Frontiers in Big Data最新文献

The rise of Large Language Models (LLMs), such as LLaMA and ChatGPT, has opened new opportunities for enhancing recommender systems through improved explainability. This paper provides a systematic literature review focused on leveraging LLMs to generate explanations for recommendations-a critical aspect for fostering transparency and user trust. We conducted a comprehensive search within the ACM Guide to Computing Literature, covering publications from the launch of ChatGPT (November 2022) to the present (November 2024). Our search yielded 232 articles, but after applying inclusion criteria, only six were identified as directly addressing the use of LLMs in explaining recommendations. This scarcity highlights that, despite the rise of LLMs, their application in explainable recommender systems is still in an early stage. We analyze these select studies to understand current methodologies, identify challenges, and suggest directions for future research. Our findings underscore the potential of LLMs improving explanations of recommender systems and encourage the development of more transparent and user-centric recommendation explanation solutions.

Introduction: Ambient acoustic detection and localization play a vital role in identifying events and their origins from acoustic data. This study aimed to establish a comprehensive framework for classifying activities in home environments to detect emergency events and transmit emergency signals. Localization enhances the detection of the acoustic event's location, thereby improving the effectiveness of emergency services, situational awareness, and response times.

Methods: Acoustic data were collected from a home environment using six strategically placed microphones in a bedroom, kitchen, restroom, and corridor. A total of 512 audio samples were recorded from 11 activities. Background noise was eliminated using a filtering technique. State-of-the-art features were extracted from the time domain, frequency domain, time frequency domain, and cepstral domain to develop efficient detection and localization frameworks. Random forest and linear discriminant analysis classifiers were employed for event detection, while the estimation signal parameters through rational-in-variance techniques (ESPRIT) algorithm was used for sound source localization.

Results: The study achieved high detection accuracy, with random forest and linear discriminant analysis classifiers attaining 95% and 87%, respectively, for event detection. For sound source localization, the proposed framework demonstrated significant performance, with an error rate of 3.61, a mean squared error (MSE) of 14.98, and a root mean squared error (RMSE) of 3.87.

Discussion: The integration of detection and localization models facilitated the identification of emergency activities and the transmission of notifications via electronic mail. The results highlight the potential of the proposed methodology to develop a real-time emergency alert system for domestic environments.

This study examines privacy calculus in online child sexual abuse material (CSAM) detection across Europe, using Flash Eurobarometer 532 data. Drawing on theories of structuration and risk society, we analyze how individual agency and institutional frameworks interact in shaping privacy attitudes in high-stakes digital scenarios. Multinomial regression reveals age as a significant individual-level predictor, with younger individuals prioritizing privacy more. Country-level analysis shows Central and Eastern European nations have higher privacy concerns, reflecting distinct institutional and cultural contexts. Notably, the Digital Economy and Society Index (DESI) shows a positive association with privacy concerns in regression models when controlling for Augmented Human Development Index (AHDI) components, contrasting its negative bivariate correlation. Life expectancy emerges as the strongest country-level predictor, negatively associated with privacy concerns, suggesting deep institutional mechanisms shape privacy attitudes beyond individual factors. This dual approach reveals that both individual factors and national contexts are shaping privacy calculus in CSAM detection. The study contributes to a better understanding of privacy calculus in high-stakes scenarios, with implications for policy development in online child protection.

Advances in high-throughput genome sequencing have enabled large-scale genome sequencing in clinical practice and research studies. By analyzing genomic variants of humans, scientists can gain better understanding of the risk factors of complex diseases such as cancer and COVID-19. To model and analyze the rich genomic data, knowledge graphs (KGs) and graph machine learning (GML) can be regarded as enabling technologies. In this article, we present a scalable tool called VariantKG for analyzing genomic variants of humans modeled using KGs and GML. Specifically, we used publicly available genome sequencing data from patients with COVID-19. VariantKG extracts variant-level genetic information output by a variant calling pipeline, annotates the variant data with additional metadata, and converts the annotated variant information into a KG represented using the Resource Description Framework (RDF). The resulting KG is further enhanced with patient metadata and stored in a scalable graph database that enables efficient RDF indexing and query processing. VariantKG employs the Deep Graph Library (DGL) to perform GML tasks such as node classification. A user can extract a subset of the KG and perform inference tasks using DGL. The user can monitor the training and testing performance and hardware utilization. We tested VariantKG for KG construction by using 1,508 genome sequences, leading to 4 billion RDF statements. We evaluated GML tasks using VariantKG by selecting a subset of 500 sequences from the KG and performing node classification using well-known GML techniques such as GraphSAGE, Graph Convolutional Network (GCN) and Graph Transformer. VariantKG has intuitive user interfaces and features enabling a low barrier to entry for KG construction, model inference, and model interpretation on genomic variants of humans.

