Machine learning with applications最新文献

Intelligent search engines including pre-trained generative transformers (GPT) have revolutionized the user search experience. Several fields including e-commerce, education, and hospitality are increasingly exploring GPT tools to study user reviews and gain critical insights to improve their service quality. However, massive user-review data and imprecise prompt engineering lead to biased, irrelevant, and impersonal search results. In addition, exposing user data to these search engines may pose privacy issues. Motivated by these factors, we present ChatReview, a ChatGPT-enabled natural language processing (NLP) framework that effectively studies domain-specific user reviews to offer relevant and personalized search results at multiple levels of granularity. The framework accomplishes this task using four phases including data collection, tokenization, query construction, and response generation. The data collection phase involves gathering domain-specific user reviews from public and private repositories. In the tokenization phase, ChatReview applies sentiment analysis to extract keywords and categorize them into various sentiment classes. This process creates a token repository that best describes the user sentiments for a given user-review data. In the query construction phase, the framework uses the token repository and domain knowledge to construct three types of ChatGPT prompts including explicit, implicit, and creative. In the response generation phase, ChatReview pipelines these prompts into ChatGPT to generate search results at varying levels of granularity. We analyze our framework using three real-world domains including education, local restaurants, and hospitality. We assert that our framework simplifies prompt engineering for general users to produce effective results while minimizing the exposure of sensitive user data to search engines. We also present a one-of-a-kind Large Language Model (LLM) peer assessment of the ChatReview framework. Specifically, we employ Google’s Bard to objectively and qualitatively analyze the various ChatReview outputs. Our Bard-based analyses yield over 90% satisfaction, establishing ChatReview as a viable survey analysis tool.

In this paper, we propose using LSTM-RNNs (Long Short-Term Memory-Recurrent Neural Networks) to learn and represent nonlinear integral operators that appear in nonlinear integro-differential equations (IDEs). The LSTM-RNN representation of the nonlinear integral operator allows us to turn a system of nonlinear integro-differential equations into a system of ordinary differential equations for which many efficient solvers are available. Furthermore, because the use of LSTM-RNN representation of the nonlinear integral operator in an IDE eliminates the need to perform a numerical integration in each numerical time evolution step, the overall temporal cost of the LSTM-RNN-based IDE solver can be reduced to $O\left(n_T\right)$ from $O\left(n_T^2\right)$ if a $n_T$-step trajectory is to be computed. We illustrate the efficiency and robustness of this LSTM-RNN-based numerical IDE solver with a model problem. Additionally, we highlight the generalizability of the learned integral operator by applying it to IDEs driven by different external forces. As a practical application, we show how this methodology can effectively solve the Dyson’s equation for quantum many-body systems.

PM_2.5, inhalable particles, with a size of 2.5 micrometers or less, negatively impact the environment as well as our health. Monitoring PM_2.5 is critical to guard against extreme events by alerting people and initiating actions to alleviate PM_2.5′s impacts. Developing PM_2.5 forecasting frameworks empowers the authorities to predict extremely polluted events in advance and gives them time to implement necessary strategies in advance (e.g., Action! Days). Understanding the spatiotemporal behavior of PM_2.5 and meteorological factors is of significance for having accurate predictions. This study utilizes EPA sensor data to quantify the PM_2.5 air quality index (AQI) and meteorological factors such as temperature over 2015–2019 across Michigan, USA. Here, a spatiotemporal deep neural structure is proposed through integrating graph convolutional neural (GCN) and exogenous long short-term memory (E-LSTM) networks to incorporate spatial and temporal patterns within PM_2.5 AQI and meteorological factors for predicting PM_2.5 AQI. Results illustrate that not only does our proposed framework outperform the traditional approaches such as LSTM and E-LSTM, but also it is robust against the network structure of EPA stations. The study's findings demonstrate that the integration of GCN with E-LSTM significantly enhances the accuracy of PM_2.5 AQI predictions compared to traditional models. This advancement indicates a promising direction for environmental monitoring, offering improved forecasting tools that can aid in timely and effective decision-making for air quality management and public health protection.

