations, Bloom filters, Spark RDD and caching techniques for iterative join computation. These significantly reduce the number of executed iterations and jobs, the amount of redundant data, and remotely accessing persistent data. Our experimental results show that the optimized recursive join is more efficient than a typical one by reducing the number of iterations to half, minimizing data transfer, and thus shorter execution time.

Thanh-Ngoan Trieu, Anh-Cang Phan, Thuong-Cang Phan

Lambda Architecture for Anomaly Detection in Online Process Mining Using Autoencoders

The analysis of event data in the context of process mining is becoming increasingly important. In particular, the processing of streaming data in the sense of an real-time analysis is gaining in relevance. More and more fields of application are emerging in which an operational support becomes necessary, i.e. in surgery or manufacturing. For proper analysis a cleanup of the streaming data in a pre-processing step is necessary to ensure accurate process mining activities. This paper presents a blueprint of a lambda architecture in which an autoencoder is embedded that is supposed to allow unsupervised anomaly detection in event streams, like incorrect traces, events and attributes, and thus will help to improve results in online process mining.

Philippe Krajsic, Bogdan Franczyk

Data Driven IoT for Smart Society

Frontmatter

Biomedical Text Recognition Using Convolutional Neural Networks: Content Based Deep Learning

Named Entity Recognition (NER) targets to automatically detect the drug and disease mentions from biomedical texts and is fundamental step in the biomedical text mining. Although deep learning has been successfully implemented, the accuracy and processing time are still major issues preventing it from achieving NMR. This research aims to upgrade the accuracy of classification while decreasing the processing time, by paying more attention to significant areas of NMR. The novel proposed system consists of a Bi-Directional Long Short-Term Memory with Conditional Random Field (BiLSTM-CRF) using dropout strategy to effectively prevent overfitting and enhancing the generalization abilities. The system built includes the attention mechanism and attention fusion for redistributing the weight of samples belonging to each class in order to compensate the problem occurring from data imbalance and to focus only on the critical areas of the observed things and ignoring non-critical areas.

Sisir Joshi, Abeer Alsadoon, S. M. N. Arosha Senanayake, P. W. C. Prasad, Abdul Ghani Naim, Amr Elchouemi

Pattern Mining Predictor System for Road Accidents

Road traffic accidents are among the major concerns that are leading for deaths and injuries in the world. Many predictive models use data mining technique to provide semi optimal solutions. Pattern identification and recognition have been used to for road accident predictions based on the critical features extracted depending on the dangerous locations and frequency of occurrences prone for accidents. The aim of this research is to propose a novel predictive model based on the pattern mining predictor which improves the accuracy of accident prediction in frequent accident locations. The proposed system consists of association rules mining technique, which identifies the correlation, frequent pattern and association among the various attributes of the road accident. Clustering technique that discriminates the data based on different patterns and classification technique that classify and predicts the severity of accident. Novel system built leads to an improvement in the accuracy of the accident prediction from 92% to 94%. Furthermore, using selective subset of features decreased the processing time and precision of classification is improved using boosting technique.

Sisir Joshi, Abeer Alsadoon, S. M. N. Arosha Senanayake, P. W. C. Prasad, Shiaw Yin Yong, Amr Elchouemi, Trung Hung Vo

Artificial Neural Network Approach to Flood Forecasting in the Vu Gia–Thu Bon Catchment, Vietnam

Flooding in the Vu Gia-Thu Bon catchment has destroyed critical facilities, such as infrastructure and housing. This study develops an application of an Artificial Neural Network (ANN) to forecast the flow at the Nong Son gauging station in the catchment. The ANN model uses rainfall data at upstream locations to estimate flows at downstream point. Architectures of the ANN model are adjusted to calculate flooding. Daily rainfall at Tra My, Tien Phuoc, Hiep Duc and Nong Son between 1991 and 2010 is used to predict flooding at Nong Son. The analysis shows that the ANN is a reliable method to forecast the flood in the Vu Gia-Thu Bon catchment, where there is a lack of a wide range of accurate data, particularly hydrological, meteorological and geological data.

