Search

A fuzzy creative works generation algorithm based on graph neural network is proposed. Firstly, the multi-label fuzzy creative data set is constructed. Secondly, fuzzy logical correlations between creative objects are dynamically calculated by graphical neural networks to capture relevant digital objects and their relationships. Thirdly, the projectiles, boundary markers and location words of the creative scene objects are generated by analyzing related attributes of each entity based on graph neural network. Fuzzy creative works are automatically generated by adjusting the appropriate location. Finally, the experiments are deployed, and experimental results show that the accuracy rate of the proposed algorithm is improved by 9.6% on average compared with the manual selection, which indicates the feasibility of the proposed algorithm replacing the manual selection.

Machine translation and extracting meaningful language descriptions has always been a challenging task in the computer vision field. This paper presents a very modern approach which uses deep learning techniques to generate image description. The concept of Capsule Neural Network is applied in Image Captioning. A capsule denotes a nested layer inside the capsule. One major difference between traditional networks and Capsule Neural Network is that the Capsule Neural Network handles better segmentation and recognition. Evaluation of the model on Flickr 8 k data-set qualitatively and quantitatively shows the proper learning and language efficiency. For model evaluation a very popular metrics BLEU score has been considered which shows significant improvement from 28.9 to 37.8 for the latest Flickr 8 k data set.

This study aims to design a complex conformal cooling channel (CCC) structure for an injection mold geometry that has eight different design variables according to multiple main objectives of the maximum temperature at the plastic interface, the difference between the maximum and minimum temperatures at the internal wall, and the pressure drop in the channel. The high number of design variables results in an unaffordable computational load for full factorial design. Hence, we present the hybrid Latin Hypercube (LHC)-Delaunay Triangulation (DT) method to create meta-model data effectively at high dimensions by using 134 different simulations in the MATLAB environment. Since the computational time is a crucial factor, we perform the hybrid LHC-DT approach with four different data simulation set of 1–30, 1–60, 1–90, and 1–134. The multi-dimensional decision-making procedure is completed with multiobjective optimization. The proposed hybrid LHC-DT approach achieves reliable and apparent results for all multiple objectives so that the Pareto frontier of the data set of 1–30 is selected for assessment. According to the data set of 1–30, the optimum values of the temperature difference, maximum temperature, and the pressure drop are found 28.88 K, 361.35 K, and 0.46 bar, respectively.

Over the last few decades, heart-related diseases or cardiovascular diseases (CVDs) are considered to be the deadly disease which affects both men and women not only in India, as well as across the world. Diagnosing heart disease in the health care field can be a challenging task and many researchers devoting their time and knowledge to develop intelligent clinical decision support systems to improve the ability of the clinicians. Out of many machine learning techniques, classification is the most powerful technique which is commonly used for prediction. The rapid growth of complex data about patient medical records is the main source of discovering hidden knowledge for error-free decision-making. The main purpose of this study is to assist non-specialists in diagnosing heart disease risk limits and in diagnosing heart disease early. This paper introduces pre-processing models that have been applied on the University of California, Irvine (UCI) repository—Cleveland heart disease database for examining the accuracy and the error rate on predictions. The results obtained are compared and then verified using the Jupyter Notebook interface with the Scikit-Learn Library in PYTHON. In order to validate the efficacy of the proposed model, 75% training–25% testing is performed on the data set. Our analysis over few classification algorithms shows the efficacy of our proposed pre-processing models.

Rainfall is the most important factor that affects Kerala’s economy. Kerala is a highly populated state when compared to its availability for water. Therefore, the need for a large amount of water can be compromised by proper management of rainfall. The main objective of this paper is to compare the three classification algorithms—SMOreg, Linear Regression and Multilayer Perceptron for Time series forecasting feature using WEKA tool. For the analysis, a data set containing 116 samples of Kerala’s annual rainfall from the year 1901–2017 that has been collected from a government website. The performance of the algorithms is checked on the basis of Root-Mean-Squared Error. The model shows that the least RMSE seems the better algorithm and found that Multilayer Perceptron is least with it. So, in the three algorithms Multilayer Perceptron is the best followed by Linear Regression and SMOreg.

