Search

Due to the sheer volume and variety of data in industrial manufacturing, manually creating features can be costly. Therefore, appropriate feature processing methods become crucial. Most existing feature processing methods abstract feature engineering as a feature search problem, i.e., finding feature transformations that optimize model performance. However, for automated feature engineering, the number of searches and the number of transformation combinations are huge. Therefore, we use a factorization-based model that measures interactions in terms of vector products, automatically learns patterns of combined features, and generalizes them to unseen features. Prior to this paper, the DeepFM algorithm (which combines an FM model with a deep neural network model) and the xDeepFM algorithm (which proposes a novel Compressed Interaction Network (CIN) designed to make feature interactions explicit) were available. The LRCIN proposed in this paper focuses on improving the CIN network in the xDeepFM method, by introducing a low-rank approximation method in the CIN network to reduce the number of parameters, and adding an attention mechanism after the CIN to ensure the accuracy of the model. The experimental results show that our method can effectively reduce the time complexity of the model and improve the model accuracy to some extent.

This study presents a deep learning research aimed at improving the performance of image classification models and increasing interpretability as well. We explore ways to improve the model by combining attention mechanisms with convolutional neural networks. This study uses garbage classification data in public datasets for training in a supervised learning manner, and employs Grad-CAM (Gradient-weighted Class Activation Mapping) with channel attention mechanism SE (Squeeze-and-Excitation) to generate heat maps for better understanding the decision-making process of the model’s classification. By using the Grad-CAM function to generate heat maps, it is possible to visualize the areas which the model focuses on during classification. This provides a method to explain the model’s classification decisions, allowing us to better understand the basis of the model’s decisions on different categories of images. The model is improved by adding attention modules to different stages of the ResNet50 (Residual Network-50) network, thereby improving the accuracy and performance of the network. We notice that within the same stage, the structure and required attention of each module are consistent, so only one attention module is added in each stage to reduce the burden of network learning and speed up learning. In order to simplify the calculation process of attention, a global tensor is introduced to store the attention of each stage, thereby avoiding repeated calculations in each module. The experimental results show that compared with traditional convolutional neural networks, our proposed method achieves better performance on garbage classification tasks. By combining the attention mechanism and heat map interpretation, our model is able to improve classification accuracy. This is of great significance for image classification tasks in practical applications and helps to promote the research progress of deep learning in interpretability.

Abstract In this paper, we propose a method for predicting epileptic seizures using a pre-trained model utilizing supervised contrastive learning and a hybrid model combining residual networks (ResNet) and long short-term memory (LSTM). The proposed training approach encompasses three key phases: pre-processing, pre-training as a pretext task, and training as a downstream task. In the pre-processing phase, the data is transformed into a spectrogram image using short time Fourier transform (STFT), which extracts both time and frequency information. This step compensates for the inherent complexity and irregularity of electroencephalography (EEG) data, which often hampers effective data analysis. During the pre-training phase, augmented data is generated from the original dataset using techniques such as band-stop filtering and temporal cutout. Subsequently, a ResNet model is pre-trained alongside a supervised contrastive loss model, learning the representation of the spectrogram image. In the training phase, a hybrid model is constructed by combining ResNet, initialized with weight values from the pre-trained model, and LSTM. This hybrid model extracts image features and time information to enhance prediction accuracy. The proposed method’s effectiveness is validated using datasets from CHB-MIT and Seoul National University Hospital (SNUH). The method’s generalization ability is confirmed through Leave-one-out cross-validation. From the experimental results measuring accuracy, sensitivity, and false positive rate (FPR), CHB-MIT was 91.90%, 89.64%, 0.058 and SNUH was 83.37%, 79.89%, and 0.131. The experimental results demonstrate that the proposed method outperforms the conventional methods.

Predicting survival outcomes in critical accidents has been a focal point in machine learning research. This study addresses several limitations of existing methods, including insufficient management of data imbalance, lack of emphasis on hyperparameter tuning, and proneness to overfitting. Many existing models struggle to generalize effectively on imbalanced datasets or depend on default hyperparameter settings, resulting in biased predictions. By integrating Principal Component Analysis (PCA), hyperparameter optimization, and resampling methods, as well as combining Edited Nearest Neighbors (ENN) with the Synthetic Minority Oversampling Technique (SMOTE), the model significantly improves predictive accuracy and model generalization. An ensemble model combining seven machine learning algorithms—Logistic Regression, Support Vector Machine, KNN, Random Forest, XGBoost, LightGBM, and CatBoost—was applied to predict survival outcomes. Stochastic Weighted Averaging (SWA) was applied to mitigate overfitting and enhance generalization. The accuracy increased from 91.97% to 94.89% after SWA was applied in this specific scenario. The combination of PCA-based dimensionality reduction, hyperparameter tuning, and resampling techniques (ENN + SMOTE) ensured the model handled data imbalance and optimized predictive accuracy. The final model demonstrated excellent performance, with Area Under the Curve (AUC) and Average Precision (AP) values both reaching 0.98, indicating high accuracy and precision. These improvements were validated using the Titanic dataset in a binary classification problem of predicting passenger survival. The results emphasize that ensemble learning, enhanced by SWA, offers a powerful framework for handling imbalanced and complex datasets, providing significant advancements in predictive modeling accuracy. This study provides insights into how machine learning techniques can be effectively combined to solve classification challenges in real-world scenarios.