Background: Leukemia is the 11^th most prevalent type of cancer worldwide, with acute myeloid leukemia (AML) being the most frequent malignant blood malignancy in adults. Microscopic blood tests are the most common methods for identifying leukemia subtypes. An automated optical image-processing system using artificial intelligence (AI) has recently been applied to facilitate clinical decision-making.

Aim: To evaluate the performance of all AI-based approaches for the detection and diagnosis of acute myeloid leukemia (AML).

Methods: Medical databases including PubMed, Web of Science, and Scopus were searched until December 2023. We used the "metafor" and "metagen" libraries in R to analyze the different models used in the studies. Accuracy and sensitivity were the primary outcome measures.

Results: Ten studies were included in our review and meta-analysis, conducted between 2016 and 2023. Most deep-learning models have been utilized, including convolutional neural networks (CNNs). The common- and random-effects models had accuracies of 1.0000 [0.9999; 1.0001] and 0.9557 [0.9312, and 0.9802], respectively. The common and random effects models had high sensitivity values of 1.0000 and 0.8581, respectively, indicating that the machine learning models in this study can accurately detect true-positive leukemia cases. Studies have shown substantial variations in accuracy and sensitivity, as shown by the Q values and I² statistics.

Conclusion: Our systematic review and meta-analysis found an overall high accuracy and sensitivity of AI models in correctly identifying true-positive AML cases. Future research should focus on unifying reporting methods and performance assessment metrics of AI-based diagnostics.

Systematic review registration: https://www.crd.york.ac.uk/prospero/#recordDetails, CRD42024501980.

Predicting the dynamics of chaotic systems is crucial across various practical domains, including the control of infectious diseases and responses to extreme weather events. Such predictions provide quantitative insights into the future behaviors of these complex systems, thereby guiding the decision-making and planning within the respective fields. Recently, data-driven approaches, renowned for their capacity to learn from empirical data, have been widely used to predict chaotic system dynamics. However, these methods rely solely on historical observations while ignoring the underlying mechanisms that govern the systems' behaviors. Consequently, they may perform well in short-term predictions by effectively fitting the data, but their ability to make accurate long-term predictions is limited. A critical challenge in modeling chaotic systems lies in their sensitivity to initial conditions; even a slight variation can lead to significant divergence in actual and predicted trajectories over a finite number of time steps. In this paper, we propose a novel Physics-Guided Learning (PGL) method, aiming at extending the scope of accurate forecasting as much as possible. The proposed method aims to synergize observational data with the governing physical laws of chaotic systems to predict the systems' future dynamics. Specifically, our method consists of three key elements: a data-driven component (DDC) that captures dynamic patterns and mapping functions from historical data; a physics-guided component (PGC) that leverages the governing principles of the system to inform and constrain the learning process; and a nonlinear learning component (NLC) that effectively synthesizes the outputs of both the data-driven and physics-guided components. Empirical validation on six dynamical systems, each exhibiting unique chaotic behaviors, demonstrates that PGL achieves lower prediction errors than existing benchmark predictive models. The results highlight the efficacy of our design of data-physics integration in improving the precision of chaotic system dynamics forecasts.

