Artificial Intelligence Review最新文献

Mental health (MeHE) is a fundamental dimension of human well-being that encompasses emotional, psychological, and social aspects. Effective MeHE management enables individuals to cope with stress, maintain healthy relationships, and achieve their personal and social goals. However, traditional approaches are often inadequate in addressing the multidimensional challenges of early detection, personalized interventions, and comprehensive MeHE education. Large Language Models offer a transformative approach to the field of MeHE. With the ability to process large and complex textual data, these models can identify behavioral patterns in patients’ responses, suggest personalized interventions, and improve access to MeHE resources. Despite these advances, significant challenges remain. Applying reinforcement learning techniques to MeHE applications necessitates addressing challenges such as model-driven bias, protecting sensitive information, and providing robust evidence of clinical performance. This review systematically examines the applications of large language models in MeHE, providing a comprehensive analysis of their capabilities and limitations. This study examined how large language models address existing challenges, including early diagnosis, personalized treatments, and effective public education. Findings show that large language models increased the accuracy of early diagnosis of mental disorders by 33%, the effectiveness of personalized treatment plans by 27%, and participation in MeHE education and awareness by 24%. Ultimately, this research underscores the pivotal role of large language models in promoting MeHE. By providing practical insights and suggesting strategies to overcome implementation challenges, this review lays the groundwork for developing innovative, effective, and equitable solutions in the field of MeHE.

Reliable 3D detection is fundamental to autonomous systems such as mobile robots, self-driving cars, and unmanned aerial vehicles (UAVs). To achieve this capability, researchers have developed and trained supervised networks, which require large amounts of diverse and precisely annotated data. Due to the complex, expensive, and time-consuming capturing and annotation process, synthetic dataset generation approaches have gained popularity over the last decade. With increasing computational resources and advances in simulation technologies, a variety of dataset generators have emerged. These methods rely on either traditional 3D modeling or neural image synthesis to generate data for specific scenarios or general-purpose 3D detection tasks. Their primary goal is to produce high-quality, annotated 3D datasets in an automated and scalable manner. In this review, we evaluate the extent to which state-of-the-art approaches fulfill this goal by introducing a categorization scheme and conducting a comprehensive analysis of both 3D modeling and neural synthesis methods. Our analysis includes techniques used to address the Sim-to-Real domain gap. Furthermore, we assess each method’s level of automation, prerequisites, and practical adoption. This review aims to guide the reader in selecting automated dataset generation workflows for specific detection problems. By considering dataset quality, prerequisites, and application scenarios, we offer practical insights into identifying suitable methods for diverse downstream tasks.

Generative Artificial Intelligence (GenAI) is rapidly redefining the landscape of work organizations and society at large. GenAI has rapidly evolved from rule-based symbolic systems ofThe 1940 s to advanced deep learning architectures capable of producing human-like content across modalities, such as text, images, audio, and video. This review focuses on current emerging trends, such as large concept models and critical comparisons of tools, including ChatGPT, Gemini, and Claude. This study synthesizes evidence of GenAI’s essential role across major industries, revealing transformative applications in the finance, cloud and IT, healthcare, education, and energy sectors. The paper also highlights the unique opportunities GenAI offers for start-ups, enabling agile projects to leverage cutting-edge technology for competitive advantage. However, the deployment of GenAI systems through edge devices also raises critical challenges related to ethics, transparency, bias, accountability, computational issues, and many more. To address these complexities, this paper examines emerging approaches such as AI agents, agentic AI, and multi-agent systems that aim to extend the functionality of GenAI through autonomy, goal-directed behavior, and collaborative intelligence. It discovers novel incorporations with agentic AI architecture, such as BabyAGI, and discusses emerging issues of coordination, hallucination, and security risks. The findings reveal persistent challenges related to scalability, interpretability, and regulatory compliance while identifying future research directions toward developing more sophisticated, ethical, and accessible GenAI systems that will continue to reshape technological landscapes and societal interactions. This systematic review informs researchers, academicians, data scientists, and developers about the latest advancements in GenAI and highlights its applications and role across various industries, as well as supporting practitioners and scholars in staying current with the rapidly evolving landscape of generative technologies.