Abstract With online work environments and other distributed computing systems—such as cloud technologies or Internet of Things systems—becoming increasingly popular today due to the COVID-19 pandemic and general technological advances, the question of how to keep them secure has also become a pertinent concern. With this increased dependence on online systems for companies, cyberattacks have also been on the rise. To protect terminal devices, many companies have resorted to implementing a single boundary-defense model. This method has yielded positive results in securing the network from external threats, but it does not effectively protect network from internal security threats. With the vulnerabilities in the internal network security in mind, a dynamic access control method used with a zero-trust software-defined perimeter security model could be a viable solution. This study proposes a dynamic access control method using an engine with a new reward and penalty point-based system (RP Engine) and a dynamic task engine (DT Engine) for a zero-trust SDP security model.

Click-through rate (CTR) prediction is a term used to predict the probability of a user clicking on an ad or item and has become a popular research area in advertising. As the volume of Internet data increases, the labor costs of traditional feature engineering continue to rise. To reduce the dependence on feature interactions, this paper proposes a fusion model that combines explicit and implicit feature interactions, called the Two-Tower Multi-Head Attention Neural Network (TMH) approach. The model integrates multiple components such as multi-head attention, residual network, and deep neural networks into an end-to-end model that automatically obtains vector-level combinations of explicit and implicit features to predict click-through rates through higher-order explicit and implicit interactions. We evaluated the effectiveness of TMH in CTR prediction through numerous experiments using three real datasets. The results demonstrate that our proposed method not only outperforms existing prediction methods but also offers good interpretability.

Abstract With the advent of the Internet, the activities of individuals and businesses have expanded into the online realm. As a result, vulnerabilities that result in actual breaches can lead to data loss and program failure. The number of breaches is increasing every year, as is the number of vulnerabilities. To address this problem, current research focuses on the detection of vulnerabilities using static analysis techniques. To prevent the propagation of vulnerabilities, a new paradigm is needed to quickly detect vulnerabilities, analyze them, and take actions such as blocking or removing them. Recently, artificial intelligence algorithms such as deep learning have been introduced for vulnerability detection. In this paper, we propose a vulnerability detection model, V-CNN, which aims to detect CWE/CVE (Common Weakness Enumeration/Common Vulnerabilities and Exposures) using CNN (convolutional neural network). We trained CWE for deep learning and redefined vulnerabilities based on CWE. We propose an experimental algorithm to improve vulnerability detection. The accuracy of the proposed V-CNN model is 98%, which exceeds the 95% of the random forest model. Therefore, our V-CNN has excellent correctness detection performance in the field of vulnerability detection. The V-CNN vulnerability detection algorithm can be used instead of static analysis to detect various security vulnerabilities.

In this paper, a medical examination algorithm is proposed that can collect users’ symptoms and automatically issue a diagnosis. The proposed algorithm makes use of “Symptom2Vec” and the “analysis model of responses on self-diagnosis questions” (AMoRSD) for real-time interviews with users. Symptom2Vec can learn about the relationship between terms related to the symptoms and disease, and establish questioning criteria to be used in patient health checkups, as well as general appropriate follow-up questions based on patient symptomology. AMoRSD analyzes the patient’s emotional expressions and responses to self-diagnostic questions, classifying them into “Sick,” “Not Sick,” and “Neutral” categories based on patterns. Compared to traditional models, Symptom2Vec earned the highest mean symptom similarity score of 0.983. Furthermore, compared to other models that only learn from patient responses, AMoRSD demonstrates an area under curves (AUC) of 0.99%, indicating that jointly learning the relationship between emotions and patient responses improves the accuracy of user response classification. The combined algorithm of Symptom2Vec and AMoRSD enhances the efficiency and accuracy of user symptom collection and appropriate diagnosis generation. The data were collected from reliable medical sources such as WebMD Dictionary, NHS inform, Snomed Ct, and Cleveland Clinic, encompassing 526 disease names and 2078 symptoms. Additional data were obtained for AMoRSD, focusing on conversations within a hospital context, and effectively trained and evaluated the model using diverse and representative datasets. This research addresses the importance of medical history-taking and contributes to the field by providing a robust framework for real-time symptom-based diagnosis in clinical environments.

Recently, various new attempts are being made to improve the quality of media content according to the expansion of the media market. Pre-visualization is one of those attempts, and the behavior of characters (agents) in virtual space is essential for pre-visualization. In this paper, a study was conducted to automatically generate behaviors of virtual characters for more efficient visualization in pre-visualization. In particular, we propose a method to automatically produce an appropriate behavior by detecting the state of the surrounding environment with a deep reinforcement learning technique. A virtual environment is created using a game engine to configure space for reinforcement learning, and a reinforcement learning model of the training environment is configured with Python and PyTorch. The virtual environment and the model training environment are communicated with the ML-agents toolkit. In the virtual environment, the character basically moves in a straight line, and three obstacles appear at random locations in front of the character. The character senses 9 states and allows 5 actions. After that, a reward is offered according to the action to proceed with learning. For performance evaluation, reinforcement learning training was conducted using the Proximal Policy Optimization (PPO) algorithm and Soft Actor-Critic (SAC) algorithm, and performance comparisons were also conducted according to the batch size. As a result, we are able to secure a reinforcement learning model with obstacle avoidance capability. Applying the model to the character proved that the character can automatically animate according to the state of the surrounding environment without explicit programming.