Duy Vu Luu, Thi Ngoc Canh Doan, Ngoc Duong Vo

Ensuring Comfort Microclimate for Sportsmen in Sport Halls: Comfort Temperature Case Study

The opinion about the difficulties of maintaining appropriate climatic conditions at sports facilities is quite popular among athletes of various disciplines. This problem can lead to health and economic problems. First of all, users of the facilities are not provided with proper conditions for sports, which can lead to bruises. Secondly, the cost of using the object increases. With the increasing number of sports facilities, maintaining internal thermal comfort in them, while ensuring low operating costs, is becoming increasingly important. The article presents the work on modeling the microclimate in a multifunctional sports hall with the most maximum mode of its use and a detailed analysis of the maintenance of thermal comfort in the hall and cognitive-utilitarian conclusions.

Bakhytzhan Omarov, Bauyrzhan Omarov, Abdinabi Issayev, Almas Anarbayev, Bakhytzhan Akhmetov, Zhandos Yessirkepov, Yerlan Sabdenbekov

Applications of Collective Intelligence

Frontmatter

Smart Solution to Detect Images in Limited Visibility Conditions Based Convolutional Neural Networks

Decrease in visibility causes many difficulties in vision, tracking. Current classic object detection techniques do not give satisfying results in less visibility. It is essential to detect and recognize the objects under such conditions and devise a better object detection mechanism. The paper proposes a solution to this problem by using a multi step approach that uses Saliency techniques and modern object detection algorithms to obtain the desired results. The distorted image is enhanced via a deep neural network for visibility enhancement. The image frame of a better quality undergoes saliency techniques so that less visible objects are visible. Faster Region-based Convolutional Neural Network (R-CNN) then runs on the saliency output to yield bounding boxes for all the objects. The coordinates of the bounding boxes are then applied on the original image thus detecting all the objects in a distorted image with less visibility.

Ha Huy Cuong Nguyen, Duc Hien Nguyen, Van Loi Nguyen, Thanh Thuy Nguyen

Experience Report on Developing a Crowdsourcing Test Platform for Mobile Applications

Crowdsourcing-based testing is a recent approach where testing is operated by volunteer users through the cloud. This approach is particularly suited for mobile applications since various users operating in various contexts can be involved. In the field of software engineering, crowd-testing has acquired a reputation for supporting the testing tasks, not only by professional testers, but also by end users. In this paper, we present TMACSTest (Testing of Mobile Applications using Crowdsourcing). This platform provides the important features for crowdsourcing testing of mobile apps by means of the following functionalities: It allows mobile app providers to register and upload mobile apps for testing, and it allows volunteering Internet users to register and test uploaded mobile apps. Expected behavior is that uploaded mobile apps are tested by many different Internet users in order to cover different runtime platforms and meaningful geographical locations.

Nguyen Thanh Binh, Mariem Allagui, Oum-El-Kheir Aktouf, Ioannis Parissis, Le Thi Thanh Binh

Vision Based Facial Expression Recognition Using Eigenfaces and Multi-SVM Classifier

Facial Expression Recognition (FER) has become one of the most popular areas of research in computer vision and biometrics authentication and it has achieved a lot of enthusiasm from researchers. The Vision based Facial Expression Recognition system intends to classify the facial expression of a given image. In this paper, the proposed system automatically classifies the facial expression. The system is composed of feature extraction and expression classification. In preprocessing, Hybrid filter (Median and Gabor) and Histogram Equalizations, is used to reduce noise and enhance images. Feature extraction is to extract feature vectors from face images using the Eigenfaces approach, based on Principal Component Analysis (PCA). To classify facial expression, extracted feature vectors are fed into a Multiclass Support Vector Machine (Multi-SVM) classifier. Experiments are performed on the standard dataset of the Japanese Female Facial Expression (JAFFE) and achieved 80% accuracy. The proposed system showed satisfying performance comparing with other methods and effects state-of-the-art performance on the JAFFE dataset.