This paper proposes a novel genetic algorithm to optimize recommendations from multiple trading strategies derived from the Directional Changes (DC) paradigm. DC is an event-based approach that differs from the traditional physical time data, which employs fixed time intervals and uses a physical time scale. The DC method records price movements when specific events occur instead of using fixed intervals. The determination of these events relies on a threshold, which captures significant changes in price of a given asset. This work employs eight trading strategies that are developed based on directional changes. These strategies were profiled using varying values of thresholds to provide a comprehensive analysis of their effectiveness. In order to optimize and prioritize the conflicting recommendations given by the different trading strategies under different DC thresholds, we are proposing a novel genetic algorithm (GA). To analyze the GA’s trading performance, we utilize 200 stocks listed on the New York Stock Exchange. Our findings show that it can generate highly profitable trading strategies at very low risk levels. The GA is also able to statistically and significantly outperform other DC-based trading strategies, as well as 8 financial trading strategies that are based on technical indicators such as aroon, exponential moving average, and relative strength index, and also buy-and-hold. The proposed GA is also able to outperform the trading performance of 7 market indices, such as the Dow Jones Industrial Average, and the Standard & Poors (S&P) 500.

The degradation of concrete structures is significantly influenced by carbonation, where atmospheric carbon dioxide (CO₂) penetrates the concrete matrix. Measuring how far carbonation penetrates into concrete plays a vital role in maintaining structural integrity and construction safety standards. Precisely forecasting the extent of carbonation penetration in recycled aggregate concrete (RAC) remains fundamental for understanding long-term performance and durability. This research is the first to introduce an innovative approach that leverages eight machine learning algorithms to estimate carbonation penetration depth. The selected techniques include NGBoost, GBRT, AdaBoost, CatBoost, XGBoost, LightGBM, HistGBRT, and MLR. Moreover, to evaluate model accuracy, four key performance indicators were employed. Additionally, SHapley Additive exPlanations (SHAP) was incorporated for enhanced model interpretability. Furthermore, the investigation examined six distinct input parameter configurations during training and testing to thoroughly assess model performance. Among the evaluated algorithms, XGBoost delivered the highest accuracy, with an RMSE of 1.389 mm, MAE of 1.005 mm, and R of 0.984. CatBoost followed closely, with RMSE of 1.772 mm, MAE of 1.344 mm, and R of 0.976. Then, the LightGBM ranked third in effectiveness, exhibiting an RMSE of 1.797 mm, MAE of 1.296 mm, and R of 0.975. These results demonstrate the reliability and interpretability of advanced machine learning models for carbonation depth estimation in RAC. The developed models offer practical tools for engineers seeking to evaluate how carbonation penetration affects structural integrity. These findings establish a strong foundation for understanding and predicting carbonation-related deterioration in concrete infrastructure.

AutoML represents a pivotal advancement in machine learning by simplifying and speeding model development. This paper provides a comprehensive survey of AutoML, tracing its evolution from early metalearning, hyperparameter optimization, and transfer learning techniques to the latest advancements in neural architecture search, automated pipeline design, and few-shot learning. It covers historical context, classical approaches, and modern applications while also addressing emerging topics. Key research directions are highlighted, focusing on enhancing model interpretability, improving generalization and robustness, expanding automated pipeline design, and ethical implications of AutoML technologies. This paper aims to provide a holistic view of the current state of AutoML, serving as a valuable resource for researchers, practitioners, and stakeholders seeking to understand and advance the capabilities of AutoML in both theoretical and practical contexts.