Aspect-based sentiment analysis (ABSA) is an important research area in natural language processing, which aims to analyze the sentiment polarity of the aspect terms present in the input sentences. In recent years, many models have focused on local text or local text-aspect relations by designing models that act directly on the local text and then fusing features of the global text. In fact, this ignores the role of the global text. This paper first proposes a masked attention mechanism that acts on the local embedding part of the global embedding, based on the global attention mechanism. Previous models use two methods, called Context-features Dynamic Mask (CDM) and Context-features Dynamic Weighted (CDW), to assign weights to text vectors based on the distance to the aspect term, these methods avoid information redundancy. In this paper, the proposed method uses this masked attention mechanism to intercept the local embedding in the global embedding and then calculate the position in the dimension of the aspect term, reorder the weights corresponding to the position, and assign them to the global embedding according to the corresponding subscripts, in this way, the proposed model not only takes into account noise reduction but can also pay more attention to the feature information of the global text. Compared with the previous embedding using two pre-training models for local and global text, the model proposed in this paper can learn features of both global and local text with only one pre-training model, so it can also improve the training efficiency of the model. The proposed model achieves good results on a total of eight datasets, including the triple-classified and quadruple-classified datasets of laptops and restaurants in SemEval2014, the restaurant dataset in SemEval2016, and the Multi-Aspect Multi-Sentiment (MAMS) dataset.

Fisheye images are attracting attention in computer vision such as autonomous vehicles and virtual reality because of the wide field of view (WFoV). However, fisheye images have geometric distortions caused by the refractive index of the lens. Conventional fisheye rectification methods require multiple images to calculate distortion coefficients and lens intrinsic parameters. This means that if the fisheye lens is changed, the same operation will have to be repeated. On the other hand, by using deep learning, images with different distortion coefficients can be rectified. Also, with end-to-end learning, no feature engineering is required. To improve the performance of fisheye image rectification, we propose global convolutional neural networks with self-attention to rectify the fisheye images. The proposed method employs dilated convolutional neural networks (D-CNNs) to enlarge receptive fields, and self-attention to extract the most important features of input images. In this way, the proposed method can extract global features from input images. To better train and evaluate the proposed method, we generate fisheye images from the Place2 dataset with Cartesian and polar coordinates, and label them with original images (ground-truth). we also schedule the learning rate with cosine annealing and use an integrated loss function. The experimental results show that the proposed method achieves an excellent performance in both qualitative and quantitative evaluations.

Internet of Things (IoT) is a network that can communicate between devices without human intervention. In this network, IoT devices collect sensor data and provide them for various applications. However, data collected from the device may include sensitive information. If this information is leaked, several serious problems will occur. Therefore, access control is presented to allow only authorized users to access important data. This paper proposes a model for organizing and managing groups with the group key (GK) so that only the authorized users can share data for access control. In particular, rekeying overhead is reduced. Because the IoT devices are resource-constrained, the network lifetime may be reduced as the rekeying overhead increases. There is a problem when group communication between multiple users is not used. If multiple users can easily access all data, the leakage of sensitive information will increase. To solve this problem, only users who need to share data join in a group based on the hyperledger fabric. Our approach groups users with the same GK and protects sensitive data by using secure communication links within the group. In addition, a trusted agent for rekeying sends a new GK to users in the group. Therefore, the security of data is guaranteed, and the network lifetime is extended. We analyze the storage cost, delay, and processing time for rekeying, and we also compare them according to the number of users and the depth of the key tree. The results of the performance analysis show that the proposed hyperledger fabric-based lightweight group management (H-LGM) outperforms the existing method in terms of storage cost, delay, and processing time.

Image processing is one of the most rapidly evolving technologies today, and it is an approach for applying operations on an image to improve it or extract relevant information from it. This is a critical research field in the engineering and computer sciences. However, analyzing a large number of variables demands a lot of memory and processing resources, which can cause a classification algorithm to overfit the training samples and underfit the test samples. As a result, various strategies, such as extraction, can be used to reduce the number of features in a dataset by producing new features from old ones. In this paper, we first propose a deep learning-based feature extraction approach with a modular neural network, where we employ a pre-trained neural architecture search net (NASNet) as a feature extractor on a custom dataset of raw copper and cobalt images. It allows the input image to be feed-forwarded while performing feature learning and feature map and then stops at a pooling layer before the fully connected (FC) layer in the NASNet to extract and save the outputs of that layer in dumped files. Second, the extracted features are used as training data to build a deep neural network and machine learning algorithms for the image classification of copper and cobalt raw minerals. The experimental results show that the NASNet extracts the features efficiently, and the proposed modular neural network performs well with the boosting-decision tree as a classifier, which gives higher accuracy of 91% than 90% of the deep neural network; moreover, the precision is 1 higher than 0.98 for the deep neural network.

