With the rapid growth of online social networks (OSNs), more and more people become concerned about their privacy exposure through the posted contents (e.g., photos and videos). Particularly, preserving the sensitive relationship contained in images has gained a lot of attention. Our recent work [1] pioneered the relationship preservation in images. However, it was limited to protecting sensitive relationships exposed in individual images. A critical and more challenging issue in practice is to preserve relationship privacy through the correlation between images. To address the problem, this work further identifies sensitive relationships among users presenting not only in each individual image but also across multiple images from the same event. A novel algorithm is developed to detect sensitive relationships among a group of images. Additionally, in order to protect the detected sensitive relationships, a linear programming algorithm and a greedy algorithm are designed to select faces to block, aimed at handling the trade-off between privacy preservation and utility loss (e.g., the aesthetic value of images). We conducted experiments on synthetic datasets to evaluate the performance of the proposed algorithms in terms of privacy protection and utility loss. Result comparison demonstrates that the enhanced algorithms provide better privacy protection given a group of images and more faces need to be blocked for preserving the detected sensitive relationships.
Traffic flow prediction (TFP) has become a crucial problem for billions of people who rely on traffic predictions to optimize their travels every day. In this paper, we present a decentralized federated meta-learning (DFML) framework for traffic flow prediction. Within this framework, traffic detectors spread across a road system can form federations with nearby peers in order to predict traffic collaboratively. A node that assembles a federation assumes the role of receiver while the remainder are assigned as broadcasters. The broadcasters learn model parameters from local data and then broadcast those models to receivers, thereby enriching the available information from which a receiver may learn. The framework combines the strength of decentralized federated learning and meta-learning frameworks to leverage distributed data sources without central aggregation, while also enabling models to generalize across diverse traffic conditions and environments efficiently. Evaluation of the DFML algorithm, using data from the Caltrans Performance Measurement System (PeMS), demonstrates initial viability for TFP and establishes a foundation for continued innovation.
Current assessments for depressive disorder are often influenced by cognitive function making them more susceptible to biases. Deep learning could provide more objective diagnoses with less access barriers for individuals who are unable to complete traditional assessments. In our study, we aim to explore the relations among speech, languages, and depression to demonstrate the feasibility of multi-lingual speech depression detection, and then build deep learning models using multi-lingual speech samples to support depression diagnosis. We first used a newly collected Chinese speech depression dataset to build a convolutional neural network (CNN) to conduct depression detection, and the accuracy of the test set reached 0.85. Besides, we tested the English depression speech dataset, DAIC-WOZ, using the same CNN model, and the accuracy of the test set was 0.73. While training the model using both Chinese and English speech samples and testing on mixture speeches, the accuracy achieved 0.74. We found that the CNN model can be applied across languages with a relatively stable performance of depression detection. This provides evidence that it is possible to develop a language-independent depression detection tool to support depression diagnostic and achieve worldwide long-term mental health monitoring.
Post-COVID the restaurant industry is experiencing a surge in demand, presenting a unique challenge of efficiently managing increased customer flow while ensuring seamless interactions. Chatbots have emerged as an innovative solution to meet the demand increase. The paper addresses the enhancement of AI chatbots through the integration of Retrieval-Augmented Generation (RAG) with the Large Language Model (LLM). This paper focuses on the development of a restaurant chatbot that not only engages in natural-language conversations but also addresses context optimization and LLM optimization for restaurant context learning. The approach uses a Neo4j Knowledge graph built using the restaurant data as an external source of knowledge. The graph is traversed to match the user question with appropriate answer tokens using Term Frequency - Inverse Document Frequency (TF-IDF) embeddings. The relevant tokens along with user questions are used to provide additional context to the T5 language model to provide nuanced responses to the users. This improvement is quantitatively evidenced by a Bilingual Evaluation Understudy (BLEU) score of 0.60, indicating a high level of precision in language understanding and generation. An extensive evaluation of the chatbot includes assessing AI testability on the level of words, sentences, and information. These evaluations include simulated dialogue assessments and performance analyses, with a focus on the chatbot's ability to retrieve and integrate information. Based on the AI testability evaluation, the models consistently produce more knowledgeable, diverse, and relevant answers as compared with state-of-the-art models with an average information score in the range of 0.6-0.8.
Investor sentiment significantly influences stock market dynamics, often displaying pronounced volatility. Con-currently, objective financial news remains a crucial determinant of market shifts. This study sourced an extensive collection of textual comments from the StockTwits trading platform. We then implemented a semi-supervised GAN-Bert methodology to perform sentiment analysis on these unlabeled remarks. Following this, the Inception V3 model facilitated sentiment evaluation of financial visuals from The Wall Street Journal (WSJ). Upon validation and integration of sentiment scores from these visuals, we amalgamated sentiments derived from both textual and visual data into a Bi-LSTM framework. This integrated model underwent training for predictive analytics, underpinning trading simulations based on anticipated returns. The experimental results demonstrated a notable return of 30.41% during our evaluation period, showcasing an improvement over single sentiment feature analysis.