Hla Myat Maw, Soe Myat Thu, Myat Thida Mon

An Effective Vector Representation of Facebook Fan Pages and Its Applications

Social networks have become an important part of human life. There have been recently several studies on using Latent Dirichlet Allocation (LDA) to analyze text corpora extracted from social platforms to discover underlying patterns of user data. However, when we wish to discover the major contents of a social network (e.g., Facebook) on a large scale, the available approaches need to collect and process published data of every person on the social network. This is against privacy rights as well as time and resource consuming. This paper tackles this problem by focusing on fan pages, a class of special accounts on Facebook that have much more impact than those of regular individuals. We proposed a vector representation for Facebook fan pages by using a combination of LDA-based topic distributions and interaction indices of their posts. The interaction index of each post is computed based on the number of reactions and comments, and works as the weight of that post in making of the topic distribution of a fan page. The proposed representation shows its effectiveness in fan page topic mining and clustering tasks when experimented on a collection of Vietnamese Facebook fan pages. The inclusion of interaction indices of the posts increases the fan page clustering performance by 9.0% on Silhouette score in the case of optimal number of clusters when using K-means clustering algorithm. These results will help us to build a system that can track trending contents on Facebook without acquiring the individual user’s data.

Viet Hoang Phan, Duy Khanh Ninh, Chi Khanh Ninh

Natural Language Processing

Frontmatter

Wordnet – a Basic Resource for Natural Language Processing: The Case of plWordNet

This paper presents a wide scope of wordnet applications on the example of applications of plWordNet – a wordnet of Polish. Wordnets are large lexical-semantic databases functioning as primary resources for language technology. They are machine-readable dictionaries. Thus, they are indispensible for tasks such as basic flow of text processing, text mining, word sense disambiguation, information extraction and retrieval. On a larger scale, wordnets are used in research, education and business. In this paper a few examples of specific plWordNet applications are described in detail.

Agnieszka Dziob, Tomasz Naskręt

KEFT: Knowledge Extraction and Graph Building from Statistical Data Tables

Data provided by statistical models are commonly represented by textual, tabular or graphical form in documents (reports, articles, posters and presentations). These documents are often available in PDF format. Even though it makes accessing a particular information more difficult, it is interesting to process the PDF documents directly. We present KEFT, a solution in the statistical domain and we describe the fully functional pipeline to constructing a knowledge graph by extracting entities and relations from statistical Data Tables. We showcase how this approach can be used to construct a knowledge graph from different statistical studies.

Rabia Azzi, Sylvie Despres, Gayo Diallo

Devising a Cross-Domain Model to Detect Fake Review Comments

The online reviews not only have huge impact on consumer shopping behavior but also online stores’ marketing strategy. Positive reviews will have positive influence for consumer’s buying decision. Therefore, some sellers want to boost their sales volume. They will hire spammers to write undeserving positive reviews to promote their products. Currently, some of the researches related to detection of fake reviews based on the text feature, the model will reach to high accuracy. However, the same model test on the other dataset the accuracy decrease sharply. Relevant researches had gradually explored the identification of fake reviews across different domains, whether the model built using comprehensive methods such as text features or neural networks, encountering the decreasing of accuracy. On the other hand, the method didn’t explain why the model can be applied to cross-domain predictions. In our research, we using the fake reviews and truthful reviews from Ott et al. (2011) and Li, Ott, Cardie, and Hovy (2014) in the three domain (hotel, restaurant, doctor). The cross-domain detect model based on Stimuli Organism Response (S-O-R) combine LIWC (Linguistic Inquiry and Word Count), add word2vec quantization feature, overcoming the decreasing accuracy situation. According to the research result, in the method one SOR calculation of feature weight of reviews, the DNN classification algorithm accuracy is 63.6%. In the method two, calculation of frequent features of word vectors, the random forest classification algorithm accuracy is 73.75%.

