Knowledge discovery researchers have recently made great progress on techniques to learn the governing equations of dynamic systems from system data. However, many of these techniques rely on significant prior knowledge about the system in question. One strong assumption is that researchers know a basis for the vector space the governing equations are in. In this paper we propose a technique to learn a set of functions that form a complete basis when only some of the basis functions are known beforehand by the researcher. We then empirically demonstrate our technique on real dynamical systems and show it is effective in learning a basis for the vector space that the governing equations are contained in.
Generative neural networks, particularly Generative Adversarial Networks (GANs), have gained a significant attention for their ability to generate new data. Moreover, transfer learning, known for its efficient processing and promising research outcomes, has emerged as a popular approach for feature extraction. This paper explores the application of the Triple-GAN model, which combines the advantages of GANs with transfer learning, particularly in the context of image classification. The study focuses on utilizing the Triple-GAN model combined with a pre-trained ResNet101 model for image classification. The evaluation of the proposed approach is conducted on the CIFAR-100 dataset, employing specific metrics to assess the model’s efficacy. The results highlight the promising nature of the proposed model, showcasing its potential for accurate image classification. Our model exhibits a remarkable level of accuracy reaching above 71.52 % outperforming other conventional ANN models. The significance of this research lies in its potential to advance image classification techniques by employing the Triple-GAN model with transfer learning as a viable alternative to traditional classification methods.
Driving behavior classification plays an important role in many real-world applications, including traffic accident prevention, driver safety, usage-based insurance, and optimizing ridesharing services. In this research, we introduce a novel approach called Hybrid ConvLSTM with Attention for Driver Behavior Classification (HCLA-DBC), designed to achieve precise and reliable driver behavior classification. Our main focus is to distinguish three fundamental driving behavior classes: normal, drowsy, and aggressive, which play a vital role in supporting advanced driver-assistance systems (ADAS). To evaluate the performance of our proposed models, we conducted experiments using the publicly available UAH-DriveSet dataset. The evaluation comprised six distinct models, encompassing two machine learning models (SVM and Gradient Boosting) and four deep learning models (CNN, MLP, LSTM, and HCLADBC). Our HCLA-DBC model outperformed all other models and previous results on the same data. On the test data, it achieved an accuracy of 94.12%, precision of 94.24%, recall score of 94.12%, F1-score of 94.12%, and a macro-average ROC AUC of 98.83. These exceptional outcomes underscore the efficacy of our hybridized approach and its potential to significantly enhance driver behavior classification. The HCLA-DBC model not only promises new possibilities for driver behavior analysis but also paves the way for elevated ADAS capabilities and safer driving experiences.
Path planning plays a pivotal role in the successful deployment of multi-agent systems. However, the adaptability and scalability of traditional path planning algorithms falter in dynamic, complex environments. This paper introduces the Fuzzy A* algorithm – an innovative path planning approach that combines the deterministic A* algorithm with the adaptability of fuzzy logic. The proposed algorithm modifies the cost and heuristic functions dynamically in response to local environmental variables, offering a more robust solution to the challenges faced in large-scale, dynamic environments. This paper explains the principles underpinning the Fuzzy A* algorithm and explores its potential in complex scenarios. Through benchmarking against traditional path planning algorithms, we underscore the superior adaptability, efficiency, and scalability of the proposed algorithm. The study presents this innovative algorithm as a promising future solution in path planning for scalable multi-agent systems.
In neuroimaging techniques, deep learning technologies are used to analyze brain functionalities and extract beneficial features. Specifically, magnetic resonance images (MRI) and computed tomography (CT) scans are utilized to classify neurological diseases with various static and deep learning approaches. This paper describes the architecture of the deep learning system for detecting neurological disorders from MRI images. The system has been implemented with multiple layers of deep convolutional neural networks. An optimization method, grey wolf optimization, is used for tuning the hyperparameters. Other existing models for medical image classification are compared with the system we designed, and our system outperforms all for this particular dataset. The system can successfully detect the six most common neurological disorders, including Cerebral Aneurysm, Alzheimer's disease, Parkinson's disease, brain stroke, and schizophrenia.