Atmospheric ozone chemistry involves various substances and reactions, which makes it a complex system. We analyzed data recorded by Switzerland's National Air Pollution Monitoring Network (NABEL) to showcase the capabilities of machine learning (ML) for the prediction of ozone concentrations (daily averages) and to document a general approach that can be followed by anyone facing similar problems. We evaluated various artificial neural networks and compared them to linear as well as non-linear models deduced with ML. The main analyses and the training of the models were performed on atmospheric air data recorded from 2016 to 2023 at the NABEL station Lugano-Università in Lugano, TI, Switzerland. As a first step, we used techniques like best subset selection to determine the measurement parameters that might be relevant for the prediction of ozone concentrations; in general, the parameters identified by these methods agree with atmospheric ozone chemistry. Based on these results, we constructed various models and used them to predict ozone concentrations in Lugano for the period between January 1, 2024, and March 31, 2024; then, we compared the output of our models to the actual measurements and repeated this procedure for two NABEL stations situated in northern Switzerland (Dübendorf-Empa and Zürich-Kaserne). For these stations, predictions were made for the aforementioned period and the period between January 1, 2023, and December 31, 2023. In most of the cases, the lowest mean absolute errors (MAE) were provided by a non-linear model with 12 components (different powers and linear combinations of NO₂, NO_X, SO₂, non-methane volatile organic compounds, temperature and radiation); the MAE of predicted ozone concentrations in Lugano was as low as 9 μgm^-3. For the stations in Zürich and Dübendorf, the lowest MAEs were around 11 μgm^-3 and 13 μgm^-3, respectively. For the tested periods, the accuracy of the best models was approximately 1 μgm^-3. Since the aforementioned values are all lower than the standard deviations of the observations we conclude that using ML for complex data analyses can be very helpful and that artificial neural networks do not necessarily outperform simpler models.

Predictions of student performance are important to the education system as a whole, helping students to know how their learning is changing and adjusting teachers' and school policymakers' plans for their future growth. However, selecting meaningful features from the huge amount of educational data is challenging, so the dimensionality of student achievement features needs to be reduced. Based on this motivation, this paper proposes an improved Binary Snake Optimizer (MBSO) as a wrapped feature selection model, taking the Mat and Por student achievement data in the UCI database as an example, and comparing the MBSO feature selection model with other feature methods, the MBSO is able to select features with strong correlation to the students and the average number of student features selected reaches a minimum of 7.90 and 7.10, which greatly reduces the complexity of student achievement prediction. In addition, we propose the MDBO-BP-Adaboost model to predict students' performance. Firstly, the model incorporates the good point set initialization, triangle wandering strategy and adaptive t-distribution strategy to obtain the Modified Dung Beetle Optimization Algorithm (MDBO), secondly, it uses MDBO to optimize the weights and thresholds of the BP neural network, and lastly, the optimized BP neural network is used as a weak learner for Adaboost. MDBO-BP-Adaboost After comparing with XGBoost, BP, BP-Adaboost, and DBO-BP-Adaboost models, the experimental results show that the R² on the student achievement dataset is 0.930 and 0.903, respectively, which proves that the proposed MDBO-BP-Adaboost model has a better effect than the other models in the prediction of students' achievement with better results than other models.

Traditional rainfall data collection mainly relies on rain buckets and meteorological data. It rarely considers the impact of sensor faults on measurement accuracy. To solve this problem, a two-layer genetic algorithm-backpropagation (GA-BP) model is proposed. The algorithm focuses on multi-source data identification and fusion. Rainfall data from a sensor array are first used. The GA optimizes the weights and thresholds of the BP neural network. It determines the optimal population and minimizes fitness values. This process builds a GA-BP model for recognizing sensor faults. A second GA-BP network is then created based on fault data. This model achieves data fusion output. The two-layer GA-BP algorithm is compared with a single BP neural network and actual expected values to test its performance. The results show that the two-layer GA-BP algorithm reduces data fusion runtime by 2.37 s compared to the single-layer BP model. For faults such as lost signals, high-value bias, and low-value bias, recognition accuracies improve by 26.09%, 18.18%, and 7.15%, respectively. The mean squared error is 3.49 mm lower than that of the single-layer BP model. The fusion output waveform is also smoother with less fluctuation. These results confirm that the two-layer GA-BP model improves system robustness and generalization.