Pneumonia is the main infectious cause of death in children under five. At present, facing a large number of patients and scarce imaging experts, rapid diagnosis of pneumonia is particularly urgent. As the texture features of pneumonia images are susceptible to noise interference, traditional classification and recognition can no longer meet the needs of existing medical treatment. Therefore, we use the powerful feature extraction capabilities of deep learning to further improve the accuracy of diagnosis. In this article, we first expand and normalize the collected pneumonia data set. Then combine the transfer learning points to retrain the three convolutional neural networks of Vgg-16, Resnet-50, and Xception. Experiments show that the convolutional neural network based on deep learning can efficiently and prepare the diagnosis of pneumonia in children, especially the classification and recognition rate of Vgg-16 is as high as 99.15%. It is of great significance for the efficient diagnosis of childhood pneumonia.

Knowledge-graph-aware recommendation often overlooks the importance of unstructured information, such as text information and picture information. In order to address the problem, we propose KGMR (Knowledge Graph for Movie Recommendation) algorithm which uses a large amount of heterogeneous information provided by the multi-modal knowledge graph to design the semantic types and semantic relations in the movie field. Constructed a knowledge graph in the film domain, and visualized it with Neo4j graph database. The text information and picture information are respectively represented by Doc2Vec and Convolutional Auto-Encode (CAE) algorithms, combined with Bayesian Personalized Ranking (BPR) to construct a recommendation system. Using the MovieLens-latest-small data set for testing, the results show that KGMR has a good improvement in the evaluation value, which proves that the knowledge graph of movies is integrated into the recommendation algorithm, and movies can be accurately recommended to target users.

With the improvement of living standards, the problem of human obesity has been getting worse. It is important to classify human obesity and determine the relevant obesity factors. In this paper, Lasso feature selection method is used for feature selection of data set to further reduce the data dimension. In addition, because the traditional LightGBM algorithm has a certain randomness in parameter selection, it is difficult to determine the optimal combination of parameters. This paper uses genetic algorithm to optimize the parameters of LightGBM algorithm. It is worth mentioning that the use of data standardization reduces the runtime of LightGBM. The LightGBM based on GA optimization compared with other common machine learning algorithms obtains good results, compared with the traditional LightGBM algorithm, the average accuracy and the average runtime are improved by 0.5% and decreased by 72.12% respectively.

In recent years, artificial intelligence technology has been widely used in the field of image classification of pneumonia. Although convolutional neural networks with deeper layers and more complex structures can classify images of pneumonia more accurately, it also brings problems such as network degradation and complex structure, which often require excessive system resources. Therefore, according to the characteristics of pneumonia images, this paper proposes an optimized residual network classification method. First, according to the characteristics of pneumonia medical images, the ResNet model with better classification effect in the deep network is selected, and the pre-processed pneumonia data set is used for sparse pre-training. Second, the channels in the network are pruned and P-ResNet is obtained model. Finally, the experimental results show that the P-ResNet18 model obtained after the pruning comparison in this paper is only 0.48% lower than the accuracy of P-ResNet50 by 0.905%, and the model parameters are reduced by 54.46%, which is much greater than that of P-ResNet50 and P-ResNet50. 47.78% and 45.72% of P-ResNet101, the reasoning time is reduced by 5.29%. The P-ResNet18 model can simultaneously ensure rapid and accurate preliminary screening and labeling of a large number of pneumonia chest radiographs.

In view of the low detection accuracy of YOLOv3_tiny algorithm on small pedestrian target, a pedestrian detection method to improve YOLOv3_tiny is proposed. Firstly, the head and shoulder of the pedestrian are taken as the detection object. The K-means++ algorithm is used to improve the k-means algorithm and cluster the data set to get the anchor frame with higher accuracy. Secondly, add a target prediction layer with a resolution of 52 × 52 to the multi-scale prediction section, which improves the detection accuracy of the small target. In addition, the backbone structure of YOLOv3_tiny is light-weighted, which further compresses the model memory and reduces the deployment burden of the model while maintaining accuracy. A comparative experiment is conducted on the self-made pedestrian dataset which tags the head and shoulder. The results show that the recall of the improved algorithm is 0.8743, which is 11.78% higher than that of the original algorithm; The average accuracy is 77.69%, which is better than the classical algorithms, such as SSD, fast RCNN and Yolo v3; the model size is 26M, and the test time of single image is 0.0789s, which can achieve efficient and real-time detection.

