Generative diffusion models have shown promise in creating realistic data, but prior work has not fully addressed the unique temporal dependencies in financial time series. In this work, we introduce Multiscale Spatial–Temporal Dynamics Diffusion (MSSTD-Diff), a new approach to generate high-quality synthetic multivariate financial time series data. MSSTD-Diff uses advanced temporal encoding and multiscale data representation to capture complex patterns and relationships in financial markets over time. Unlike standard diffusion models, we use a one-step generation process and adjust the training to emphasize these temporal relationships. We evaluate the model using moments, autocorrelation, correlation, distributional properties, and an explainability index to assess how well the synthetic data mimics real market behavior. Our experiments show that MSSTD-Diff outperforms previous diffusion-based methods in reproducing the temporal dynamics of financial time series. These results suggest that our approach can produce higher-quality synthetic financial data, benefiting applications such as risk modeling, portfolio optimization, and algorithmic trading.

Model-based evaluation of synthetic financial time series data: A comparative study with multi-metric validation

Industry 4.0 technologies demand robust fault detection and diagnosis systems distinguishing genuine causal relationships from spurious correlations in complex industrial processes. Traditional correlation-based approaches exhibit significant limitations with nonlinear dynamics, temporal dependencies, and uncertain operational conditions. This paper presents a physics-guided variational attention framework for causal discovery, integrating log-normal variational attention mechanisms with probabilistic interventions and domain expertise. The dual-attention architecture utilizes multivariate log-normal distributions to model asymmetric, positive-valued causal strengths, addressing limitations of symmetric Gaussian parameterizations. Physics-informed priors from operator knowledge are incorporated through Gaussian Mixture Models and transformed via moment-matching. Uncertainty quantification employs Monte Carlo sampling and conformal filtering for statistically rigorous causal validation. Evaluation across synthetic time-series data, Australian Refinery Process oscillation diagnosis, and Tennessee Eastman Process demonstrates superior performance versus baseline approaches. Log-normal variational attention consistently outperforms Gaussian alternatives, with physics-informed priors providing improvements under high-uncertainty conditions, establishing a robust foundation for industrial causal discovery applications.

Inertial measurements play a vital role in motion tracking and navigation systems. In recent years, inertial sensing and state estimation pipelines have included deep learning algorithms to enhance their performance. The accuracy of such algorithms depends on the quality and diversity of the training dataset. Yet, the collection of large-scale inertial datasets remains a time-consuming and resource intensive process. This challenge has limited the advancement of robust inertial sensing learning models. Recently, diffusion models have emerged as a groundbreaking class of generative models that transform artificial data generation. They have demonstrated superior performance to adversarial generative networks and other state-of-the-art techniques in handling complex tasks. In this work, we propose a diffusion-driven framework for producing synthetic inertial data applicable to inertial-based quadrotor positioning. We introduce a novel three-stage pipeline that transforms inertial signals into image representations using delay embedding. It applies vision-based diffusion models conditioned on the target regression variables, and reconstructs synthetic time-series data. To show the effectiveness of our approach, we evaluated it on the EuRoC MAV dataset. Our comprehensive cross-validation evaluation across all 11 sequences of the EuRoC MAV dataset demonstrates that incorporating synthetic inertial data during training significantly improves regression performance, reducing the displacement error by 10.6% and the average position error by 31.1%. These results confirm that our diffusion-based generative model successfully captures the complex dynamics of inertial motion patterns, enabling the creation of diverse, realistic synthetic data that reduces dependence on extensive real-world data collection while providing high-quality training augmentation for neural inertial regression networks. Our method offers a scalable solution for data-constrained quadrotor applications by reducing costly real-world data collection.