In collaborative filtering, matrix factorization and collaborative metric learning are challenged by situations where non-preferred items may appear so close to a user in the feature embedding space that they lead to degrading the recommendation performance. We call such items ‘potential impostor’ risks. Addressing the issues with ‘potential impostor’ is important because it can result in inefficient learning and poor feature extraction. To achieve this, we propose a novel loss function formulation designed to enhance learning efficiency by actively identifying and addressing impostors, leveraging item associations and learning the distribution of negative items. This approach is crucial for models to differentiate between positive and negative items effectively, even when they are closely aligned in the feature space. Here, a loss function is generally an objective optimization function that is defined based on user–item interaction data, through either implicit or explicit feedback. The loss function essentially decides how well a recommendation algorithm performs. In this paper, we introduce and define the concept of ‘potential impostor’, highlighting its impact on learned representation quality and algorithmic efficiency. We tackle the limitations of non-metric methods, like the Weighted Approximate Rank Pairwise Loss (WARP) method, which struggles to capture item–item similarities, by using a ‘similarity propagation’ strategy with a new loss term. Similarly, we address fixed margin inefficiencies in Weighted Collaborative Metric Learning (WCML), through density distribution approximation. This moves potential impostors away from the margin for more robust learning. Additionally, we propose a large-scale batch approximation algorithm for increased detection of impostors, coupled with an active learning strategy for improved top-$N$ recommendation performance. Our extensive empirical analysis across five major and diverse datasets demonstrates the effectiveness and feasibility of our methods, compared to existing techniques with respect to improving AUC, reducing impostor rate, and increasing the average distance metrics. More specifically, our evaluation shows that our two proposed methods outperform the existing state-of-the-art techniques, with an improvement of AUC by 3.5% and 3.7%, NDCG by 1.0% and 9.1% and HR by 1.3% and 3.6%, respectively. Similarly, the impostor rate is decreased by 35% and 18%, and their average distance is increased by 33% and 37%, respectively.

Drilling and blasting operations are pivotal for productivity and safety in hard rock surface mining. These operations are restricted due to complexities such as site-specific uncertainties, safety risks, and environmental and economic constraints. Machine Learning (ML) is a transformative approach to tackle these complexities resulting in significant cost reductions. ML applications can reduce overall blasting costs by up to 23% and decrease the amount of explosives by as much as 89% compared to traditional methods. This survey presents a comprehensive review of how ML can be applied to optimize drill and blast designs while accounting for its operational challenges. Our research highlights the difficulties in collecting quality site-specific data, the complexity of interpreting this data into insightful information, the selection of ML models relating to mining objectives, and the need for established methods to assess blast efficiency quantitatively. We provide a synthesis of ML model development practices in drilling and blasting and demonstrate the value of ML methodologies. Based on our survey, we present actionable recommendations for developing ML methodologies to improve safety, reduce costs, and enhance efficiency in drilling and blasting processes. This includes establishing standardized data schematics, multiobjective model optimization, and comprehensive evaluation metrics. These benefits can guide mine management and engineers to adopt ML techniques and improve on-ground operational practices. This survey aims to serve as a resource for both practitioners and researchers shaping the future research direction in ML applications for drilling and blasting practices.

Background

This study aimed to identify the most impactful set of intrinsic and extrinsic fall risk factors and develop a data-driven inpatient fall risk assessment tool (FRAT).

Methods

The dataset used for the study comprised in-hospital fall records from 2012 to 2017. Four machine learning (ML) algorithms, Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting (Gboost), and Deep Neural Network (DNN) were utilized to predict the inpatient fall risk level. To enhance the performance of the prediction models, two approaches were implemented, including (1) feature selection to identify the optimal feature set and (2) the development of three distinct shift-wise models. Furthermore, the optimal feature sets in the shift-wise models were extracted.

Results

According to the results, DNN outperformed other methods by reaching an accuracy, sensitivity, specificity, and AUC of 0.71, 0.8, 0.6, and 0.7, respectively, considering the full set of features. The performance of the models was further improved (by 3-5 %) by conducting a feature selection process for all models. Specifically, the DNN model achieved an accuracy of 0.74 while considering the optimal set of predictors. Moreover, the shift-wise RF models demonstrated higher accuracies (by 4-10 %) compared to the same model using a full feature set.

Conclusions

This study's outcome confirms ML models' compelling capability in developing an inpatient FRAT while considering intrinsic and extrinsic factors. The insight from such models could form a foundation to (1) monitor the inpatients’ fall risk, (2) identify the major factors involved in inpatient falls, and (3) create subject-specific self-care plans.