The graph neural network (GNN) can significantly improve the performance of few-shot learning due to its ability to automatically aggregate sample node information. However, many previous GNN works are sensitive to noise. In this paper, a few-shot image classification algorithm (Proto-GNN) based on the prototypical graph neural network is presented. First, convolutional neural network (CNN) is used to obtain the feature vectors of the support set samples, which can be used to calculate the prototype of each category. Then, the feature vectors of the support set, prototype vectors and the feature vectors of the query set are used to form a completely connected graph. Finally, Proto-GNN is built and trained. Due to the addition of the prototype vectors, the query set samples obtain more category information, which is more conducive to classification. The experimental results show that the model in this paper has a better classification performance on the Omniglot data set.

Estimation of recovery rate of COVID-19 positive persons is significant to measure the severity of the disease for mankind. In this work, prediction of the recovery rate is estimated based on machine learning technology. Standard data set of Kaggle has been used for experimental purpose, and the data sets of COVID cases in Italy, China and India for these countries are considered. Based on that data set and the present scenario, the proposed technique predicts the recovery rate.

Gesture recognition is a technology that uses optical character recognition to identify gestures, hand movements, etc. It has seen applications on various platforms like VR, AR, gaming consoles (Xbox Kinect), etc. The game of hand cricket has been very popular among school-going children. The game will allow the person to play a game of hand cricket against the computer. The system will use gesture recognition in real-time similar to that found in gaming consoles. Libraries like OpenCV help in real-time computer vision, convolutional neural network or CNN help in analysing the image. In our proposed work, we have developed a model that could recognize the gestures of a user correctly. To achieve this, we have implemented a classification model using an image classifier which groups a linear stack of layers into a Keras model. We have created our data set which consists of 1800 images in total. For every gesture, we have taken the 300 images using the webcam and converted those images into the binary image which consists of pixels that have only two colours.

Behavior recognition is an important research topic in the field of artificial intelligence and computer vision. In daily life, smart devices with behavior recognition capabilities are widely used in video retrieval, video surveillance, human-computer interaction and other fields, and they play an important role in the landscape. This paper studies the concept, function and application of dynamic recognition technology. At the same time, the basketball action recognition is summarized, and the saliency detection algorithm is proposed. The results show that the ucf-101 data set of 25 iterations has the largest proportion, 23.4%; the hmdb-51 data set of 10 iterations has the smallest proportion, 23.6%. As the number of iterations increases, the accuracy continues to improve. Finally, based on stability, the number of iterations basically no longer increases. At the same time, considering the issue of training time, the final number of iterations is selected as 10 times.

Medical practitioners depend on medical diagnosis systems for detection, diagnosis, and treatment of various diseases in recent years. Genetic algorithms play a vital role as an essential optimization approach for problems involving classification in machine learning. Genetic algorithms can also achieve a high level of prediction and accuracy. Coronary heart disorder is a major heart disorder that narrows the blood vessels that supply oxygen to the heart. In this paper, we analyze and predict heart diseases among patients using genetic algorithms. The heart disease data set from the UCI machine learning repository data set is used. The proposed method utilizes the data set on heart disease available at the UCI machine learning repository and provides better classification accuracy and prediction among the patients with various heart disorders. Implementation is carried out using Python language.

Skin burn identification is a very critical job to identify the burn location and its impact on the body. The current paper aims with the objective to identify the burn location and its impact so that the severity can be measured to provide effective treatment. The solution is derived using the Machine Learning model using CNN. The treatment can be provided after taking the right direction from the model. The relative features are identified and then based on the model the burn identification and its impact can be identified as well as we can measure the impact on it. The proposed approach based on CNN is tested on a standard burn data set of burns—BIP_US database. Training is done by classifying images into two groups. The test data set images are analyzed using the proposed CNN-based approach and 93% accuracy was achieved for the CNN-based model. The current method can perform better with a state-of-the-art machine learning technique on the burn images. Finally, COCO and BIS dataset is used to check the accuracy of the model. The final result illustrates the performance of the model, which is very effective in terms of accuracy.

Now days, enormous amount of the data are generated and collected from different domains and analyze of all these data manually is tedious task. Therefore, Knowledge Discovery in Databases (KDD) process have to used to find potentially, valid, novel, and explicable patterns in data. Lately, delicate registering turned out to be increasingly more alluring for the specialists, who work in the related research field of data mining. The main objective of this paper is about how to use soft computing model to pull out knowledge from database. The created knowledge base will predict /inference the data for the new data element of same data set. The concepts of data mining pre processing techniques have applied on the data to make the data in suitable form for knowledge extraction. Thereafter, statistical analysis and soft computing techniques have been applied on the information to pick the ideal model for the data set. The choice of the optimal model decided based on the residual analysis and average error of the model. The preferable model has been used to pull out the knowledge from the data mining later on. The objective of the paper is to develop an integrated system to extracting exact information (knowledge) from the data. The quality of extracted information based on the particular model has also been reviewed. Here Iris flowers data have been used for getting knowledge related to quality of Iris flowers based on certain primary information.