For connected and autonomous electric vehicles (CAEVs), drivers must choose a power-efficient route because the energy source of a CAEV is a battery. The driver can find the power-efficient route and reduce the risk of the battery running out if the CAEV navigation system can accurately predict the battery’s state of charge (SOC) based on the power consumption prediction for the selected route. As the driver’s driving pattern and the vehicle’s battery capacity are different for individual vehicles, it is important to predict the power consumption and battery SOC for the selected route. In this paper, the system model of CAEV to collect on-board diagnostics (OBD), OBD data cloud (ODC) to manage the OBD data, and the sequence of message exchange between them is proposed. Further, we propose an algorithm that can learn power consumption based on machine learning (ML) in the ODC. It predicts the battery SOC at the end of a trip using the current SOC when the onboard navigation system requests prediction for a specific route of the CAEV. The algorithm’s performance has been evaluated by the OBD dataset of personal vehicles collected in Michigan based on three road types: downtown, urban road, and highway. The results confirmed that the proposed MLR+DNN model could predict battery consumption with an average accuracy of 92%. Additionally, an average accuracy of 94% was obtained in selecting the most similar personal driving cycle (PDC) using hierarchical grey relational analysis (GRA), which is about 15% higher than the accuracy obtained by using the dynamic time warping (DTW) method.

Abstract Hybrid access point (HAP) is a node in wireless powered communication networks (WPCN) that can distribute energy to each wireless device and also can receive information from these devices. Recently, mobile HAPs have emerged for efficient network use, and the throughput of the network depends on their location. There are two kinds of metrics for throughput, that is, sum throughput and common throughput; each is the sum and minimum value of throughput between a HAP and each wireless device, respectively. Likewise, two types of throughput maximization problems can be considered, sum throughput maximization and common throughput maximization. In this paper, we focus on the latter to propose a deep learning-based methodology for common throughput maximization by optimally placing a mobile HAP for WPCN. Our study implies that deep learning can be applied to optimize a complex function of common throughput maximization, which is a convex function or a combination of a few convex functions. The experimental results show that our approach provides better performance than mathematical methods for smaller maps.

Artificial intelligence (AI) has made rapid progress in recent years, but as the complexity of AI models and the need to deploy them on multiple platforms gradually increases, the design of network model structures for specific platforms becomes more difficult. A neural network architecture search (NAS) serves as a solution to help experts discover new network structures that are suitable for different tasks and platforms. However, traditional NAS algorithms often consume time and many computational resources, especially when dealing with complex tasks and large-scale models, and the search process can become exceptionally time-consuming and difficult to interpret. In this paper, we propose a class activation graph-based neural structure search method (CAM-NAS) to address these problems. Compared with traditional NAS algorithms, CAM-NAS does not require full training of submodels, which greatly improves the search efficiency. Meanwhile, CAM-NAS uses the class activation graph technique, which makes the searched models have better interpretability. In our experiments, we tested CAM-NAS on an NVIDIA RTX 3090 graphics card and showed that it can evaluate a submodel in only 0.08 seconds, which is much faster than traditional NAS methods. In this study, we experimentally evaluated CAM-NAS using the CIFAR-10 and CIFAR-100 datasets as benchmarks. The experimental results show that CAM-NAS achieves very good results. This not only proves the efficiency of CAM-NAS, but also demonstrates its powerful performance in image classification tasks.

Deep learning researchers believe that as deep learning models evolve, they can perform well on many tasks. However, the complex parameters of deep learning models make it difficult for users to understand how deep learning models make predictions. In this paper, we propose the specific-input local interpretable model-agnostic explanations (LIME) model, a novel interpretable artificial intelligence (XAI) method that interprets deep learning models of tabular data. The specific-input process uses feature importance and partial dependency plots (PDPs) to select the “what” and “how”. In our experiments, we first obtain a basic interpretation of the data by simulating user behaviour. Second, we use our approach to understand “which” features deep learning models focus on and how these features affect the model’s predictions. From the experimental results, we find that this approach improves the stability of LIME interpretations, compensates for the problem of LIME only focusing on local interpretations, and achieves a balance between global and local interpretations.

The Korean e-commerce market represents a large percentage of the global retail distribution market, a market that continues to grow each year, and online payments are rapidly becoming a mainstream payment method. As e-commerce becomes more active, many companies that support electronic payments are increasing the number of franchisees. Electronic payments have become an indispensable part of people’s lives. However, the types of statistical information on the results of electronic payment transactions are not consistent across companies, and it is difficult to automatically determine the error status of a transaction if no one directly confirms the error messages generated during payment. To address these issues, we propose an optimized LSTM model. In this study, we classify the error content in statistical information based on natural language processing to determine the error status of the current failed transaction. We collected 11,865 response messages from various vendors and financial companies and labelled them with an LSTM classifier model to create a dataset. We then trained this dataset with simple RNN, LSTM, and GRU models and compared their performance. The results show that the optimized LSTM model with the attention layer added to the dropout layer and the bidirectional recursive layer achieves an accuracy of about 92% or more. When the model is applied to e-commerce services, any error in the transaction status of the system can be automatically detected by the model.

