Temporal Heterogeneous Networks (THNs) are evolving networks that characterize many real-world applications such as citation and events networks, recommender systems, and knowledge graphs. Forecasting THNs involves predicting future connections within a network that evolves over time and comprises diverse types of nodes and interactions with varying temporal dynamics. Although some Graph Neural Networks (GNNs) models have been successfully applied to forecast THNs, there is a lack of a general overview of how the message-passing computation could be extended to treat THNs. Moreover, most of the current solutions exhibit pitfalls in their training and evaluation strategies. Hence, in this work, we propose a graph deep learning framework for THN forecasting. Our framework decomposes the computation of a GNN layer into multiple components and introduces two different schemes to update embedding representations for THNs. This design allows the classification of existing solutions into special instances of our framework and highlights their potential limitations. We also extend the set of benchmarks for THNs by introducing two novel high-resolution temporal heterogeneous graph datasets derived from an emerging Web3 platform and a well-established e-commerce website. Overall, we conducted the first massive evaluation of THNs solutions over four temporal heterogeneous network datasets on two different future link prediction tasks using a fair newly introduced evaluation setting that considers the evolving nature of the data. Based on the limitations of existing solutions, we develop a new model that combines working techniques from previous models and leverages a new embedding update scheme. Experiments show the prediction power of our model compared to current solutions for link prediction in temporal graphs. Moreover, the experimental evaluation highlights the strengths and weaknesses of the different solutions and shows the effectiveness of our framework design.
As deep learning models grow in complexity and parameter, many researchers have attracted their attention to distributed training and hybrid parallel strategies. However, current methods are often based on dynamic programming, leading to a large time cost. There exist only a limited number of approaches to exploring deep learning techniques, and these approaches demand considerable time for online training. In this paper, we develop the Automatic 3D Parallelism Strategies Search (APSS), which is an efficient strategy search algorithm based on a pre-trained neural network and a heuristic policy. APSS initially employs a heuristic policy in conjunction with the training cluster environment to derive candidate strategies. Subsequently, a Deep Pipeline Strategy Network (DPSN) is proposed to provide a detailed pipeline partition for each candidate strategy. APSS utilizes an offline pre-training method called Contrastive Reinforcement Learning (CRL) to pre-train DPSN, eliminating the need for real data collection and fine-tuning during subsequent application. In search speed, APSS achieves enhancements by an order of magnitude faster than AMP, a prominent framework based on the open-source DeepSpeed. In terms of the quality of parallel strategies, APSS can still identify the optimal strategy. The source code will be made publicly available on GitHub1.
Graph Neural Networks (GNNs) have attracted significant attention for their effectiveness in fake news detection, particularly due to their capability to leverage the social context embedded within news dissemination. Most existing studies, however, operate on static heterogeneous information networks (HINs), assuming these graphs adequately capture the complex interactions among social entities. In practice, the dynamic nature of real-world social networks presents a substantial challenge, as training on suboptimal or outdated graph structures can severely limit the expressiveness of GNNs. Although various approaches have been proposed to model dynamic HINs, they often rely on computationally expensive message-passing mechanisms to update node embeddings, which hinders scalability to large social graphs. In this paper, we introduce DHGNN (Dynamic Heterogeneous Graph Neural Network), a novel model designed to address these challenges. DHGNN simplifies traditional GNN message-passing by employing a dynamic propagation scheme inspired by the personalized PageRank tracking process in HINs. Extensive experiments on three real-world benchmark datasets demonstrate the effectiveness and efficiency of DHGNN in detecting fake news.
While large language models (LLMs) have demonstrated impressive results on public benchmarks, their effectiveness in structured, real-world problems like behavioral analytics remains underexplored. This work assesses the out-of-the-box performance of LLMs for industry-specific downstream tasks, with player churn prediction as a representative task. Evaluating LLMs on public benchmarks risks data leakage and task-specific overfitting, so instead we perform experiments on a novel self-compiled dataset for churn prediction, a task not part of any standard benchmark. We compare the performance of OpenAI's GPT-4.1 with traditional machine learning models, such as XGBoost and MLPs, and analyze the impact of the LLM's extensive internal world knowledge and reasoning capabilities. With few-shot prompting, GPT-4.1 achieves a weighted F1 score of 0.787, matching the performance of XGBoost on the same set of samples. We show that the LLM can compensate for missing information with its internal world knowledge and reasoning capabilities, performing best if it can leverage both. Our results highlight the potential of LLMs for cross-game churn prediction and other structured, industry-specific tasks.