Diagnosing and detection of ophthalmology related disorders, extraction of vessels from retinal fundus images is very important task in Medical Science. Locating retinal vessels manually is not only tough and time taking task for ophthalmologist but also it is prone to have human error because of complex nature of blood vessels in Retina. Across the globe different techniques are being applied to overcome this challenge by deploying different strategies and techniques. This paper proposes a new method for extraction of vessels from retinal images. Proposed method first extracts thick vessels from the fundus image. Finally, flood filling algorithm in-conjugation with BFS is applied to discover the tiny blood vessels and reconstruct the final result by combining it with the thick vessels extracted in the first place. Experiment is performed on publicly available DRIVE data set and it is found that proposed method produced quite outstanding results.

Remote sensing image classification has long attracted the attention of the remote‐sensing community because classification results are the basis for many environmental and socioeconomic applications. The classification involves a number of steps, one of the most important is the selection of an effective image classification technique. This paper provides a comparative study of the supervised learning techniques for remote sensing image classification. The study is being focused on classification of land cover and land use. Supervised learning is a branch of machine learning and is used in this study. The comparison is made among the different techniques of pixel-based supervised classification used for remote sensing image classification. The study has been made on a labelled data set. After the implementation, support vector machine has been found to be the most effective algorithm among the five algorithms of pixel-based supervised classification (i.e. maximum likelihood estimation, minimum distance classifier, principal component analysis, isoclustering and support vector machine).

In the current scenario, the classification methods are needed to reduce the possible errors. It will help the physicians to take suitable decisions with speedy manner. Here, the prime motto of this paper is to achieve an efficient classification method for liver disease. So, authors have used random forest, support vector machine and AdaBoost methods on the Indian Liver Patient Disease (ILPD) data set where random forest gives the highest accuracy of 93%. Finally, authors would like to conclude that the proposed classification methods have not improved but sustained the accuracy compared to the existing and could also be implemented in other medical diseases.

With the exponential growth of the internet among users worldwide, network engineers pose a great challenge in network security to identify intrusion activities. Intrusion Detection System (IDS) and Intrusion Prevention System (IPS) are the tools used to defend against these intrusion activities. IDS has sometimes been prone to false alarms. Therefore, to improve IDS, machine learning method is used. The largest number of IDS data sets are available till date where some are unreliable to use, whereas some are out of date and some does not cover common updated attacks. CICIDS-2017 data set overcome above major flaws [1]. Consequently, this paper assesses the performance of CICIDS-2017 data set by applying various machine learning algorithms such as Convolution Neural Network (CNN), Naive Bayes (NB) and Random Forest (RF), RF with highly ranked features, RF with feature reduction techniques (PCA and SVD). Then the comparison study is done which shows Random Forest gives good result when compared with other algorithms.

Image segmentation has proven to be beneficial in many applications, including medical imaging, object detection and scene detection. The future of image segmentation is scene prediction based on the previously segmented images. In this paper, author presents a survey comparing image segmentation architectures based on convolutional neural network (CNN). The paper explains the basic layers of CNN, its depth, general representation and the working of CNN. The paper also discusses the challenges faced by the architectures and their probable solutions. This survey highlights the segmentation network (SegNet). SegNet is a technique of encoder-decoder pair for image segmentation and classification. SegNet architecture is built using CNN and Visual Geometry Group (VGG16) network. VGG16 is a network with 16 deep layers and can classify images into 1000 categories. This architecture resolves the problem of high computational time and high requirement of memory. The comparison of fully convolutional network (FCN) and SegNet is also presented at the end. CamVid dataset is used for testing SegNet for scene segmentation. This data set is captured with sensors shown in SUN RGB-D and implemented using PASCAL VOC challenges on road scene segmentation and indoor scene segmentation. The paper tries to explore how the SegNet architecture works efficiently in case of road scenes and faces accuracy issues with indoor scene segmentation.