With the development of artificial intelligence technology, machine learning models are becoming more complex and accurate. However, the explainability of the models is decreasing, and much of the decision process is still unclear and difficult to explain to users. Therefore, we now often use Explainable Artificial Intelligence (XAI) techniques to make models transparent and explainable. For an image, the ability to recognize its content is one of the major contributions of XAI techniques to image recognition. Visual methods for describing classification decisions within an image are usually expressed in terms of salience to indicate the importance of each pixel. In some approaches, explainability is achieved by deforming and integrating white-box models, which limits the use of specific network architectures. Therefore, in contrast to white-box model-based approaches that use weights or other internal network states to estimate pixel saliency, we propose the Optimized Input Sampling Explanation (OISE) technique based on black-box models. OISE uses masks to generate saliency maps that reflect the importance of each pixel to the model predictions, and employs black-box models to empirically infer the importance of each pixel. We evaluate our method using deleted/inserted pixels, and extensive experiments on several basic datasets show that OISE achieves better visual performance and fairness in explaining the decision process compared to the performance of other methods. This approach makes the decision process clearly visible, makes the model transparent and explainable, and serves to explain it to users.

A proper battery management system (BMS) plays a vital role in ensuring the safety and reliability of electric vehicles (EVs) and other electronic products. Accurate State-of-Health (SOH) estimation of Lithium-ion (Li-ion) batteries is a key factor in a BMS. It is difficult to determine SOH because of the complexity of the electrochemical reactions within the battery. To improve the accuracy of SOH estimation, a dynamic spatial-temporal attention-based gated recurrent unit (DSTA-GRU) model is proposed in this paper. First, we extract six features from the battery's charging and discharging processes that can reflect the aging degree of the battery to some extent. Second, this paper proposes a model to combine spatial attention and temporal attention that can not only consider the effects of states at different time step on the results, but also consider the effects of different features in the space domain. Third, the proposed model is trained and tested on NASA battery datasets and compared with other conventional models. Experiments carried on these data sets demonstrate that our model achieves higher accuracy than other conventional models.

Many machine learning algorithms and almost all deep learning architectures are incapable of processing plain texts in their raw form. This means that their input to the algorithms must be numerical in order to solve classification or regression problems. Hence, it is necessary to encode these categorical variables into numerical values using encoding techniques. Categorical features are common and often of high cardinality. One-hot encoding in such circumstances leads to very high dimensional vector representations, raising memory and computability concerns for machine learning models. This paper proposes a deep-learned embedding technique for categorical features encoding on categorical datasets. Our technique is a distributed representation for categorical features where each category is mapped to a distinct vector, and the properties of the vector are learned while training a neural network. First, we create a data vocabulary that includes only categorical data, and then we use word tokenization to make each categorical data a single word. After that, feature learning is introduced to map all of the categorical data from the vocabulary to word vectors. Three different datasets provided by the University of California Irvine (UCI) are used for training. The experimental results show that the proposed deep-learned embedding technique for categorical data provides a higher F1 score of 89% than 71% of one-hot encoding, in the case of the Long short-term memory (LSTM) model. Moreover, the deep-learned embedding technique uses less memory and generates fewer features than one-hot encoding.

Abstract As the 4th industrial revolution using information becomes an issue, wireless communication technologies such as the Internet of Things have been spotlighted. Therefore, much research is needed to satisfy the technological demands for the future society. A LPWA (low power wide area) in the wireless communication environment enables low-power, long-distance communication to meet various application requirements that conventional wireless communications have been difficult to meet. We propose a method to consume the minimum transmission power relative to the maximum data rate with the target of LoRaWAN among LPWA networks. Reinforcement learning is adopted to find the appropriate parameter values for the minimum transmission power. With deep reinforcement learning, we address the LoRaWAN problem with the goal of optimizing the distribution of network resources such as spreading factor, transmission power, and channel. By creating a number of deep reinforcement learning agents that match the terminal nodes in the network server, the optimal transmission parameters are provided to the terminal nodes. The simulation results show that the proposed method is about 15% better than the existing ADR (adaptive data rate) MAX of LoRaWAN in terms of throughput relative to energy transmission.

In a vehicular delay tolerant network (VDTN), there is no static connection, and the network behavior is highly temporal. This makes determining the routing protocol critical for the performance of a network. However, traditional routing technology rarely considers the influence of a VDTN’s selfish nodes. Under ideal conditions, all nodes will try to store and transfer as many messages as possible. However, selfish nodes may not transfer messages to other nodes due to limited resources. Taking selfish nodes into account is necessary for improving the performance of a VDTN. In this paper, we calculate each node’s credit value by recording each node’s behavior when messages are transferring to avoid selecting selfish nodes as the relay. We present an efficient VDTN specific multi-copy routing algorithm that combines Q-learning and the credit value to determine whether a candidate node is suitable for delivering a message. Because multi-copy routing protocols require a high buffer to store messages, we also optimize the node buffer to reduce network congestion. The proposed algorithm is evaluated by a number of different performance metrics, such as delivery probability, network overhead, and message latency. The proposed algorithm achieves better results in different configurations and provides improved delivery probability, low message latency, and network overhead.