As wireless technologies increasingly occupy the digital landscape, wireless charging services for smartphones and related accessories have taken a leap forward. In this context, the ability to authenticate and distinguish devices has become a paramount necessity for maintaining consistent service quality and user satisfaction. In this study, we implement an efficient and cost-effective side-channel based device authentication system. Our approach primarily focused on harvesting side-channel data from Qi wireless charger voltage measurements, thus developing a procedure that is not confined to a specific protocol. To implement this authentication framework, we developed a model that classifies devices by continuously monitoring voltage fluctuations in a wireless charging environment, and learns from these observations. We conducted validation experiments using four products from two different smartphone manufacturers. Our results show that the manufacturer of a mobile device can be identified with an exceptional 100% accuracy by relying solely on the voltage readings obtained from the Qi wireless charger. Moreover, our method can distinguish among four devices with different models with an accuracy of 92%. By adopting such a side-channel based device authentication system, service providers can mitigate the risks associated with unauthorized access by unauthenticated users. Our groundbreaking findings may also contribute to the development of a robust and scalable architecture for device classification and authentication in wireless charging environments.
Mathematical equations (MEs) are the main substances in a scientific document that constitute many types of technical discussions. Each ME implicitly carries different weights throughout the whole content. In this paper, we present an algorithm to rank the MEs based on their importance within a scientific document. The score of an ME is based upon the following four major features: (1) the number of times the ME occurred throughout the content; (2) the density of mathematical objects in the sentence where the ME is embedded; (3) the number of mathematical identifiers the ME depends on; (4) the length of the ME based on the number of mathematical identifiers. The above features require the content of mathematical objects and the dependencies among them. We tackled the issue of consecutive multiple identifiers (CMI) to ensure accurate analysis of the dependencies among mathematical objects. An aggregator consisting of PCA and z-score was also proposed to transform the four features into one scalar score. As a result, the experimental results have shown an approximation of a 10% improvement in normalized discounted cumulative gain over the baseline method, suggesting that the proposed algorithm can be used to distinguish important versus unimportant MEs.
One of the main areas where generative AI models thrive is image synthesis or generation. This work highlights the importance of quality prompts in generating compelling artworks and delves into four principal methodologies for generating prompt recommendations: text embeddings, ensemble models, text with image embeddings and object detection for feature extraction. Multiple traditional and neural network-based models are explored for feature vector representation. Furthermore, the study explores the incorporation of image embeddings, the user’s preferred art styles for tailored recommendations, and the inherent challenges in evaluating these systems. We also propose a novel methodology for evaluating such systems, in the absence of ratings or preference scores, using graph analysis and community detection algorithms. This work distinctly contributes to the prompt recommendation domain and complements previous works in the AI art generation landscape.
Traditional approaches to container orchestration, such as distributed threshold-based policies, often result in isolated or underutilized resources. This can negatively affect application availability and performance. This paper presents a new orchestration framework that solves the cloud-computing resource-allocation problem, focusing on containerized environments. This framework focuses on the runtime phase of the container lifecycle. It uses visualized resource information to optimize resource reallocation in cloud services, and introduces a systematic orchestration approach that enables efficient resource utilization and real-time allocation. Its goal is to improve the smooth operation and quality of service (QoS) of containers and microservices in a cloud environment. Our method improved the experimental allocation from 24 container blocks to 67. The results improved linearly until 30k epochs and then plateaued.