Natural disasters are an unpredictable one, but the damages caused by the disaster is severe. It causes hazards to both humans and their properties. Among many hazards like earthquake, eruption and flood, the flood prediction is a quite predictable one. But, it requires proper learning for predicting the floods. Because the flood occurs due to the overflow of water from dams and rivers which is caused by heavy rainfall. Due to this, proper learning based on the weather conditions and previous flood data, the possibility of flood range and area can be detected. Such a flood prediction is performed on the real-time data set collected from the Columbia province. It uses data mining approach on the collected information for forecasting the flood level. But, it is not only sufficient for real time. Hence, the image-based flood prediction is proposed to analyse the flood level in the particular region using a satellite image of the area. Yet, both the techniques offer good prediction in the individual unit it suffers from predicting the nearby flood areas. Hence, in this, a combined approach of image processing and data mining is proposed to forecast the flood level. For data mining, the regression learning approach is used to forecast future flood levels. It is combined with the corresponding image processing of the particular area to accurately the flood level based on extracting the depth of water in the area. Finally, the predicted level is broadcasted to the people through social media for immediate action to save their lives. The proposed regression-based data mining and image analysis help to forecast the level of flood in the area accurately along with broadcasting saves precious lives of people.

Among the chronic nervous system diseases, Parkinson’s disease (PD) is known for its progressiveness in impairing the speech ability, gait as well as complex muscle and nerve actions. Hence an early diagnosis of PD will help in reducing the symptoms. Telemedicine offers a cost-effective and convenient approach, and several studies have used dysphonic features to remotely detect PD. In this study, we have used a data set from Kaggle, which included voice measurements from 31 people of whom 23 were diagnosed with PD. The data set included 22 different attributes pertaining to voice measurements, including the pitch period entropy with 195 voice recordings for each of the individuals. In the data pre-processing, the correlated attributes were removed and we used 10 non-correlating attributes (< 0.7) along with individual status (0 and 1 for healthy and PD, respectively). The data set after pre-processing was split into 70:30 ratio and also ascertained that the number normal versus PD are in equal ratios in both the training and testing data sets, respectively. The data set was evaluated with four different supervised classification machine learning (ML) models, namely random forest, XGBoost, SVM and decision tree. The XGBoost classifier model was found to be highly efficient in precise classification of PD with an accuracy of 0.93.

Bad road conditions, such as cracks and potholes, can cause passenger discomfort, vehicle damage, and accidents. Condition of roads indirectly effects on growth of the country. Hence, there is a need for such a system that can detect potholes. It would allow vehicles to issue alerts to identify potholes so that drivers can reduce the speed and avoid them and make the ride smooth. Many researchers had developed various algorithms to become aware of potholes on roads. In this paper, the proposed system detects the potholes using You Only Look Once version 2(YOLOv2) and a convolutional neural network (CNN). The predefined CNN, namely resnet50, is used to extract the features of testing images and training images. Kaggle data set is used to evaluate the proposed algorithm. The experimental results are evaluated in terms of precision rate and recall rate. The proposed approach precision rate is 94.04% for test images.

The increase in number of disease challenges medical practitioner in making right decisions. As most diseases have the same set of symptoms, the medical practitioner struggles to take decision on recognizing disease as well as right treatment methods. A number of approaches are available for the disease identification and providing treatment, but finding the right approach is what matters. To solve this issue, an inter-/intra-disease impact, the disease-based prediction and recommendation generation method, is presented. The method first reads the input data set and produces a series of clusters with the samples obtained. In the second level, the method estimates inter-disease impact measure and intra-disease impact measure on various disease classes for every data point of the data collection. Using these two measures, the method computes I2-Disease weight for each data point in assigning a label to the data points. For the classification, the method estimates symptomatic disease weight based on inter-/intra-symptom correlation assessment. Based on the selected disease class, a set of treatment samples is populated and ranked according to their curing rate.

Aiming at the problem that the entities in the field of Wa cultural information resources have long length, parallel entities, and no public data set, this paper uses the text information in the Chinese ethnic dictionary Wa nationality volume as the data set, and uses the BERT model to pre-trained the word vector, then extracting the semantic features with the attention mechanism based on the BiLSTM network model, finally, the CRF model is used to predict and output the optimal tag sequence. A method of named entity recognition of Wa cultural information resources based on the attention mechanism is proposed. The experimental results show that the model can effectively identify the entities in the Wa cultural information resources and alleviate the problem of inconsistent entity labels. The recognition accuracy, recall rate, and F value of the Wa cultural information resources corpus are 92.67%, 90.06%, and 91.34% respectively.