eXplainable Artificial Intelligence (XAI) is a new trend of machine learning. Machine learning models are used to predict or decide something, and they derive output based on a large volume of data set. Here, the problem is that it is hard to know why such prediction was derived, especially when using deep learning models. It makes the models unreliable in the case of reliability-critical applications. So, it is required to explain how they derived such output. It is a reliability-critical application for self-driving cars because the mistakes made by the computers inside them can lead to critical accidents. So, it is necessary to adopt XAI models in this field. In this paper, we propose an XAI method based on computing and explaining the difference of the output values of the neurons in the last hidden layer of convolutional neural networks. First, we input the original image and some modified images of it. Then we derive output values for each image and compare these values. Then, we introduce the Sensitivity Analysis technique to explain which parts of the original image are needed to distinguish the category. In detail, we divide the image into several parts and fill these parts with shades. First, we compute the influence value on the vector indicating the last hidden layer of the model for each of these parts. Then we draw shades whose darkness is in proportion to the influence values. The experimental results show that our approach for XAI in self-driving cars finds the parts needed to distinguish the category of these images accurately.

Hardware bottlenecks can throttle smart device (SD) performance when executing computation-intensive and delay-sensitive applications. Hence, task offloading can be used to transfer computation-intensive tasks to an external server or processor in Mobile Edge Computing. However, in this approach, the offloaded task can be useless when a process is significantly delayed or a deadline has expired. Due to the uncertain task processing via offloading, it is challenging for each SD to determine its offloading decision (whether to local or remote and drop). This study proposes a deep-reinforcement-learning-based offloading scheduler (DRL-OS) that considers the energy balance in selecting the method for performing a task, such as local computing, offloading, or dropping. The proposed DRL-OS is based on the double dueling deep Q-network (D3QN) and selects an appropriate action by learning the task size, deadline, queue, and residual battery charge. The average battery level, drop rate, and average latency of the DRL-OS were measured in simulations to analyze the scheduler performance. The DRL-OS exhibits a lower average battery level (up to 54%) and lower drop rate (up to 42.5%) than existing schemes. The scheduler also achieves a lower average latency of 0.01 to >0.25 s, despite subtle case-wise differences in the average latency.

Recently, game companies have been increasingly offering a variety of content in their games. The more this happens, the more players will need to consider what is best for them. Players who have played such a game may not find it difficult to play, but those who are not used to play may have a hard time finding content. Therefore, in this paper, we try to give a customized guide to players in Massively Multiplayer Online Role-Playing Games (MMORPGs). We compare the similarity of growth speeds and visited areas, and then utilize this information to recommend the most similar characters. In this work, the K-means algorithm is used for clustering based on location, the Euclidean distance is calculated to recommend similar characters with similar growth speeds. In addition, Jaccard Similarity is introduced to recommend similar characters with similar access areas. Finally, we propose a method to recommend suitable areas by applying the access speed to the recommended characters in the previous steps. Our method achieves Precision and Recall of 0.74 and 0.81, respectively, on the real-life PvE (Player VS Environment) dataset.

For most machine learning and deep learning models, the selection of hyperparameters has a significant impact on the performance of the model. Therefore, deep learning and data analysis experts have to spend a lot of time on hyperparameter tuning when building a model for accomplishing a task. Although there are many algorithms used to solve hyperparameter optimization (HPO), these methods require the results of the actual trials at each epoch to help perform the search. To reduce the number of trials, model-based reinforcement learning adopts multilayer perceptron (MLP) to capture the relationship between hyperparameter settings and model performance. However, MLP needs to be carefully designed because there is a risk of overfitting. Thus, we propose a random forest-enhanced proximal policy optimization (RFEPPO) reinforcement learning algorithm to solve the HPO problem. In addition, reinforcement learning as a solution to HPO will encounter the sparse reward problem, eventually leading to slow convergence. To address this problem, we employ the intrinsic reward, which introduces the prediction error as the reward signal. Experiments carried on nine tabular datasets and two image classification datasets demonstrate the effectiveness of our model.

Deep neural network models perform well in a variety of domains, such as computer vision, recommender systems, natural language processing, and defect detection. In contrast, in areas such as healthcare, finance, and defense, deep neural network models, due to their lack of explainability, are not trusted by users. In this paper, we focus on attention-map-guided visual explanations for deep neural networks. We employ an attention mechanism to find the most important region of an input image. The Grad-CAM method is used to extract the feature map for deep neural networks, and then the attention mechanism is used to extract the high-level attention maps. The attention map, which highlights the important region in the image for the target class, can be seen as a visual explanation of a deep neural network. We evaluate our method using two common metrics: average drop and percentage increase. For a more effective experiment, we also propose a new metric to evaluate our method. The experiments were carried out to show that the proposed method works better than the state-of-the-art explainable artificial intelligence method. Our approach can provide a lower average drop and higher percent increase when compared to other methods and find a more explanatory region, especially in the first twenty percent region of the input image.