Modern power grids are increasingly complex due to growing integration of renewable energy sources, power electronic devices, and bidirectional power flows challenging traditional fault detection and stability monitoring methods. In this study, we propose a hybrid convolutional neural networks -sCNN Transformer model for predictive fault detection and real time grid optimization. Using synthetic time series data representing normal operation and a variety of fault conditions (including amplitude spikes, harmonics, and voltage sag/swell), we perform four comprehensive analyses: time domain waveform and spectrogram inspection; frequency and harmonic domain analysis; deep model feature and attention map visualization; and grid level stability simulation (load, voltage, frequency). Our results show the hybrid model successfully distinguishes faults from normal conditions, separates different fault types in latent space, and yields robust classification performance under noisy and distorted signals. The study demonstrates the feasibility of combining spatial feature extraction and temporal sequence modeling for smart grid fault detection, and highlights the potential for real time monitoring and proactive grid management. Future work will target real world grid datasets, renewable integration scenarios, and extension to multi class/multi fault localization tasks.

In the current artificial intelligence (AI) era, the scale and quality of the dataset play a crucial role in training a high-quality AI model. However, often original data cannot be shared due to privacy concerns and regulations. A potential solution is to release a synthetic dataset with a similar distribution to the private dataset. Nevertheless, in some scenarios, the attributes required to train an AI model are distributed among different parties, and the parties cannot share the local data for synthetic data construction due to privacy regulations. In PETS 2024, we recently introduced the first Vertical Federated Learning-based Generative Adversarial Network (VFLGAN) for publishing vertically partitioned static data. However, VFLGAN cannot effectively handle time-series data, which contains both temporal and attribute dimensions. In this article, we proposed VFLGAN-TS, which combines the ideas of attribute discriminator and vertical federated learning to generate synthetic time-series data in the vertically partitioned scenario. The performance of VFLGAN-TS is close to that of its centralized counterpart, which represents the upper limit for VFLGAN-TS. To further protect privacy, we apply a Gaussian mechanism to make VFLGAN-TS satisfy an (ε ,δ)-differential privacy. Besides, we develop an enhanced privacy auditing scheme to evaluate the potential privacy breach through the framework of VFLGAN-TS and synthetic datasets.

Abstract Generating synthetic financial time series data that accurately reflects real-world market dynamics holds tremendous potential for various applications, including portfolio optimization, risk management, and large scale machine learning. We present an approach that uses structured noise for training generative models for financial time series. The expressive power of the signature transform has been shown to be able to capture the complex dependencies and temporal structures inherent in financial data when used to train generative models in the form of a signature kernel. We employ a moving average model to model the variance of the noise input, enhancing the model’s ability to reproduce stylized facts such as volatility clustering. Through empirical experiments on S&P 500 index data, we demonstrate that our model effectively captures key characteristics of financial time series and outperforms comparable approaches. In addition, we explore the application of the synthetic data generated to train a reinforcement learning agent for portfolio management, achieving promising results. Finally, we propose a method to add robustness to the generative model by tweaking the noise input so that the generated sequences can be adjusted to different market environments with minimal data.

Recent advances in time-series anomaly detection have leveraged artificial intelligence (AI) to improve detection performance. In industrial control systems (ICSs), however, acquiring training data is challenging due to operational constraints and the difficulty of system shutdowns. To address this, many countries are developing ICS simulators and testbeds to generate training data. This study uses a publicly available ICS testbed dataset as a benchmark for the discriminator in a Semi-Supervised Generative Adversarial Network (SGAN). The goal is to generate large volumes of synthetic time-series data through adversarial training between generator and discriminator networks, thereby mitigating data scarcity in ICS anomaly detection. Comparative experiments were conducted using this synthetic data to evaluate its impact on existing detection models. Using the HAI 22.04 dataset from the National Security Research Institute, this study performed feature engineering and preprocessing to identify correlations and remove irregularities. Various models, including One-Class SVM, VAE, CNN-GRU-Autoencoder, and CNN-LSTM-Autoencoder, were trained and tested on the dataset. A synthetic dataset was then generated via SGAN and validated using PCA and t-SNE. The results show that applying SGAN-generated data to time-series anomaly detection yielded significant performance improvements in F1 score. Additional validation using the SWaT dataset from the National University of Singapore confirmed similar gains. These findings indicate that synthetic data generated by SGANs can effectively enhance semi-supervised learning for anomaly detection, classification, and prediction in data-constrained environments such as medical, industrial, transportation, and environmental systems.