Graph data has emerged in numerous scientific domains and machine learning techniques have been widely used for analysis and learning of diverse data for prediction and decision. Machine learning techniques can readily address complex problems by leveraging their structural information. But graphs cannot be directly used for existing machine learning algorithms unless encoded as vectors. The problem of efficient representation of graphs is a substantial challenge in graph machine learning. In this paper, we propose a novel two-stage framework for the representation of chemical molecule graphs based on the strengths of Graph Isomorphism Networks (GINs) and Siamese autoencoders. In the first stage, the GIN model is constructed and trained using the structural information of chemical molecule graphs. Node attributes, edge attributes, and edge indices are used as input data, while graph attributes are used as labels. The GIN model effectively captures the structural characteristics of graphs and can accurately predict graph attributes, i.e., molecular properties. It also generates Graph Embeddings, represented as vectors that encode the structural information of graphs. In the second stage, Graph Embedding vectors are further optimized for downstream similarity tasks while preserving the graph structural information. The Siamese autoencoder is constructed and trained, which reduces the dimensionality of the Graph Embedding vectors, while maximizing the preservation of structural information in the original high-dimensional vectors. The resulting low-dimensional Graph Embeddings can be effectively utilized for tasks such as approximate nearest neighbor search. The experimental results demonstrate the effectiveness of our proposed framework in accurately predicting graph similarity.
Content-Based Image Retrieval (CBIR) systems are currently a widely used solution for image retrieval tasks with various applications. Despite the advances achieved, one of the central issues is the need for methods capable of handling the scarcity or absence of labeled data. In this scenario, Query Performance Prediction (QPP) approaches represent a successful technique in the effectiveness estimation of retrieval results. In this work, we propose a novel self-supervised framework, named Regression for Query Performance Prediction Framework - RQPPF, which is flexible and can be used with different regression models. Among the contributions, our training relies only on synthetic data and rank-based features. An experimental evaluation was conducted on 4 different retrieval datasets, considering 14 visual features and 11 regression models. The results indicate highly effective predictions and most of them are greater than recent baselines.
In a previous work, we presented Agora, a stock recommendation system based on sentiment analysis. One of the potential areas for improvement we recognized was the accuracy of the supervised machine learning model. We had previously employed a logistical regression machine learning model to generate our predictions. This paper aims to detail the improvements in accuracy we made by training and deploying ensemble models for our application. We detail our improved methodology in training Random Forest and XGBoost Classifier models on similar datasets from our original publication. We performed a comparison of the accuracies between the two models to show how our improved models lead to better results in stock market prediction. We have also provided sufficient context along the way to help a reader understand what we are attempting to achieve with this paper. Our Random Forest Classifier model outperformed the logistic regression model by 10.4%, which marked a significant improvement. Detailing how we beat our original accuracy is the major takeaway of this paper.
This document is a model and instructions for LATEX. This and the IEEEtran.cls file define the components of your paper [title, text, heads, etc.]. *CRITICAL: Do Not Use Symbols, Special Characters, Footnotes, or Math in Paper Title or Abstract.
As brand marketing budgets increasingly focus on partnering with social content creators, selecting the appropriate talent to align with their brand becomes increasingly critical. This research paper focuses on presenting a novel technique for finding similar users, or ”social lookalikes,” in order to better identify partners. The system’s faster speed and interpretability are also significant in this process. The novel technique employs the core concepts of semi-supervised learning, i.e., clustering and embedding methodology. The advanced methodology is evaluated using a specially designed numerical validation technique. The method has the capability to deliver a set of similar social profiles and similar promotional performers with reasonably low variance.
Anomaly Detection is an important problem that has been well-studied within diverse research areas and application domains. However, within the field of Semantic Web and Knowledge Graphs, anomaly detection has been relatively overlooked. Additionally, the existing literature on anomaly detection over Knowledge Graphs lacks proper organization and poses challenges for new researchers seeking a comprehensive understanding. In light of these gaps, this paper aims to offer a well-structured and comprehensive overview of the existing research conducted on anomaly detection over Knowledge Graphs. In this overview, we review the quality metrics of KGs and discuss the possible errors which may occur in different parts of the RDF data. Additionally, we outline a generic conceptual framework for the execution pipeline of Anomaly Detection over KGs. Moreover, we study the anomaly detection techniques, along with their variants, and present key assumptions, to differentiate between normal and anomalous behavior. Finally, we outline open issues in research and challenges encountered while adopting anomaly detection techniques for KGs.