Traditional Noise, Vibration, and Harshness (NVH) testing is costly and complex, requiring manual limit setting by skilled NVH engineers. Additionally, production line data is mostly imbalanced, leading to bias toward the majority class in classification models. We addressed these problems by (a) proposing a machine learning (ML)-based solution for the classification of defect transmissions using multivariate time series data (vibration and encoder signals) from a production line test bench, and (b) introducing a new over-sampling method “Cyclic Exponential Averaging of Minority Data”. Our novel over-sampling method uses exponential average weights, with different time constants to generate new data points between minority data points of consecutive data pairs. Our classification model employs a Multi-layer Perceptron (MLP) for fault detection, using spectral order tracks and Short Time Fourier Transform (STFT). The spectral order tracks were extracted from the time series data as follows: First, using encoder pulses, we estimated the angle of the transmission's input shaft at each given time. Next, the vibration data were resampled into angle domain. Then, the Fast Fourier Transform (FFT) of the resampled vibration signal was computed to transform the signal to order domain. Finally, the relevant spectral order was tracked, and then the decibel (dB) average computed. To extract the STFT, first, the vibration signal of each transmission was divided into several small segments of equal length, then multiplied by a sliding window of same length, and then the FFT was computed. Next, the window was moved with a 50 percent overlap. Finally, the aforementioned steps were repeated until the end of the vibration signal is reached. To verify the effectiveness of our proposed over-sampling method with our developed model, we performed 6 experiments, comparing the performance of our proposed over-sampling method with the performance of 4 state-of-the-art over-sampling methods (SMOTE, Borderline-SMOTE, SVM-SMOTE and KMeans-SMOTE) and without over-sampling. The results demonstrate that our method performs better than other methods.
As per the statistics provided by the U.S. Depart-ment of Energy, the transmission system of the United States com-prises roughly 160,000 miles of high-voltage transmission lines. Regular maintenance and monitoring of power line damages are essential. The power transmission lines and tower market size value are set to grow at a 4% CAGR between 2021 and 2028. A defective electrical line that goes undetected could cause wildfires, fatalities, and other lethal devastation. The traditional manual inspection method may cause a delay in response to anomalies as it may not be efficient in covering vast areas and is expensive. This made it necessary to establish a system that works effectively with minimal human intervention, cost, and time. In this paper, we propose a machine learning-based approach to eliminate the need for manual inspection using UAVs. The use of UAVs helps prevent any human error, and due to their GPS functionality, the UAV enables precise tracking of power line faults. We used deep learning techniques and models such as YOLOv8 to train the system for powerline fault detection and classification, achieving an average accuracy of up to 84%. A streamlit user interface has also been developed for this system, where users can upload the captured data. Using this input, the trained machine learning model will carry out the subsequent tasks, such as detecting the power line components (transmission tower, conductor, or insulator plate), identifying the faults by further classifying the faults into their sub-classes (broken wires, missing or broken insulator plates, vegetation on power lines, etc.), and reporting the details of the detected anomalies on the user interface.
Predictive maintenance is a critical task for modern systems and in the case of electric vehicles, it mostly focuses on predicting the lifespan of their rechargeable batteries. In this study, we explore the potential of machine learning to accurately predict the lifespan of lithium-ion batteries and evaluate different implementation alternatives that can boost prediction performance. Using Microsoft's BatteryML1 framework, we experimentally assess various predictive models and parameters to establish a baseline for lifespan prediction models. Subsequently, we conduct a comprehensive study on model hyperparameters and examine techniques to fine-tune them and enhance the model's performance. To further refine the performance of our approach, we implement several ensemble methods that combine the strengths of multiple models. Our results demonstrate that this approach improves the accuracy of predicting the remaining lifespan of lithium-ion batteries. This research highlights the importance of robust early prediction techniques, which can help in better battery management and lifecycle planning.
Named Entity Recognition (NER), as a fundamental task in the field of natural language processing, directly deter-mines the efficiency and accuracy of semantic processing and inference in subsequent tasks. In previous studies, the Bi-LSTM model has demonstrated relatively mature and stable performance in entity extraction tasks. However, with the continuous evolution of technology, pre-trained models have now become the preferred choice for mainstream NLP tasks. In this study, based on five publicly annotated datasets, we conducted an in-depth comparison between the Bi-LSTM and BERT models in the task of named entity recognition. We found that in four out of the five datasets, the performance of the BERT model was superior to that of Bi-LSTM. This result fully demonstrates the excellent performance of pre-trained models in the field of named entity recognition, especially when dealing with large-scale and complex textual data, where their performance is even more outstanding. Thanks to pre-trained models, which are trained on large-scale corpora, they can learn rich language rules and knowledge, thereby more accurately capturing deep semantic information in text and improving the accuracy and efficiency of recognition.
