Zhiyuan Chen

Existing machine learning methods for financial anomaly detection (FAD) typically overlook the complex spatio-temporal interactions inherent in financial data, significantly limiting detection accuracy and interpretability. While graph-based approaches have shown promise in modeling relational structures, existing methods mainly rely on static or overly simplistic dynamic graph constructions, struggling to capture these complex spatio-temporal interactions. To address these limitations, we propose SF-DynFinGC, a novel dynamic graph construction framework specifically tailored for event-driven spatio-temporal learning in FAD. SF-DynFinGC integrates scale-free structural regularization with a decay-governed global Top-k edge selection strategy, effectively reflecting real-world hierarchical and dispersed financial network characteristics. Through dynamically evolving graph structures triggered by financial events rather than fixed intervals, SF-DynFinGC accommodates irregular temporal and spatial data distributions. Comprehensive experiments on various benchmark datasets demonstrate that SF-DynFinGC consistently achieves state-of-the-art performance across graph neural network (GNN) and traditional classifiers, significantly outperforming existing static and dynamic graph construction approaches. Structural analyses further reveal that SF-DynFinGC generates interpretable and appropriately sparse graphs, highlighting critical transactional relationships indicative of complex anomaly patterns. This study provides a robust and interpretable dynamic graph construction framework that enhances both detection accuracy and analytical transparency, thereby extending the practical applicability of graphbased methods in financial risk management. Our code and data resources are available at https://github.com/AnguoCYF/SF-DynFinGC.

Unsupervised continual learning (UCL) aims to develop learning systems that can acquire knowledge from a sequence of unlabeled and potentially non-stationary data while retaining previously learned information, without relying on task labels or explicit supervision. By combining unsupervised learning with continual learning (CL), UCL provides a framework for autonomous adaptation in dynamic environments. This review presents a comprehensive synthesis of the state of the art in UCL, covering fundamental principles, representative methods, key challenges, and evaluation practices. Despite recent progress, major obstacles remain, including catastrophic forgetting, distribution shift in streaming data, and scalability in memory and computation.We summarize these challenges and organize mainstream solutions accordingly, and we review commonly used benchmarks and metrics to clarify how UCL methods are evaluated and compared. Finally, we identify key gaps and suggest directions for future UCL research.

Accurate identification of subtle faults in industrial manufacturing remains a critical challenge, driving increased adoption of machine learning (ML) techniques. However, classical ML models often overlook complex inter-sample relationships rooted in shared physicochemical properties, thereby compromising diagnostic accuracy. Addressing this, we propose Adaptive Synergistic Similarity Graph Construction (A-SSGC), a novel algorithm that adaptively fuses multiple graph construction methods. A-SSGC employs an adaptive sparsification strategy, guided by node degrees, to capture physicochemical commonalities among samples effectively. A-SSGC significantly outperforms traditional ML models, basic graph construction techniques, and both unsupervised and semi-supervised deep graph construction approaches. It consistently outperforms these baselines across representative graph neural networks on multiple industrial manufacturing datasets. Visualization of the constructed graphs confirms the ability of A-SSGC to reveal physicochemical commonalities, thereby enhancing interpretability and supporting deeper analytical insights. By effectively capturing these commonalities, A-SSGC improves diagnostic performance. It also shows strong potential as a versatile tool for industrial data analysis, contributing to improved automation and reliability in manufacturing processes. Our code and datasets are available at https://github.com/AnguoCYF/A-SSGC.

The design of public transit networks is a complex and computationally challenging problem faced by urban planners worldwide, which traditionally focuses on incremental improvements or hypothetical replacements of existing public transit networks, neglecting data-scare and non-urbanised regions. This work introduces the Modular Adaptive Network for Transit Infrastructure Solutions (MANTIS), an artificial intelligence-based framework for transit network design, addressing the multi-objective challenge of optimising station locations, route structures, and service frequencies, with a focus on full control for urban developers, and adaptability for data-scarce regions. The proposed methodology integrates three key algorithmic components: K-means clustering for data-driven station placement, a gravity model for estimating travel demand between candidate stations, and a genetic algorithm for simultaneous optimisation of routes and frequencies across multiple lines, using a range of supported and optional geographic and passenger data. The framework is evaluated using agent-based simulation in Simulation of Urban MObility, enabling quantitative analysis of network performance. By combining clustering, demand modelling, and evolutionary optimisation, this approach provides a systematic and modular foundation for developing efficient, demand-responsive transit networks in rapidly evolving urban environments.

Cureus Journals Exchange-traded funds (ETFs) have emerged as a favored investment vehicle due to their liquidity, affordability, and diversified exposure across asset classes. However, their price dynamics, influenced by macroeconomic trends, market sentiment, and real-time news, challenge traditional forecasting methods. This study advances ETF price prediction by integrating the probabilistic forecasting with autoregressive recurrent networks (DeepAR) with normalized sentiment scores (NSS) and a novel temporal keyword-enhanced sentiment framework. Temporal keyword-enhanced sentiment combines trending keywords from Google Trends with user-defined macroeconomic terms to enrich news sentiment capture, while NSS consolidates conflicting sentiments into a daily signal. Applied to eight U.S. ETFs spanning bonds, equities, and commodities, the enhanced DeepAR-NSS model consistently outperformed the baseline DeepAR, reducing mean absolute percentage error by 22-49.7% and root mean square error by up to 40.6%. Notable improvements were observed in bond ETFs (e.g., GOVT) and volatile assets (e.g., GLD), though high-risk ETFs like JNK exhibited increased root mean square error variability. These findings highlight the efficacy of sentiment integration in financial forecasting and underscore the need for asset-specific model refinement.

Changes in macroeconomic policies and market news have considerable influence over financial markets and subsequently impact their predictability. This study investigates whether incorporating sentiment analysis can enhance the accuracy of ETF price predictions. Specifically, we aim to predict ETF price movements using sentiment scores derived from news article summaries. Utilizing FinBERT for sentiment analysis, we quantify the sentiment of these summaries and integrate these scores into our predictive models. We employ DeepAR as a probabilistic model and compare its performance with LSTM in predicting ETF prices. The results demonstrate that DeepAR generally outperforms LSTM and that integrating sentiment scores significantly improves prediction accuracy. Given the promising outcomes, we also introduce a fixed “Seed” approach to ensure greater reliability and stability in our probabilistic predictions, addressing the need for robust sampling techniques in practical applications.