The Zero Trust security model is being applied to IT systems recently, leading to changes in the security model. Zero Trust requires continuous authentication for all actions and elements, including users, computing devices, and data. This necessitates additional costs and policy changes, which can impact user convenience. In this paper, the GEA system is proposed, which performs geographical-electrical authentication between a building’s power system and computing devices. By utilizing the building’s power system to generate power patterns and detecting them on computing devices, geographical-electrical authentication is achieved. The feasibility and efficiency of this system are validated through experiments. The GEA system provides transparent authentication to users and offers advantages of cost savings and maintenance without modifying existing infrastructure. Additionally, it analyzes power patterns to determine geographical location and provides security measures.
Query optimization aims to select a query execution plan among all query paths for a given query. The query optimization of traditional relational database management systems (RDBMSs) relies on the estimation of the cost of the alternative query plans in the query plan search space provided by a cost model. The classic cost model (CCM) may lead the optimizer to choose query plans with poor execution time due to inaccurate cardinality estimations and simplifying assumptions [7, 8, 14]. A learned cost model (LCM) based on machine learning does not rely on such estimations and learns the cost from runtime [5, 10, 11]. While learned cost models are shown to improve the average performance, they may not guarantee that optimal performance would be consistently achieved. In addition, the query plans generated using the LCM may not necessarily outperform the query plans generated with the CCM. In this paper, we propose a hybrid approach to solve this problem by striking a balance between the LCM and the CCM. The hybrid model uses the LCM when it is expected to be reliable in selecting a good plan and falls back to the CCM otherwise. The evaluation results of the hybrid model demonstrate promising performance, indicating potential for successful use in future applications.
Over the last decades, there has been growing interest in research in multiple and interdisciplinary fields of human-AI computing. In particular, approaches integrating the intersecting design with reinforcement learning (RL) have received more attention. However, the current research on RL may need to consider its enhancement from a human-inspired approach further. In the present work, we focus on enabling a meta-reinforcement learning (meta-RL) agent to achieve adaptation and generalization according to modeling Markov decision processes using Bayesian knowledge and analysis. By introducing a novel framework called human-inspired meta-RL (HMRL), we incorporate the agent performing resilient actions to leverage the dynamic dense reward based on the knowledge and prediction of a Bayesian analysis. The proposed framework can make the agent learn generalization and prevent the agent from failing catastrophically. The experimental results show that our approach helps the agent reduce computational costs with learning adaptation. Finally, we conclude and anticipate that integrating human-inspired meta-RL can enable learning more formulations relating to robustness and scalability, leading to promising directions and more complex AI goals in the future.
This paper aims to apply knowledge graph construction techniques to textbooks, explicitly focusing on the challenge of the absence of domain-specific schema for each textbook. Various entity and relation extraction models are utilized to capture logical and semantic information related to the textbook’s topic. These models include a Text-Encoding-Initiative (TEI) model to extract hierarchical concepts, spaCy Natural Language Processing (NLP), and Google Cloud Natural Language to extract semantic information from the main textual content. The study includes a case study on a cloud computing textbook, where each approach is evaluated and analyzed. Ultimately, the goal is to create knowledge graphs of textbooks, enabling the completion task of predicting missing entities or relations in a low-dimensional space.
Scientists and policy makers have become interested in ways to limit the harm caused by machine learning methods. Algorithmic recourse attempts to limit harm done by making action recommendations that change the undesirable output from a machine learning model. A major challenge with recourse recommendations is their tendency to suggest actions that change the model output without changing the target being predicted. Recourse recommendations that improve both the model prediction and the underlying target are referred to as meaningful. In this paper, we use a technique from knowledge discovery as the basis for the first machine learning algorithm designed so that recourse recommendations generated for its prediction will be meaningful. We empirically compare our algorithm to other common classification algorithms and show it has significantly better performance in terms of meaningful recourse without sacrificing accuracy.