Motivated by the large number of wearables offering geolocation, human mobility mining has emerged as an novel research field within AI. The study of mobility creates increasingly predictable models in which it is easy to find patterns of behaviour. However, this data is not publicly available and access to it is restricted to large telecommunications operators. In this context, this paper aims to solve one of the main problems of human mobility databases, i.e. the scarcity of data for the generation of human mobility models. For this purpose, Generative adversarial network (GANs) have been proposed to generate synthetic time-series mobility data. Moreover, several neural network models are proposed to assess the impact of synthetic data generation on the prediction of human mobility. Our results show that the use of synthetic data improves predictions of human mobility compared to models based on available measured data. Specifically, the reinforcement learning with synthetic data benchmark, when compared to using only ground truth data, achieved a 1.22% improvement in R2, a 0.70% reduction in RMSE, a 2.97% decrease in MAE, a 27.07% reduction in MAPE, and an 18.18% improvement in CVRMSE, demonstrating its effectiveness in enhancing predictive accuracy.

Dear Readers, I am very pleased to announce today the tenth J.UCS issue of 2025. In this issue, various topical aspects of computer science are covered in 5 articles by 13 authors from 5 countries (Brazil, Croatia, Germany, India, Spain). As always, I would like to thank all the authors for their sound research and the editorial board for their highly valuable review effort and suggestions for improvement. These contributions sustain the quality of our journal. I would also like to express my sincere thanks to the KOALA Initiative and its team for their financial support, without which the J.UCS team would not be able to publish our journal. In an ongoing effort to further strengthen our journal, I would like to expand the editorial board: If you are a tenured associate professor or above with a strong publication record, you are welcome to apply to join our editorial board. We are also interested in receiving high-quality proposals for special issues on new topics and trends. Please consider yourself and encourage your colleagues to submit high-quality articles or special issue proposals for our journal. In this regular issue, I am very pleased to introduce the following 5 accepted articles: Rodrigo Costa Camargos and Ismar Frango Silveira from Brazil explore in their research the application of Explainable Artificial Intelligence (XAI) techniques to mitigate cognitive biases in predicting student dropout comparing Explainable Boosting Machine (EBM), Logistic Regression and XGBoost models. Sorav Kumar Singh, Alak Roy and Rajneesh Raushan from India focus their research on underwater wireless sensor networks and propose a Residual Energy-Aware Fuzzy-Based Clustering Algorithm (REAFCA), which presents an enhanced framework to improve network performance and addresses issues with energy usage. Juan Morales-García, Fernando Terroso-Sáenz, Andrés Bueno-Crespo, and  José M. Cecilia from Spain discuss in their research the analysis of synthetic timeseries as an enabler to improve region-based human mobility forecasting by applying Generative adversarial network (GANs) to generate synthetic time-series mobility data. Igor Tomičić,  Petra Grd, and Andrija Bernik from Croatia present in their research a comprehensive analysis of the integration of artificial intelligence into threat intelligence (TI) systems focusing on its potential to enhance cybersecurity operations by an extensive literature review including machine learning, deep learning, and natural language processing for automating threat detection, classification, and analysis. And last but not least, Daniel Spiekermann from Germany investigates in his research the real-world behaviour of network traffic within virtualized environments to identify the key factors that impact packet dynamics, including VM operations, multi-tenancy, user customization, and hardware adjustments. Enjoy Reading! Best regards, Christian Gütl, Managing Editor-in-Chief