Traditional data mining technology is based on data repository and related data to calculate data, and find data relationship and data model. Therefore, when processing cloud computing platform network data, it will consume a lot of storage resources and computer resources, which not only has low mining efficiency, but also has high material and software costs, which to a certain extent limits the development of cloud computing platform, traditional data processing technology has been unable to adapt to the current era of large amount of information. Based on this, this paper studies the network data processing based on cloud computing platform. This paper designs k-means clustering algorithm to process network data, and selects 32G, 16G, 8G, 4G, 2G, 1G and other data groups to test the performance of clustering algorithm for network data processing. The experimental results show that with the increase of the size of the data set, the acceleration efficiency of the algorithm will also be improved.

Object detection of a moving object under variable illumination condition is one of the inspirational tasks for the researchers. In order to do so, background subtraction plays a vital role. In this paper, we constructed the background model by computing the geometric mean (GM) of log-normal features of frames in temporal direction. This is followed by subtracting the background model with the input frame. Finally, for object detection, we have proposed three clustering algorithm, i.e. fuzzy C-mean (FCM), fast fuzzy C-mean (FFCM) and adaptive fuzzy C-mean (AFCM) algorithm. We have implemented the proposed algorithm on the standard PET data set. Partition coefficient (PC), partition entropy (PE) and Xie Beni index (XBI) are used to validate the performance. The simulation result of our proposed algorithm has been compared with the existing work, and by using different validity index, we concluded that GM-based AFCM is better than GM-based FFCM and GM-based FCM.

With the development of network technology, the importance of intrusion detection has gradually increased. At the same time, due to the continuous increase in the number of network connections, the efficiency of traditional intrusion detection technologies is low. In order to solve this problem, this article uses GA-XGB algorithm for intrusion detection. The model uses genetic algorithm for feature selection, remove redundant and low-relevant features and XGBoost algorithm for final classification. Experiments conducted with the KDD data set prove that the accuracy, recall, F1 score and ROC score of the GA-XGB algorithm are improved compared to other traditional machine learning algorithms.

When the task model or data set changes, the active learning strategies based on heuristics are difficult to perform well all the time. In this paper, we propose a task-agnostic and model-agnostic active learning method based on meta-feature. This method draws on the idea of meta-learning. First, we designed the meta-features of an unlabeled sample at the current learning stage. These designed meta-features have nothing to do with the feature space of the data set or the form of the task model. So our method can be applied to any data set and any model. Second, we regard the active learning query selection procedure as a regression problem. We design a meta regressor that predicts the improvement of model performance for a candidate sample in a particular learning state. And we train the regressor on the experience from previous active learning outcomes. Experimental results show that our method is more stable and effective than the heuristic active learning method.

Big data is a collection of data which increases exponentially; generally, big data is complex in nature due to its dimensional characteristics. Present data managing tools do not efficiently process and store huge data. In an incomplete data set, there will be missing nodes, which will be randomly distributed in its dimensions. When the data set is large, it is very difficult to get the information. So the dominance value in the data set is considered as most significant value. An in-depth study is essential to obtain these k-dominant values from a data set. Algorithms such as skyline-based, bitmap index guided, upper-bound-based and pairwise comparisons are some of the familiar models available to identify the dominance values. Due to its slow data processing characteristics, those models are suitable to process small data sets and it requires numerous comparisons between values, complexity increases with increasing data, respectively. Considering these issues, this research work proposed a novel algorithm dominance on incomplete big data using scoring method (DIBSM) to obtain the dominance values quickly using a bit string representation of each user and the scores. Based on the score, we will get the top-k dominance values. MapReduce method is used to process the data in the proposed algorithm which helps to enable parallel data processing and obtains the results more quickly. As compared to the existing method, using this algorithm dominance values got more quickly.