Low-lying coastal cities, exemplified by Norfolk, Virginia, face the challenge of street flooding caused by rainfall and tides, which strain transportation and sewer systems and can lead to personal and property damage. While high-fidelity, physics-based simulations provide accurate predictions of urban pluvial flooding, their computational complexity renders them unsuitable for real-time applications. Using data from Norfolk rainfall events between 2016 and 2018, this study compares the performance of a previous surrogate model based on a random forest algorithm with two deep learning models: Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). The comparison of deep learning to the random forest algorithm is motivated by the desire to utilize a machine learning architecture that allows for the future inclusion of common uncertainty quantification techniques and the effective integration of relevant, multi-modal features.

Following the very recent launch of the ChatGPT chatbot, numerous comments and speculations were posted concerning the potential aspects of society that are expected to benefit from this AI revolution. In particular, the education sector is considered as one of the primary domains affected by this application, the impact of which remains yet to be fully understood. Furthermore, many Higher Education institutions are required to get to terms with its impact on teaching and learning, and to clarify their stances on the use of ChatGPT software. This study was developed to investigate some critical case studies considered as relevant to the inevitable re-evaluation of educational aspects needed, ranging from academic missions to student and course learning outcomes and its ethical uses. Following a review of some of the pros and cons of ChatGPT in the higher educational sector, this paper shall demonstrate several case studies of early trials in teaching and learning assessments related to various specializations. Next, the ability of some well-known AI detector software and analyzed in terms of their capacity to successfully detect AI-generated content. Analysis shall be made of the foreseen impact on important aspects including challenges and benefits related to its use in course assessments as well as academic integrity and ethical use. The study concludes with a set of recommendations made from our findings and benchmarks obtained from top universities in order to assist faculty members and decision makers at Higher Education institutions concerning their response strategy and use of ChatGPT.

Predictive modeling in healthcare continues to be an active actuarial research topic as more insurance companies aim to maximize the potential of Machine Learning (ML) approaches to increase their productivity and efficiency. In this paper, the authors deployed three regression-based ensemble ML models that combine variations of decision trees through Extreme Gradient Boosting (XGBoost), Gradient-boosting Machine (GBM), and Random Forest (RF) methods in predicting medical insurance costs. Explainable Artificial Intelligence (XAi) methods SHapley Additive exPlanations (SHAP) and Individual Conditional Expectation (ICE) plots were deployed to discover and explain the key determinant factors that influence medical insurance premium prices in the dataset. The dataset used comprised 986 records and is publicly available in the KAGGLE repository. The models were evaluated using four performance evaluation metrics, including R-squared (R²), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). The results show that all models produced impressive outcomes; however, the XGBoost model achieved a better overall performance although it also expanded more computational resources, while the RF model recorded a lesser prediction error and consumed far fewer computing resources than the XGBoost model. Furthermore, we compared the outcome of both XAi methods in identifying the key determinant features that influenced the PremiumPrices for each model and whereas both XAi methods produced similar outcomes, we found that the ICE plots showed in more detail the interactions between each variable than the SHAP analysis which seemed to be more high-level. It is the aim of the authors that the contributions of this study will help policymakers, insurers, and potential medical insurance buyers in their decision-making process for selecting the right policies that meet their specific needs.

Many clinical spaces are outfitted with centralized video recording systems to monitor patient–client interactions. Considering the increasing interest in video-based machine learning methods, the potential of using these clinical recordings to automate observational data collection is apparent. To explore this, seven patients had videos of their functional assessment and treatment sessions annotated by coders trained by our clinical team. Commonly used clinical software has inherent limitations aligning behavioral and video data, so a custom software tool was employed to address this functionality gap. After developing a Canvas-based coder training course for this tool, a team of six trained coders annotated 82.33 h of data. Two machine learning approaches were considered, where both used a convolutional neural network as a video feature extractor. The first approach used a recurrent network as the classifier on the extracted features and the second used a Transformer architecture. Both models produced promising metrics indicating that the capability of detecting aggression from clinical videos is possible and generalizable. Model performance is directly tied to the feature extractor’s performance on ImageNet, where ConvNeXtXL produced the best performing models. This has applications in automating patient incident response to improve patient and clinician safety and could be directly integrated into existing video management systems for real-time analysis.