Providing a user-friendly interface for the operators of robotic devices is challenging but can save them precious time when reprogramming a device to perform different tasks under varying conditions. This paper presents a novel pipeline that adds context awareness at every step of the task execution and employs LLMs to develop a textual interface that supports commands and directions in natural language. Contextual awareness is achieved by analysing input from a camera that supervises the task execution and leveraging 6D pose estimation and scene graph generation models, to create an enriched representation of the task's environment. The rich scene graph incorporates spatial information and refined relationships between objects, and along with the user prompt in natural language is given as input to an LLM-powered chatbot, thus enabling a more intuitive interaction between the user and the robot that requires the minimum information from the user. The first experiments in a simulated environment demonstrate the effectiveness of the proposed approach in facilitating task execution through natural language inputs, empowering users without programming expertise to interact seamlessly with robotic systems and achieve complicated tasks with minimum effort.
This study considers the problem of selecting images that enhance the perception of attractiveness from a series of 3D captures of Buddhist statues, utilizing the advanced visualization capabilities of Grad-CAM++, an enhancement over traditional Gradient-weighted Class Activation Mapping (Grad-CAM). By leveraging convolutional neural networks (CNNs), Grad-CAM++ enables the identification of features that contribute to at-tractiveness in images, as perceived by observers. To validate the effectiveness of this approach, the study incorporates a comprehensive survey involving human participants to assess their reactions to the images selected by our method. This combination of computational analysis with human feedback not only showcases the capabilities of Grad-CAM++ in selecting appealing images from complex 3D datasets but also validates its utility through empirical evidence. The findings provide signifi-cant insights into the intersection of computer vision technology and cultural heritage, proposing a novel method for appreciating and evaluating the aesthetic qualities of cultural artifacts.
The goal of sparse online learning is to induce sparsity in the weights of online learning algorithms, ensuring the prediction model only contains a limited size of active features. However, the adaptability of sparse online learning has been ignored by most existing studies. In this work, we propose a novel adaptive sparse online learning framework named Asymmetric Truncated Gradient (ATG) for handling high dimensional data streams, which is implemented by an adaptive sparsity-promoting strategy to truncate the gradients in an online fashion. Our main idea is to exploit adaptive sparse learning for the current weight vector by introducing an extra parameter in TG to asymmetrically control the sparsity level for positive and negative weights, achieving adaptive sparse solutions of the online weight updating. Compared with projection-based or l(1)-norm based sparsity-introducing mechanisms, our proposed ATG algorithm is able to achieve stable sparse solutions by considering the sparsity structure of the current weight vector. Theoretical regret bound analysis of ATG provides a solid mathematical support for its wide applications. Extensive empirical studies have demonstrated the effectiveness and efficiency of the proposed ATG method.
This paper delves into the fascinating integration of Retrieval-Augmented Generation (RAG) with Large Language Models (LLMs) for optimizing supply chain management operations. RAG combines the robust retrieval capabilities of information retrieval systems with the generative prowess of neural language models to create a powerful tool that bolsters data protection while expanding the knowledge base to capture supply chain intricacies. This innovative methodology revolves around a dual-component system that employs a retrieval module to pinpoint relevant information from a knowledge base, while a generation module crafts contextualized responses using large language models. Through iterative retrieval strategies and tailored chunk optimization techniques, RAG enables contextualized analysis, predictive insights, and data-driven decision-making that streamlines processes from demand forecasting to inventory optimization. An experimental setup mimicking enterprise data classification assesses RAG's efficacy, employing recursive retrieval, multi-hop querying, and integration of generative and retrieval processes. Results showcase RAG's potential to revolutionize supply chain logistics, enhancing operational agility, minimizing disruptions, and fortifying data security. The impacts span improved forecasting accuracy, inventory level optimization, supplier risk assessment, and comprehensive supply chain reporting. However, RAG necessitates stringent ethical considerations and robust countermeasures against exploitation. Future work centers on system scalability, advanced evaluation metrics, and interdisciplinary collaboration between machine learning, retrieval systems, and supply chain domains. Overall, this paper presents a groundbreaking approach to optimizing supply chain management operations that could significantly impact the industry's future.