In the case of information overload of academic papers, the demand for academic paper recommendation is increasing. Most of the existing paper recommendation methods only utilize scholar friendship or paper content information, and ignore the influence of temporal weight on research interest, and hence they are hard to obtain good recommendation quality. Aiming at this problem, the method HNTA for academic paper recommendation based on the combination of heterogeneous network and temporal attributes is proposed. HNTA firstly constructs a heterogeneous network composed of different types of entities to calculate the similarity between two papers, and then the temporal attribute is introduced into scholars’ research interests which are divided into instant interests and continuous interests to calculate the similarity between scholars and papers. Finally, by weighting the above two similarities, the purpose of recommending papers to scholars is achieved. Overall, HNTA can not only comprehensively utilize both relationships of scholars and the content information of papers, but also it considers the impact of the temporal weight of scholars' research interests. By conducting comparative experiments on the data set of the real academic social network: SCHOLAT, the results show that HNTA performs better than traditional paper recommendation methods.

Now a days, enormous quantity of data is being generated everyday from various sources such as online data, social media, scientific data, text data, images etc. When data becomes large in Volume, generated in high Velocity and Variety of data in the form of structured data or unstructured data is known as Big data. Traditional database tool such as RDBMs cannot deal with Big data. Challenges in big data are information storage, search and retrieval of data. Since most of the data are in different formats, we need tools and technology for transforming, analyzing, visualising data and extracting information for decision making. Collection of large data set is a significant issue and to make sense from that big data is a problem. We need ways to find semantic information from the data to help users make better decisions making. To facilitate better data management, Semantics can be used. Semantic can be efficiently use for management of large data sets and better information retrieval process. Semantic annotation of heterogeneous data sets could facilitate the search and retrieval using the concepts contained in them. This paper proposes a new methodology for designing a Learning based Method for Smart Semantic Data Management System.

The heart disease is considered as the most widespread disease. It is challenging for most of the physicians to diagnose at an early stage to avoid the risk of death rate. The main objective of this study involves the prediction of heart disease by using efficient techniques based on feature selection and classification. For feature selection, the enhanced genetic algorithm (GA) and particle swarm optimization (PSO) have been implemented. For classification, the recurrent neural network (RNN) and long short term memory (LSTM) has been implemented in this study. The data set used is the Cleveland heart disease data set available on UCI machine learning repository, and the performance of the proposed techniques has been evaluated by using various metrics like accuracy, precision, recall and f-measure. Finally, the results thus obtained have been compared with the existing models in terms of accuracy. It has been observed that LSTM when combined with PSO showed an accuracy of 93.5% whereas the best known existing model had an accuracy of 93.33%. Therefore, the proposed approach can be applied in the medical field for accurate heart disease prediction.

Gait is very important to identify person from distance. It requires very less interaction with human participants. Gait is considered the popularly known visual identification technique. The major challenges associated with gait-based person identification are high variability, gait occlusion, pose and speed variance and uniform gait cycle detection, etc. In this research work, the CASIA-A, B and C data set is explored for the view, cloth and speed invariant person identification to address the challenged associated with gait-based person identification. In this work, the very important technique of computer vision for object identification is being explored. It included feature extraction techniques, namely gait energy image(GEI) for cloth invariance, histogram of gradients(HOG) for multiview invariance and Zernike moment with random transform for crossview invariance. To classify data, SVM, ANN and XGBoost-based machine learning algorithms are used on the CASIA gait data set and achieved 99, 96 and 67% identification accuracy, respectively, for three different scenarios of invariance, i.e. speed, cloth and pose.

Runoff estimation, as well as forecasting, is a challenging hydro-climatological topic since governing physical processes is complex, and in reality, it is hardly represented by a system of the equations. Due to the complex nature and extreme spatial-temporal variability of the processes which control runoff, it is difficult to set up a reliable framework for runoff prediction and forecasting based on available observations only. In this research, two kinds of methods have been approached. The first one is a conceptual method, Soil Conservation Service Curve Number (SCS-CN) method, which combines the climatic factors and watershed parameters in one unit called the Curve Number (CN). The other method is the Artificial Neural Network (ANN) modeling, where two different kinds of models, Feed Forward Back Propagation (FFBP) and Cascade Forward Back Propagation (CFBP) model have been applied. The runoff-rainfall coefficient has been chosen as the standard parameter of the study result. Among 16 years, the year 2000 has the highest annual, seasonal monthly total runoff (monsoon season, July to Sept.). In artificial neural network models, generated coefficient correlation (R) values varied from 0.96 to 0.99 range, which indicated a good correlation between the rainfall-runoff data set. The models developed for the present study can be utilized for further basin hydrologic analysis.