Chen-Shan Wei, Ping-Yu Hsu, Chen-Wan Huang, Ming-Shien Cheng, Grandys Frieska Prassida

Low Resource Languages Processing

Frontmatter

Towards the Uzbek Language Endings as a Language Resource

The Uzbek language belongs to low-resource languages. It is very important to increase the number of language resources such as dictionaries, corpora (monolingual and bilingual) for the Uzbek language. Dictionaries may be different kinds: monolingual, orthographical, bilingual, grammar special dictionaries: stems dictionaries, affixes dictionaries, etc. For different NLP tasks of agglutinative languages, such as morphological analysis, information retrieval, machine translation (segmentation preprocessing) in some cases needs a dictionary of words’ endings. In this paper, we proposed the first electronic dictionary of Uzbek words’ endings in variants for morphological segmentation preprocessing useful for neural machine translation. The resource analysed by the initial version of the Lexicon free stemming tool [3] created by authors. For creation of Uzbek words’ endings’ electronic dictionary, it was used a combinatorial approach inferring apply for part of speech of the Uzbek language: nouns, adjectives, numerals, verbs, participles, moods, voices.

Sanatbek Matlatipov, Ualsher Tukeyev, Mersaid Aripov

Inferring the Complete Set of Kazakh Endings as a Language Resource

The Kazakh language belongs to low-resource languages. For application of actual modern branches as artificial intelligence, machine translation, summarization, sentiment analysis, etc. to the Kazakh language needs increasing the number of electronic language resources. Although neural machine translation (NMT) has shown impressive results for many world languages, it does not solve the problem of low-resource languages. Therefore, the development of resources and tools perfecting the use of NMT for low-resource languages is relevant. For perfect use of NMT for the Kazakh language needs bilingual parallel corpora, but also needs a perfect method of the segmentation source text. By the opinion of authors, one of the effective ways for source text segmentation is morphological segmentation. The authors propose to use for morphological segmentation of Kazakh text a table of a complete set of Kazakh words’ endings. In this paper is described the inferring of the complete set of Kazakh words’ endings. Development of the table of the complete set of word’ endings of the Kazakh language will allow in one-step (by reference to the table of endings of the language) to perform the segmentation of the word’s ending into suffixes. The complete set of endings of the Kazakh language allows guaranteeing the analysis of any word of the Kazakh language, as this is determined by the inferring of the complete system of words’ endings of the language.

Ualsher Tukeyev, Aidana Karibayeva

A Multi-filter BiLSTM-CNN Architecture for Vietnamese Sentiment Analysis

Feedback is information about reactions to a product or a person’s performance of a task. It is powerful as it serves as a guide to assist people to know how others perceive their performance and helps them meet standards. This paper concentrates on the use of natural language processing and deep learning. It combines the advantages of these approaches to perform sentiment analysis on student and customer feedback. Furthermore, word embedding is also applied to the model to add complementary effectiveness. The preliminary findings show that the use of BiLSTM-CNN–the first to catch the temporary information of the data and the second to extract the local structure thereof–outperformed other algorithms in terms of the F1-score measurement, with 93.55% for the Vietnamese Student’s Feedback Corpus (VSFC) and 84.14% for the Vietnamese Sentiment (VS). The results demonstrate that our method is an improvement compared to the best previously proposed methods on the two datasets.