Large Language Models (LLMs) have demonstrated remarkable capabilities but their significant computational and memory demands hinder widespread deployment, especially on resource-constrained devices. Quantization, the process of reducing the numerical precision of model parameters, has emerged as a critical technique for compressing LLMs and accelerating inference. This paper provides an overview of LLM quantization, with a particular focus on the Post-Training Quantization (PTQ) methods implemented within the popular llama. cpp framework and its GGUF file format. We begin by covering quantization fundamentals, including the distinction between PTQ and Quantization-Aware Training (QAT). We then describe the specific PTQ schemes employed by llama. cpp, including legacy methods, advanced K-quants, and recent IQ-quants, along with their underlying mathematical principles. The paper also discusses the impact of these techniques on model fidelity, hardware requirements, inference speed, and traces the adoption of GGUF as a de facto standard in the open-source community. This work serves as a practical guide and comprehensive reference for researchers aiming to deploy LLMs on resource-constrained hardware. By systematically documenting and comparing the PTQ methods within llama. cpp, we provide the necessary insights to navigate the trade-offs between model fidelity, inference speed, and memory footprint. This enables informed decision-making for real-world applications, from local CPU-based inference to efficient edge deployment.
A careful analysis of the Large Language Model (LLM) results, generated through anonymized representations of the original dataset, is crucial to precisely evaluate the data-sharing procedure's limitations and facilitate valuable collaborations among Internet-based cognitive behavioral therapy (ICBT) companies and third parties. This paper presents an experimental study of fine-tuning 27 LMs for a multiclass classification task to identify depression severity using 40,191 tweets labeled by human annotators. We fine-tune 14 Bidirectional Encoder Representations from Transformers (BERT), 6 Robustly Optimized BERT Pretraining Approaches (RoBerta), 3 Generative Pretraining (GPT), and 4 Text-to-Text Transfer Transformer (T5) based LMs to classify confidential and anonymized tweets. We report that T5, through conditional generation, outperforms widely adopted BERT, RoBerta, and GPT types for classifying confidential and anonymized tweets. Anonymizing personal information safeguards user privacy and often increases LM performance. Case sensitivity can potentially improve or harm the performance of domain-specific LMs for original and anonymized text.
Cross-validation is a critical component for robust machine learning evaluation. In imbalanced learning, stratified cross-validation is commonly recommended to preserve class distribution. However, it neglects the underlying distribution of instance hardness, which can introduce distribution shifts between training and testing folds, ultimately compromising the validity of performance evaluation. This paper proposes a stratified cross-validation informed by the hardness distribution for robust imbalanced medical diagnosis. The proposed multi-level cross-validation (MLCV) retains jointly the class distribution and instance hardness, maintaining equivalent distribution of hardness levels across folds. This strategy enables the model to encounter a more realistic version of the medical data for a reliable performance evaluation. Experimental work demonstrates that the hardness distribution shift exists; the (MLCV) not only enhances classification performance in imbalanced medical data but also improves the results of balancing methods, as measured by classification performance indicators such as recall, precision, and F1-measure.
Relation extraction is an important task for understanding relationships between entities, building knowledge graphs, and facilitating knowledge discovery. Pre-trained models can be fine-tuned for relation extraction if a substantial amount of labeled data is available. However, acquiring extensive labeled data is generally challenging. Semi-supervised techniques for low-resource relation extraction, such as self-training, offer a promising solution by leveraging both limited labeled data and vast unlabeled data to mitigate this challenge. Traditional selftraining methods use a teacher-student framework, where a student is iteratively trained with pseudo-labels generated by the teacher. This may lead to noisy pseudo-labels and impact performance. To address this limitation, we introduce a new model called RE-AUM-LLM that generates high-quality pseudolabels using self-training combined with Area Under the Margin (AUM) and Large Language Models (LLMs), such as Llama 3.1. Experimental results on two benchmark datasets show that the proposed approach achieves state-of-the-art results for lowresource relation extraction by comparison with several strong baselines. We will make the code publicly available to enable reproducibility and further research in this area.