Automatic melodic phrase detection and segmentation is a classical task of content-based music information retrieval and also a key problem of automatic music structure analysis. In this paper, we apply neural network architectures with conditional random field (CRF) to produce satisfactory melodic phrase segmentation. To tackle the problem of the sparse labelling problem of data, we design two tailored label-engineering techniques with corresponding training techniques for different neural networks. We compare the performance of the traditional model on the public data set Essen Folksong Database. The experimental results show that the performance of the model using the neural CRF architecture far exceeds that of the traditional method. The results of ablation experiments on Essen Folksong Database and POP909 dataset show that the improvement of performances mainly comes from the introduction of CRF structure. Besides, our labelling techniques also improve model performance and make training process more robust (Codes and data are available at https://github.com/ldzhangyx/music-melody-segmentation-using-neural-CRF).

The small round blue cell tumors are childhood tumors, they are named so because they appear like round masses of cells, blue in color under the histological observations. These tumors are of four types—Ewing’s tumors, neuroblastoma, non-Hodgkin lymphoma, and rhabdomyosarcoma. Although these tumors appear similar yet they are molecularly different and need different courses of treatment, therefore the task of differentiating among the four subtypes of this tumor is an important and challenging task in computational as well as clinical cancer research. In this paper, we have presented an approach to distinguish between the four types of small round blue cell tumors from the analysis of gene expression data set using the steps of feature selection using mutual information and classification using regularized support vector machine with Radial basis function kernel. We have presented the results and discussed their significance. The comparison with the existing method shows that the proposed method performs better with respect to classification accuracy, F-score, and other confusion matrix parameters.

Segmentation is a process of detecting boundaries of an object to extract the object of interest within a given image. There are different techniques like CT scan, MRI scan, X-ray scan, and so on those can be used to get these medical images. Processing of these medical images is laborious because of variation in size and shape, and contrast. Hence, Skin lesion segmentation became a challenging task for researchers and dermatologists. The Segmentation of medical images plays a vital role in medical diagnosis and further treatment. Although there are many proposed image segmentation techniques, there is no perfect segmentation method that supports different datasets. This paper presents an efficient skin lesion segmentation model using a Convolutional Deconvolutional Neural Networks (CDNN). The proposed framework is developed based on Convolutional Neural Networks (CNN) by replacing the classification network with a segmentation network. The proposed model has used International Skin Imaging Collaboration (ISIC) 2017 challenging data set and PH2 dataset, and results are compared with State of Art models U-Net and SegNet.

There is the need for a device to identify and forecast crimes at a complex period because of the escalation in the occurrence of crimes. The goal of this survey is to learn different deep learning techniques to help identify and forecast crimes using Deep Learning. Noticeable findings from this survey are that pre-processing becomes an essential activity as dataset instances have a significant amount of missing values, and violence does not arise consistently through metropolitan environments but concentrates in particular locations. Predicting crime hotspots is also a very important activity, and adding post-processing would help lower crime rates. The Deep Learning algorithm was commonly used in many areas, including image recognition and natural language processing. With fine tuning, the Deep Learning algorithm produces improved predictive performance. The overall detection depends on the crime data sets which could be found from the UCI library or any standard depository. Through Deep learning algorithm on the data set, it could be analyzed that how the crime patterns are in the particular demography. So, applying deep learning in crime prevention is highly advantageous and could be further reinforced by offering more resources to intelligent systems. This will certainly help to get better findings about crime through deep learnings. This paper explores mainly genetic algorithms to find the crime detection.

This paper presents a system for the intelligent search of shipwrecks using Autonomous Underwater Vehicles (AUVs). It introduces a machine learning approach to the automatic identification of potential archaeological sites from AUV-obtained side-scan sonar (SSS) data. The site identification pipeline consists of a series of stages that set up for, run, and process the output of a convolutional neural network (CNN). To alleviate the issue of training data scarcity, i.e. the lack of SSS data that includes shipwrecks, and improve the performance at testing time, a data augmentation stage is included in the pipeline. In addition, edge detection and other traditional image processing feature extraction methods are used in parallel with CNN to improve algorithmic performance. Experiments from two multi-deployment shipwreck search expeditions involving actual AUV deployments along the coast of Malta for data collection and processing demonstrate the pipeline’s usefulness. Results from these two field expeditions yielded a precision/recall of 29.34%/97.22% and 32.95%/80.39%, respectively. Despite the poor precision, the pipeline filters out 99.79% of the area in data set A and 99.31% of the area in data set B.