Lac Si Le, Dang Van Thin, Ngan Luu-Thuy Nguyen, Son Quoc Trinh

Computational Collective Intelligence and Natural Language Processing

Frontmatter

Causality in Probabilistic Fuzzy Logic and Alternative Causes as Fuzzy Duals

Causal inference is an essential part of human reasoning. Computer scientists often use inferential logic and probability theories to find the causes of the events. In causal inference, researchers are interested in finding the relationship between two observable events. However, most of the time, we can only perceive the co-occurrences of events or usual successions of events. So, we must be careful in trying to establish causality between events. Also, many alternative causes or conjoint causes can be at work. So, we need good tools to deal with the complex problem of causality. In this paper, we will explore probabilistic fuzzy logic (PFL) as a sophisticated alternative to establish causal inference and fuzzy duals as tools for the expression of alternative possible causes.

Serge Robert, Usef Faghihi, Youssef Barkaoui, Nadia Ghazzali

Enhance Trend Extraction Results by Refining with Additional Criteria

Traders make their investment based on stock trends or price directions to get more profit. Many researchers have tried to retrieve the interesting features for trends on large financial data such as news. Data obtained from stock market is highly volatile and correlated. It is characterized with high dimensionality to make prediction of stock trends a challenging. Feature extraction is an important part in developing a fully automated stock market prediction system. Features in stock news have Multiple Interdependent Nature (MIN) which are the relationship between two or more features. Rule and Syntactic Feature based relation extraction have already proposed to get trend related features. Our previous work is still required to solve MIN for stock trend prediction. This paper proposes the additional criteria to improve the trend extraction on MIN. These criteria are date conversion, main verb identification, stock reference and advanced trend extraction. According to the experiment, the trend extraction with additional criteria on stock news gets more extracted trends than the extraction without criteria. The stock trend extraction results with additional criteria become more consistent with the real stock price movement.

Ei Thwe Khaing, Myint Myint Thein, Myint Myint Lwin

Image Captioning in Vietnamese Language Based on Deep Learning Network

Image captioning is an underlying and crucial problem in artificial intelligence. This is challenging since it requires advanced research in computer vision to detect objects and the correlation of these objects in the image, and text-mining to convert these relationships into words. Although some based deep-learning and machine translation approaches have been achieved state-of-art results in English recently, it is missing an approach to generate the caption in Vietnamese, which is a local language in Vietnam with complex grammar and variable meaning in simple words. Moreover, machine translation is affected negatively by a significant issue called unknown words, which is caused by both large vocabulary size and unbalanced dataset. In this paper, we propose a new approach to generate the Vietnamese caption of the image and also a simple and effective solution to tackle unknown words problem of machine translation. In general, the results of these methods achieved in the self-build testing database are promising.

Ha Nguyen Tien, Thanh-Ha Do, Van-Anh Nguyen

Textual Clustering: Towards a More Efficient Descriptors of Texts

In this paper, we present how itemsets can be used as a discriminating descriptor in a textual clustering process. We implemented a platform named “IDETEX” capable of extracting itemsets from textual data and using them for the experimentation in different types of clustering methods, such as K-Medoids, Hierarchical clustering and Self-Organizing Map (Kohenon). To some extent the experimentations performed reveal promising results with different classifiers either “Hierarchical”, “Non-hierarchical” or “Neural network”.

Ayoub Bokhabrine, Ismaïl Biskri, Nadia Ghazzali

Artificial Intelligence in Detecting Suicidal Content on Russian-Language Social Networks

Due to the anonymity of online media and social networks, people tend to Express their feelings and suffering in online communities. To prevent suicides, it is necessary to detect messages about suicides and user perceptions of suicides in cyberspace using natural language processing methods. We focus on the social network Vkontakte and classify users’ messages with potential suicide and without suicidal risk using text processing and machine learning methods.In this paper, we tell about suicidal and depressive ideation detection in Russian Language. For this purpose, we create a dataset that consists of 64,000 posts that collected from Russian language social network Vkontakte. The dataset was applied to eight machine learning algorithms.

Sergazy Narynov, Kanat Kozhakhmet, Daniyar Mukhtarkhanuly, Aizhan Sambetbayeva, Batyrkhan Omarov

Backmatter

Weitere Informationen

Premium Partner

    Bildnachweise