The autonomous flight of an unmanned aerial vehicle refers to creating a new flight route after self-recognition and judgment when an unexpected situation occurs during the flight. The unmanned aerial vehicle can fly at a high speed of more than 60 km/h, so obstacle recognition and avoidance must be implemented in real-time. In this paper, we propose to recognize objects quickly and accurately by effectively using the H/W resources of small computers mounted on industrial unmanned air vehicles. Since the number of pixels in the image decreases after the resizing process, filtering and object resizing were performed according to the altitude, so that quick detection and avoidance could be performed. To this end, objects up to 60 m in height were classified by subdividing them at 20 m intervals, and objects unnecessary for object detection were filtered with deep learning methods. In the 40 m to 60 m sections, the average speed of recognition was increased by 38%, without compromising the accuracy of object detection.

In the field of advertising technology, it is a key task to forecast posterior click distribution since 66% of advertising transactions depend on cost per click model. However, due to the General Data Protection Regulation, machine learning techniques to forecast posterior click distribution based on the sequences of an identified user's actions are restricted in European countries. To overcome this barrier, we introduce a contextual behavior concept for the advertising network environment and propose a new hybrid model, which we call the Long Short Term Memory-Hawkes model by combining a stochastic-based generative model and a machine learning-based predictive model. Also, to meet the computational efficiency for the heavy demand in mobile advertisement market, we define gradient exponential kernel with just three hyper parameters to minimize residuals. We have carefully tested our proposed model with production data and found that the LSTM-Hawkes model reduces the Mean Squared Error by at least 27.1% and up to 83.8% on average in comparison to the existing Hawkes Process based algorithm, Hawkes Intensity Process, as well as 39.77% on average in comparison to Multivariate Linear Regression. We have also found that our proposed model improves the forecast accuracy by about 21.2% on average.

This paper proposes an unmanned aerial vehicle (UAV)-enabled wireless-powered communication network (WPCN), which consists of a hybrid access point (H-AP), a UAV, and nodes. The H-AP broadcasts energy to all nodes, and the nodes harvest and use the energy for information transmission. However, far-apart nodes from the H-AP hardly harvest the energy and they require more energy for the same throughput as near-apart nodes due to distance-dependent signal attenuation, which is called the doubly near–far problem. To overcome the doubly near–far problem, we propose a weighted harvest-then-transmit protocol. In the proposed protocol, we consider that the channel power gain changes according to the location of nodes, whereas it has remained constant in most conventional WPCNs. The UAV acts as a mobile H-AP, where the UAV performs weighted energy transfer and receives information to/from all encountering far-apart nodes with the better channel power gain. For the UAV, we consider the flight path optimization by implementing a regression algorithm in terms of energy efficiency. Under these considerations, we aim to maximize the sum-throughput of all nodes based on the weighted harvest-then-transmit protocol, by using convex optimization techniques. The optimal time allocation is investigated for far-apart nodes and near-apart nodes sequentially, to solve the doubly near–far problem. Simulation results show that the proposed UAV-enabled WPCN outperforms the conventional WPCN with a fixed H-AP in terms of the sum-throughput maximization.

Abstract This paper proposes a large-scale wireless powered communication network (WPCN), which consists of a hybrid access point (H-AP) and numerous nodes. The H-AP broadcasts energy to all nodes, and the nodes harvest the energy and then use the harvested energy to transmit information. For the large-scale WPCN, we also propose a novel hybrid multiple access protocol, termed hybrid multiple access-based dual harvest-then-transmit. The proposed protocol utilizes both time division multiple access (TDMA) and carrier-sense multiple access (CSMA), in order to deal with various traffic patterns and transmission reliability of different applications in large-scale networks. We consider a dual wireless energy transfer (WET) at the H-AP. The main WET is performed in TDMA and the other WET is performed at space holes in CSMA, to increase the channel utilization and harvested energy. For the considerations, we study the sum-throughput maximization in the large-scale WPCN based on hybrid multiple access-based dual harvest-then-transmit. Simulation results show that the proposed protocol outperforms the conventional protocol in large-scale WPCN.

An accurate prediction of the State of Charge (SOC) of an Electric Vehicle (EV) battery is important when determining the driving range of an EV. However, the majority of the studies in this field have either been focused on the standard driving cycle (SDC) or the internal parameters of the battery itself to predict the SOC results. Due to the significant difference between the real driving cycle (RDC) and SDC, a proper method of predicting the SOC results with RDCs is required. In this paper, RDCs and deep learning methods are used to accurately estimate the SOC of an EV battery. RDC data for an actual driving route have been directly collected by an On-Board Diagnostics (OBD)-II dongle connected to the author’s vehicle. The Global Positioning System (GPS) data of the traffic lights en route are used to segment each instance of the driving cycles where the Dynamic Time Warping (DTW) algorithm is adopted, to obtain the most similar patterns among the driving cycles. Finally, the acceleration values are predicted from deep learning models, and the SOC trajectory for the next trip will be obtained by a Functional Mock-Up Interface (FMI)-based EV simulation environment where the predicted accelerations are fed into the simulation model by each time step. As a result of the experiments, it was confirmed that the Temporal Attention Long–Short-Term Memory (TA-LSTM) model predicts the SOC more accurately than others.