The humongous volumes of data utilized to train the machine learning models are vulnerable to leakage by model inversion attacks and membership inference attacks. These days, massive amounts of research are being conducted to leverage differential privacy to safeguard the privacy of users. Tabular data generation from differentially private generative adversarial networks is still an untapped area. This work suggests a framework to enhance privacy protection in generating synthetic data by utilizing Wasserstein distance. The developed architecture generated synthetic data that replicated the time series relations of real-world data without compromising identifiable features of members of the input data. Results obtained from the architecture were compared with two other current GAN frameworks, DP-WGAN, and Time GAN. The privacy vs. utility tradeoff was found to be improved in the case of the architecture under discussion, as can be seen from the RMSE scores and Overall Quality Report.

This study addresses the challenge of generating high-quality synthetic financial time series data, acritical issue in financial forecasting due to limited access to complete and reliable historical datasets.The aim of this research was to compare the performance of the standard Variational Autoencoder andthe Vector Quantized Variational Autoencoder (VQ-VAE) in generating synthetic multivariate time seriesdata using the Adaro Energy Indonesia stock dataset. The VQ-VAE incorporates a discrete latentspace to improve the structure and control of the data generation process, whereas the standard VAEutilizes a continuous latent space. This research method was based on the implementation of bothmodels, followed by a quantitative evaluation using statistical metrics, including mean absolute error(MAE), mean squared error (MSE), root mean squared error (RMSE), and R² score. This researchshowed that the VQ-VAE outperformed the standard VAE in replicating the statistical characteristicsof stock prices, as shown by lower error values and higher R² scores across all tested features. The discretelatent space of the VQ-VAE led to the generation of more structured and statistically consistentsynthetic data. The implications of these findings suggest that the VQ-VAE model is highly suitablefor financial forecasting applications and indicate the potential for future enhancements throughintegration with hybrid models, such as attention mechanisms or generative adversarial networks.

Digital energy systems rely on actual data about power consumption and generation, which are not always available and, in certain situations, can be replaced with synthetic forms. This study presents a methodology for generating synthetic time-series data of electrical power consumers. It is based on the Markov chains theory, and unlike previous studies, the data are divided into hourly and hour-change monthly records, which leads to the generation of 48 transition matrices for each month. This study aimed to ensure statistical and probabilistic similarity between the original and synthetic data, which was assessed using the Frobenius distance, the coefficient of determination, variance, and standard deviation. The methodology was applied to three load profiles obtained from different types of consumers—domestic, agricultural, and industrial. In all three cases, the statistical and probabilistic characteristics of the generated data were very similar to those of the original datasets; however, the visual comparison showed that it is recommended to increase the number of states to lower the data scattering. Based on the results, recommendations are proposed on choosing the number of states for the transition matrices to optimize the statistical and probabilistic similarity. The described methodology can be used by experts involved in the design of systems with renewable energy sources and by scientists dealing with long-term studies.

This paper presents the results of a novel scoping review on the practical models for generating three different types of synthetic health records (SHRs): medical text, time series, and longitudinal data. The innovative aspects of the review, which incorporate study objectives, data modality, and research methodology of the reviewed studies, uncover the importance and the scope of the topic for the digital medicine context. In total, 52 publications met the eligibility criteria for generating medical time series (22), longitudinal data (17), and medical text (13). Privacy preservation was found to be the main research objective of the studied papers, along with class imbalance, data scarcity, and data imputation as the other objectives. The adversarial network-based, probabilistic, and large language models exhibited superiority for generating synthetic longitudinal data, time series, and medical texts, respectively. Finding a reliable performance measure to quantify SHR re-identification risk is the major research gap of the topic.