Breathing patterns understanding plays a crucial role in assessing human health, notably in the early detection and prediction of asthma attacks. Identifying these patterns, including inhalation and exhalation, is essential for doctors. However, the current methods, such as invasive techniques like belts or nasal probes, often cause considerable discomfort to patients. Given the rapid advancements in artificial intelligence (AI) and computer vision, there is a pressing need to leverage these technologies for developing non-invasive diagnostic solutions. Leveraging the precision of AI, this study introduces an innovative asthma attack prediction model employing contactless thermal imaging to evaluate asthma severity. By meticulously analyzing thermal patterns emanating from the nasal area, our model promises a groundbreaking shift in asthma diagnostics. Validation on a robust benchmark dataset collected at the University of Tech-nology Malaysia showcased exceptional performance metrics, with accuracy, precision, recall, and F1-score each reaching an unprecedented 99.49%, 98.01 %, 98.4%, and 99.2% respectively. Beyond its high accuracy, this method emphasizes patient comfort and accessibility, potentially democratizing asthma management across varied healthcare settings. This research not only under-scores the feasibility of non-invasive asthma assessment but also sets a precedent for future explorations in leveraging computer vision for respiratory condition diagnostics.
Due to the rapid growth of research papers on biomedical topics, it has become increasingly important to make advancements in biomedical Natural Language Processing (NLP). Biomedical NLP enables us to extract important information from text, such as new insights into the role of different genes in disease susceptibility, or the potential for drug therapies that are effective against one disease to work effectively against another. In this paper, we present a comparative evaluation of the binary relation classification capabilities of the current state-of-the-art binary relation classifier, BioBERT, against recently released open-source large language models, Gemma-2b, Gemma-7b, and Llama2-7b, which we fine-tune with the benchmark GAD and EU-ADR datasets. In addition, we quantify the potential of discovering new relationships by utilizing knowledge graphs built out of known binary relations.
Personalized and feedback-based learning frequently produces better learning outcomes in the classroom. Digital adaptive learning systems rely upon the categorization of problems and course materials to accurately recommend personalized content to users. While adaptive learning systems have already been widely adopted in math, the problems are often manually categorized and created. This paper discusses the multi-label classification application for categorizing competitive math problems into widely recognized categories using problem text, answer choices, and user-generated solutions with DistilBERT, a Transformer-based language representation model. Additionally, by adding LaTeX control sequences to the tokenizer vocabulary, the trained model was able to recognize symbolic math equations and thus more accurately predicted problem categorization compared to using a tokenizer without additional vocabulary. The trained model with additional vocabulary mostly accurately predicted problem labels across all five general categories, thus demonstrating DistilBERT's ability for problem categorization and the significance of representing symbolic equations with vocabulary. Based on the classification results, a problem recommendation algorithm adaptive to user performance is introduced, to illustrate an application of categorization in adaptive learning systems.
Artificial Intelligence models have been employed in various fields, leading to a growing interest in the subject and in the development of the models. The direct involvement of complex AI models in decision-making processes stressed the needs to explain the rationales behind the results, globally and locally for each prediction/result via eXplainable Artificial Intelligence (XAI) techniques. This paper compared three XAI techniques (SHAP, LIME and IG) with aim of using them for temporal explainability of predictive results regarding time-series in order to understand if these methods are able provide temporal explanation of deep learning AI models. The comparison provided has been qualitative and quantitative and addressing computational performance. This work has been partially supported by the CN MOST, national center on sustainable mobility in Italy, on CAI4DSA of FAIR, and has been developed on the Snap4City platform.
Developing a stock trading bot capable of navigating high-entropy environments like stock markets presents both challenges and opportunities in the field of machine learning. While quantitative algorithms currently dominate global stock trading, creating a consistently profitable prediction model remains elusive due to market unpredictability. This paper explores the feasibility of training an agent using reinforcement learning (RL) and artificial neural networks (ANN) to make optimal decisions in stock trading. We evaluate various RL techniques and their modified variants to assess the agent's performance across different stock prices and portfolio settings. While our approaches show promise in specific historical periods, they also exhibit fragility, with inconsistent results across different time periods and uncertain performance in future data. Nevertheless, this work provides valuable insights into the challenges and potential of employing RL in stock trading algorithms.
This paper surveys the latest unsupervised anomaly detection methodologies applied to health insurance fraud, covering studies from 2017 to 2024. Our review includes a variety of machine-learning approaches, evaluating their effectiveness in handling complex, high-dimensional, and imbalanced healthcare datasets. Techniques such as Isolation Forest, Bayesian hierarchical models, and deep autoencoders demonstrate superior performance compared to traditional methods. Despite significant advancements, gaps remain with regard to transfer learning, interpretability and explainability of models, and the development of real-time, incremental learning algorithms. Future research should focus on these areas to enhance fraud detection accuracy and trust. Our work aims to provide a valuable resource for researchers and practitioners, supporting the development of more robust and adaptive fraud detection systems to protect healthcare integrity and reduce financial losses.