Classifier ensembles are mainly used for classification problems exhibiting complex data characteristics, e.g., a high class imbalance. They consist of a set of classifiers and a decision fusion method that combines the predictions of the classifiers. However, creating an accurate ensemble is a challenging optimization problem, since the individual classifiers and the decision fusion method must be selected from a huge search space that consists of numerous classification and fusion algorithms. Thereby, the selected classifiers have to be both accurate and diverse, so that they complement each other and make correct predictions on different data subsets. In addition, the selected decision fusion method has to accurately combine the predictions of individual classifiers into a consensus decision. Current literature focuses on AutoML approaches that reduce the complexity of the optimization problem by omitting the optimization of the classifier diversity and of the decision fusion. In this paper, we propose Auto-CEn as a novel ensemble approach based on AutoML that efficiently solves the optimization problem by selecting a set of accurate and diverse classifiers for the ensemble and by optimizing the decision fusion. In our extensive evaluation on 20 real-world datasets, we show that Auto-CEn outperforms several state-of-the-art baselines and that this improvement is mainly attributed to its novel contributions of diversity-based classifier selection and decision fusion optimization.
In the era of data-driven decision-making, selecting appropriate nonlinear modeling techniques is critical for building robust and interpretable predictive systems. While both Takagi-Sugeno (T-S) fuzzy models and Back-Propagation (BP) neural networks are well-established universal approximators in the data science domain, their fundamentally different structural characteristics lead to varied performance across application scenarios. This paper presents a systematic comparative study of these two modeling approaches through a series of simulation experiments designed to reflect key tasks in predictive analytics, including static function approximation, dynamic system modeling and forecasting, and real-time state tracking of time-varying systems. By evaluating performance across multiple dimensions modeling accuracy, robustness to noise, and adaptability to temporal dynamics, this work provides actionable insights into the practical strengths and limitations of each model type. The results show that T-S fuzzy models offer superior accuracy in clean, stable environments, while BP neural networks demonstrate strong resilience to noise and generalization ability in uncertain conditions. Additionally, T-S fuzzy models exhibit higher adaptability in real-time, dynamic contexts, making them valuable for time-sensitive predictive applications. This study contributes to the broader data science community by offering a structured framework for model selection based on application-specific demands, helping practitioners and researchers alike to optimize predictive modeling strategies for complex, nonlinear systems.
Air pollution, particularly particulate matter 2.5 (PM2.5), poses a significant public health challenge in densely populated developing regions. Moreover, deploying an extensive ground sensor network to monitor PM2.5 accurately is economically unfeasible in such regions. To address this problem, we utilize Sparse Variational Gaussian Process (SVGP) models to generate approximate data using the limited ground sensor data. Since SVGPs use computational approximators for Gaussian Process modeling, we hypothesize that their inducing points can be trained to adapt spatially, i.e., these points, when optimized, can spread over the region of interest. Hence, well-initialized inducing points allow SVGPs to model PM2.5 data by capturing spatial variations of the region. We evaluate our hypothesis using PM2.5 data from Lima, Peru, one of the most polluted cities in the Americas, and with very few PM2.5 ground sensors. Our experiments qualitatively validate our hypothesis of spatial adaptation and provide a quantitative justification of improved performance over the baseline models.
This study emphasizes recent advances in learning analytics and explainable AI to provide meaningful insights for data-driven interventions. We explore how ICT-based Smart Learning Environments (SLE) can enhance decision-making through predictive models with the aim of supporting learners at risk of underperformance. Our approach introduces an analytical framework that uses raw interaction data extracted from Learning Management Systems (LMS), focusing on both cognitive and social dimensions of the collaborative learning process. The system incorporates interactive dashboards and dynamic visualizations that inform tutors about group dynamics and individual engagement, allowing for early detection of learning difficulties. The first experimental results underline the effectiveness of tracking behavioral patterns and generating multi-dimensional learning indicators, which serve to improve instructional support and reflective learner feedback. This contribution emphasizes the role of predictive analytics and interpretability in the facilitation of adaptive and scalable learning interventions within SLE. This predictive model is enhanced by the integration of SHAP values of explainable AI techniques to describe the most influential learning behaviors that conduct the model's decisions. In this case, the tutors receive early risk alerts and understand each prediction.