This paper introduces a novel framework for generating synthetic time-series data from turbine engine sensor readings using a text-conditioned diffusion model. The approach begins with dataset preprocessing, including correlation analysis, feature selection, and normalization. Principal Component Analysis (PCA) transforms the normalized signals into three components, mapped to the RGB channels of an image. These components, combined with engine identifiers and cycle information, form compact 19 × 19 × 3 pixel images, later scaled to 512 × 512 × 3 pixels. A variational autoencoder (VAE)-based diffusion model, fine-tuned on these images, leverages text prompts describing engine characteristics to generate high-quality synthetic samples. A reverse transformation pipeline reconstructs synthetic images back into time-series signals, preserving the original engine-specific attributes while removing padding artifacts. The quality of the synthetic data is assessed by training Remaining Useful Life (RUL) estimation models and comparing performance across original, synthetic, and combined datasets. Results demonstrate that synthetic data can be beneficial for model training, particularly in the early epochs when working with limited datasets. Compared to existing approaches, which rely on generative adversarial networks (GANs) or deterministic transformations, the proposed framework offers enhanced data fidelity and adaptability. This study highlights the potential of text-conditioned diffusion models for augmenting time-series datasets in industrial Prognostics and Health Management (PHM) applications.

In human behavior recognition using machine learning, model performance degrades when the training data and operational data follow different distributions which is a phenomenon known as domain shift. This study proposes a method for domain adaptation in the hidden semi-Markov model (HSMM) by modifying only the emission probability distributions. Assuming that the state transition probabilities remain unchanged, the method updates the emission probabilities based on the posterior distribution of the target domain. This approach enables domain adaptation with minimal computational cost without requiring model retraining. The effectiveness of the proposed method was evaluated on synthetic time-series data from different domains and actual care work data, achieving recognition performance comparable to that of models retrained for each domain. These findings suggest that the proposed method applies to various time-series data analysis tasks requiring domain adaptation.

Advanced graph embedding for intelligent heating, ventilation, and air conditioning optimization: An ensemble learning-based recommender system

There are many interesting machine learning (ML) applications in underwater acoustics. However, some ML algorithms require large amounts of training and testing data and there is a lack of open-source data for these purposes. The Point-based Sonar Signal Model (PoSSM) is a useful tool that generates synthetic time-series data appropriate for coherent signal processing applications. One underwater acoustics application is the identification of objects in synthetic aperture sonar (SAS) imagery. This paper introduces multiple datasets that provide a collection of SAS imagery from a generic SAS sonar above multiple seafloor textures and bathymetries. A variety of objects (e.g., cylinders, rocks, lobster traps) are contained in the imagery. This collection of synthetic data is suitable for training and testing ML algorithms that span a range of complexity, from constant false alarm rate (CFAR) automated detectors to convolutional neural networks (CNNs). These algorithms can perform a variety of tasks that include object detection and classification. In this paper, we test a CFAR detector on image data to estimate object locations. An example use-case of the synthetic data is shown via the training and evaluation of CNN-based classifiers for object recognition.

Temporal spiking generative adversarial networks for heading direction decoding.

In cloud computing, ensuring the high availability and reliability of data is dominant for efficient content delivery. Content replication across multiple clouds has emerged as a solution to achieve the above. However, managing optimal replication while considering dynamic changes in data popularity and cloud resource availability remains a formidable challenge. In order to address these challenges, this article employs TFT-based Dynamic Data Replication Strategy (TD2RS), leveraging the Temporal Fusion Transformer (TFT), a deep learning temporal forecasting model. This proposed system collects historical data on content popularity and resource availability from multiple cloud sources, which are then used as input to TFT. Then TFT is used to capture temporal patterns and forecasts future data demands. An intelligent replication is performed to optimize content replication across multiple cloud environments based on these forecasts. The framework's performance was validated through extensive experiments using synthetic time-series data simulating with varied cloud resource characteristics. Some of the findings include that the proposed TFT approach improves the availability of data by 20% when compared to traditional replication techniques and also cuts down the latency level by 15%. These outcomes indicate that the TFT-based replication strategy targets to improve content delivery efficiency in the dynamic cloud computing environment, thus providing effective solution to dynamically address the availability, reliability, and performance challenges.