The number of people diagnosed with epilepsy as a common brain disease accounts for about 1% of the world’s total population. Seizure prediction is an important study that can improve the lives of patients with epilepsy, and, in recent years, it has attracted more and more attention. In this paper, we propose a novel hybrid deep learning model that combines a Dense Convolutional Network (DenseNet) and Long Short-Term Memory (LSTM) for epileptic seizure prediction using EEG data. The proposed method first converts the EEG data into the time-frequency domain through Discrete Wavelet Transform (DWT) for use in the input of the model. Then, we train the previously transformed image through a hybrid model combining Densenet and LSTM. To evaluate the performance of the proposed method, experiments are conducted for each preictal length of 5, 10, and 15 min using the CHB-MIT scalp EEG dataset. As a result, we obtained a prediction accuracy of 93.28%, a sensitivity of 92.92%, a specificity of 93.65%, a false positive rate of 0.063 per hour, and an F1-score of 0.923 when the preictal length was 5 min. Finally, as the proposed method is compared to previous studies, it is confirmed that the seizure prediction performance was improved significantly.

A cyber physical system (CPS) is a distributed control system in which the cyber part and physical part are tightly interconnected. A representative CPS is an electric vehicle (EV) composed of a complex system and information and communication technology (ICT), preliminary verified through simulations for performance prediction and a quantitative analysis is essential because an EV comprises a complex CPS. This paper proposes an FMI-based distributed CPS simulation framework (F-DCS) adopting a redundancy reduction algorithm (RRA) for the validation of EV simulation. Furthermore, the proposed algorithm was enhanced to ensure an efficient simulation time and accuracy by predicting and reducing repetition patterns involved during the simulation progress through advances in the distributed CPS simulation. The proposed RRA improves the simulation speed and efficiency by avoiding the repeated portions of a given driving cycle while still maintaining accuracy. To evaluate the performance of the proposed F-DCS, an EV model was simulated by adopting the RRA. The results confirm that the F-DCS with RRA efficiently reduced the simulation time (over 30%) while maintaining a conventional accuracy. Furthermore, the proposed F-DCS was applied to the RRA, which provided results reflecting real-time sensor information.

Cognitive Radio Sensor Network (CRSN), incorporating cognitive radio capability in wireless sensor networks, is a new paradigm of the next-generation sensor network. Sensor nodes are usually battery powered and hence have strict energy constraints. As a result, energy efficiency is also a very critical problem in the CRSN. In this paper, we focus on energy consumption because of spectrum sensing. Furthermore, we present an adaptive spectrum sensing time interval strategy, in which SUs can adjust the next spectrum sensing time interval according to the current spectrum sensing results (namely, channel status). In order to find an optimal spectrum sensing time interval, we introduce the Markov model. Then, we establish a Markov model-based mathematical modeling for analyzing the relationship between spectrum sensing time interval and prior spectrum sensing results. Finally, numerical results demonstrate that the proposed strategy with dynamic adaptive spectrum sensing time interval exceeded listen before talk (LBT) strategy which is widely used for traditional wireless sensor networks.

Distributed beamforming can significantly improve the reliability of the link and the capacity and the coverage of wireless networks. Using a subset number of nodes from a network of sensors, they collectively transmit a common message to an intended destination. In distributed beamforming, the maximum channel capacity could be changed according to the number of cooperating source nodes and the distance (between the average source nodes and destination). Therefore, the scheme is necessary to guarantee the required channel capacity. However, it is difficult to adapt the practical environment due to signal fading, interference, and low quality of sensor nodes in WSNs. Therefore, we studied about the channel characteristic and required transmission power according to the number of cooperating nodes and the distance theoretically to overcome these problems. As a result, we propose an Intelligent Transmission Power Allocation (ITPA) algorithm to guarantee the required channel capacity considering dynamic channel statement, the number of cooperating source nodes, and the distance between the average source nodes and destination with simplicity computation. In addition, ITPA distinguishes noise data (using an exponential weighted received power average) from the estimated original data. From that the system can satisfy requirements of the user without wasting power by itself.

In the dual wireless radio localization (DWRL) technique each sensor node is equipped with two ultra-wide band (UWB) radios; the distance between the two radios is a few tens of centimeters. For localization, the DWRL technique must use at least two pre-localized nodes to fully localize an unlocalized node. Moreover, in the DWRL technique it is also not possible for two sensor nodes to properly communicate location information unless each of the four UWB radios of two communicating sensor nodes cannot approach the remaining three radios. In this paper, we propose an improved DWRL (I-DWRL) algorithm along with mounting a magnetometer sensor on one of the UWB radios of all sensor nodes. This addition of a magnetometer helps to improve DWRL algorithm such that only one localized sensor node is required for the localization of an unlocalized sensor node, and localization can also be achieved even when some of the four radios of two nodes are unable to communicate with the remaining three radios. The results show that with the use of a magnetometer a greater number of nodes can be localized with a smaller transmission range, less energy and a shorter period of time. In comparison with the conventional DWRL algorithm, our I-DWRL not only maintains the localization error but also requires around half of semi-localizations, 60% of the time, 70% of the energy and a shorter communication range to fully localize an entire network. Moreover, I-DWRL can even localize more nodes while transmission range is not sufficient for DWRL algorithm.