Many companies provide their customers with digital services for analytical purposes that are backed by modern expert machine learning models specifically trained for individual appliances. These models are often easy to train, but the deployment and operation of numerous individual machine learning models is a resource-intensive challenge. Due to hidden features in the individual appliances, consolidating all expert models into one model is often not possible. However, certain groups of models with similar appliances usually can be combined without (significant) loss in performance. To find these groups without knowledge of the actual hidden features, this paper proposes the consolidation algorithm ModelForge, which is based on a novel embedding strategy for model clustering. The Prediction Loss strategy embeds arbitrary models into Euclidean space in a way that close models share similar properties and can, therefore, effectively be consolidated. We validate ModelForge across four diverse domains, which are energy forecasting, timeseries anomaly detection, weather postprocessing, and house price prediction, to show that it yields more accurately consolidated models than previous works and alternative embedding strategies.
The offshore energy industry faces challenges in maintaining ageing infrastructure, with over half of North Sea platforms past their 25-year design life. This creates a need for scalable inspection methods beyond traditional manual reviews. We present a collaborative Human-AI framework that assists in detecting structural defects whilst preserving expert oversight under challenging marine conditions. Our two-stage system first employs a lightweight classifier to filter video frames by risk level. High-risk frames are then analysed by a modified Pyramid Attention Network that performs precise defect localisation. Experts validate the results at both stages, ensuring the system's continuous improvement. To better identify rare but critical flaws, we design Enhanced Tversky, a composite loss function to mitigate severe class imbalance by explicitly prioritising rare yet safety-critical defects like cracks. Evaluation on 5,525 classified frames and 1,013 segmented images demonstrates F2 score of 87.59% and a mean IoU of 78.73%, with crack detection reaching a crucial 89.62% F2 score. The framework reduces expert review time by 3.88 x whilst maintaining safety standards, offering a practical approach to scaling offshore inspection capabilities through Human-AI collaboration.
In 2025, Indonesia launches the flagship free lunch program, named Makan Bergizi Gratis (MBG, or free nu-tritious meal). The initiative accordingly targets the improvement of human development quality, in achieving the vision of Golden Indonesia 2045. However, this program is receiving mixed reviews, with both pros and cons among societies. This study aimed to examine public opinions regarding the implementation of MBG using sentiment analysis. We utilise Latent Dirichlet Allocation (LDA) to identify the dominant topics and employ sentiment analysis using BERT to comprehend public opinions regarding the MBG program. This study found five key topics: 1) Economic impact; 2). Budget and policy prioritisation; 3). Food quality; 4). Program implementation, and 5). Community engagement. Sentiment analysis reveals negative sentiment around budget cuts and imbalanced food quality (top-ics 2 and 3). In contrast, economic impact, program implementation, and community engagement (topics 1, 4, and 5) are per-ceived positively. However, the tweet in the economic impact topic is notably shaped by buzzer (an Indonesian term used to refer to a social media user who works to “buzz” or influence public opinion regarding a particular issue) activities, particu-larly the hashtag #MBGDorongEkonomi, which influences pub-lic discourse. This result indicates that many positive tweets are possibly from buzzers, manipulating the original sentiment. The shift from positive to negative majority sentiment indicates how much the Indonesian sentiment toward the MBG program was artificially amplified. The sentiment distribution of filtered data is more reflective of the actual public perception of the MBG program.
Graph Neural Networks (GNNs) have emerged as powerful tools for modeling financial data as networks, effectively capturing both individual attributes and complex relationships. However, their inherent message-passing and aggregation operations introduce significant inference latency, limiting their applicability in latency-sensitive domains such as finance, healthcare, and robotics. Recent efforts have attempted to mitigate this limitation by distilling GNN knowledge into more efficient Multi-Layer Perceptrons (MLPs). While promising in reducing inference costs, existing GNN-to-MLP distillation approaches face three critical challenges: (1) reliance on labeled data, (2) limited robustness to noisy or perturbed inputs due to the absence of structural information, and (3) the existence of representational bias. To address these issues, we propose REFINE, a novel self-supervised GNN-to-MLP knowledge distillation framework. Our method enhances model stability and fairness through structure-free feature augmentations, including noise injection and counterfactual generation. Extensive experiments on two real-world financial datasets and one social network benchmark demonstrate that our approach consistently outperforms existing distillation baselines, achieving a favorable trade-off between predictive utility, stability, and fairness.