Intensity-modulated radiation therapy (IMRT) is an advanced technique for cancer treatment that uses a computer-controlled linear accelerator to customise beams’ radiation intensities for patients, optimising the treatment effectiveness. The complexity of IMRT planning requires sophisticated algorithms to solve the different optimisation problems that arise in the context of IMRT treatment planning. One of those optimisation problems is the Direct Aperture Optimisation (DAO). The DAO problem aims to find a set of aperture shapes for each beam angle to enhance precision and improve clinical outcomes. However, this process is computationally intensive and thus, heuristic approaches have been proposed to balance computational efficiency and solution quality, offering nearly optimal solutions within clinically acceptable times. This systematic literature review aims to trace the development and application of heuristic algorithms for the DAO problem in the context of IMRT over the past two decades. We synthesised 41 studies published between 2002 and 2023, sourced from seven major databases (ACM, IEEE Xplore, PubMed, ScienceDirect, Springer, Scopus, and Web of Science). The review highlights key trends, innovations, and future directions in using heuristic methods for DAO, providing valuable insights for researchers and practitioners in radiotherapy optimisation.

Image segmentation aims to partition an image into non-overlapping regions that are coherent in appearance. Although the fuzzy C-means (FCM) algorithm is widely used for its simplicity and efficiency, it treats each pixel independently and is therefore sensitive to noise. We propose LIFWCM, a local information-based fuzzy weighted C-means algorithm that assigns a single-pass, data-driven weight to each pixel by aggregating neighborhood intensity variation and positional overlap, and then integrates these weights into the standard FCM objective and a spatially aware membership refinement. This design suppresses the influence of noisy and boundary pixels while preserving details with low computational overhead. Across six experiments on synthetic images and natural images from the Image Processing Toolbox and BSDS500, LIFWCM consistently improves segmentation quality under heavy noise. On the BSDS500 image with 30% salt-and-pepper noise, LIFWCM attains 98.96% segmentation accuracy, exceeding the best baseline, and surpassing classical FCM variants. LIFWCM also achieves higher MPA (0.94) and MIoU (0.82) than competing methods, while converging in a few iterations. These results demonstrate that LIFWCM is robust to high-intensity noise, preserves fine structures, and remains efficient due to one-time weight computation, making it suitable for real-world noisy images with complex structures.

Many real-world machine learning classification problems suffer from imbalanced training data, where the least frequent label has high relevance and significance for the end user, such as equipment breakdowns or various types of process anomalies. This imbalance can negatively impact the learning algorithm and lead to misclassification of minority labels, resulting in erroneous actions and potentially high unexpected costs. Most previous oversampling methods rely only on the minority samples, often ignoring their overall density and distribution in relation to the other classes. In addition, most of them lack in the oversampling method’s explainability. In contrast, this paper proposes a novel oversampling method that considers a subspace of the feature-set for the creation of synthetic minority samples using nonlinear optimization of a class-sensitive objective function. Suitable subspaces for oversampling are identified through a Bayesian reinforcement strategy based on Dirichlet smoothing, which may be useful for explainable-AI. An empirical comparison of the proposed method is performed with 10 existing techniques on 18 real-world datasets using two traditional machine learning classifiers and four evaluation metrics. Statistical analysis of cross-validated runs over the 18 datasets and four metrics (i.e. 72 experiments) reveals that the proposed approach is among the best performing methods in 6 and 2 instances when using random forest classifier and support vector machine classifier, thus placing it at the top. The study also reveals that some feature combinations are more important than others for minority oversampling, and the proposed approach offers a way to identify such features.

Traditional variable importance measures quantify overall feature contributions but often overlook individual-level heterogeneity. Several new procedures attempt to address this limitation but remain model dependent and may introduce biases. We propose individual variable priority (iVarPro), an extension of the Variable Priority (VarPro) framework, which uses rule-based, data-driven partitioning to estimate the gradient of the conditional mean function. By focusing on gradients, iVarPro captures how small perturbations in a variable influence an individual’s outcome, providing a more precise and interpretable measure of importance. To demonstrate its advantages, we conducted simulations and analyzed a real-world survival dataset. Our results show that iVarPro more accurately captures the true functional relationship by flexibly leveraging local samples.