Extrinsic regression of time series data consists in predicting the value of a numerical target variable using an input vector which is a time series. The target variable is considered as “extrinsic” as it is not of the same nature as the series values and may not necessarily follow the temporal continuity of the series. This formalization addresses a wide range of problems in different application areas, such as environmental, health or sentiment analysis. In line with the literature on supervised classification of time series, some classification methods have been adapted to the task of regression. Existing regression methods are diverse and use different paradigms, e.g. distance-based methods, interval-based or neural network-based approaches. In parallel to these developments, several libraries for unsupervised feature extraction from time series data have been developed, primarily for descriptive analysis and visualization purposes. In this paper, we combine existing regression methods with signal processing libraries that extract features from time series. To that purpose, the potential of 10 libraries, for the extrinsic regression task, across a set of 61 datasets and six usual regressors is evaluated. The comparative analysis of results from over 3,000 learning ex-periments suggests that unsupervised feature extraction achieves competitive performance for extrinsic regression.
Anomaly detection is a cornerstone of machine learning with applications spanning healthcare, fraud detection, and scientific discovery. Despite extensive research, fair bench-marking remains a significant challenge due to the unsupervised nature of anomaly detection. Hyperparameter selection, a crucial determinant of algorithm performance, is often overlooked or bi-ased, leading to inflated or misleading results. Current practices, including reliance on default configurations, random choices, or limited optimization, hinder reproducibility and impede progress. This work presents a novel pipeline for standardized hyperparameter optimization in anomaly detection. Leveraging a curated collection of nearly 500 datasets, the largest of its kind, our approach systematically optimizes over 80 hyperparameters for 13 widely used anomaly detection algorithms. Our comparison re-veals that the performance variance from hyperparameters often surpasses inter-algorithm differences, emphasizing the need for hyperparameter-specific evaluations. We establish a reproducible foundation for anomaly detection research by providing open-access datasets and code. Our findings not only challenge existing evaluation norms but also pave the way for more robust and reliable comparisons toward better anomaly detection research.
Data-driven decision models based on Artificial Intelligence (AI) are increasingly adopted across domains. However, these models are susceptible to bias that can result in unfair or discriminatory outcomes. Recent research has explored causal discovery methods as a promising way to understand and improve fairness in decision-making systems. In this work, we investigate how different conditional independence tests used in constraint-based causal discovery algorithms, specifically the PC algorithm, affect fairness and performance. We perform an empirical evaluation on several datasets, including Portuguese public contracts, COMPAS, and the German Credit dataset. Using seven conditional independence tests, we assess model behavior under fairness (demographic parity, accuracy parity, equalized odds and predictive rate parity) and performance (accuracy, F1-score, AUC) metrics. Our findings reveal that some tests, due to their statistical properties, fail to expose unfairness detectable via causal structures, even when performance metrics appear acceptable. Furthermore, we highlight significant differences in computational efficiency among the tests, with x2-adf, sp-mi, and sp-x2 being the least efficient. This study underscores the need for careful selection of conditional independence tests in causal discovery to ensure both fairness and reliability in data-driven decision systems.
Surgical gauze is an essential part of surgical procedures, which is primarily used for controlling bleeding and absorbing bodily fluids. The post-surgical retention of gauze can lead to serious complications in the patient's health and necessitate additional surgery for gauze removal. In the wake of data scarcity, the research on gauze segmentation on the real-world surgical data remains underexplored. In this work, we investigate the use of deep learning methods for gauze segmentation in robotassisted minimally invasive abdominal surgeries, utilizing an inhouse surgical dataset prepared at a university hospital. The training data reflects a realistic surgical setting and extensive diversity in spatial, morphological, and visual attributes of three different gauze categories. We have investigated prevalently used segmentation architectures, including CNN-based, transformer-based, and hybrid architectures, to provide a proof-of-concept for gauze segmentation in a realistic setting. Besides, we investigate the influence of additional sub-optimally annotated, auto-tracked segmentation masks to address the bottleneck of data scarcity and performance optimization. Our results demonstrate the efficacy of real-world data to counter the main challenge reported by prior works - the trade-off between blood presence and gauze detection. The incorporation of auto-track annotations enables performance enhancements, particularly in generic cases. The integration of effective segmentation approaches will benefit robotguided surgical procedures and various downstream applications by providing a precise delineation of foreign objects, enhancing patient safety and surgical outcomes.
