Single Machine Learning Model Research Articles

Experimental analysis of intrusion detection systems using machine learning algorithms and artificial neural networks

Since the invention of the internet for military and academic research purposes, it has evolved to meet the demands of the increasing number of users on the network, who have their scope beyond military and academics. As the scope of the network expanded maintaining its security became a matter of increasing importance. With various users and interconnections of more diversified networks, the internet needs to be maintained as securely as possible for the transmission of sensitive information to be one hundred per cent safe; several anomalies may intrude on private networks. Several research works have been released around network security and this research seeks to add to the already existing body of knowledge by expounding on these attacks, proffering efficient measures to detect network intrusions, and introducing an ensemble classifier: a combination of 3 different machine learning algorithms. An ensemble classifier is used for detecting remote to local (R2L) attacks, which showed the lowest level of accuracy when the network dataset is tested using single machine learning models but the ensemble classifier gives an overall efficiency of 99.8%.

Exchange rate prediction research based on LSTM-ELM hybrid model

The fluctuation of exchange rates holds paramount importance for a country's economic and trade activities. Due to the non-stationary and nonlinear structural characteristics of exchange rate time series, accurately predicting exchange rate movements is a challenging task. Single machine learning models often exhibit lower precision in exchange rate prediction compared to combined machine learning models. Hence, employing a combined model approach aims to enhance the predictive performance of exchange rate models. Both Long Short-Term Memory (LSTM) and Extreme Learning Machine (ELM) exhibit intricate structures, making their direct integration challenging. To address this issue, an innovative weighted approach is adopted in this study, combining LSTM and ELM models and further refining the combination weights using an improved Marine Predators Algorithm. This paper encompasses both univariate and multivariate prediction scenarios, employing two distinct allocation strategies for training and testing datasets. This is done to investigate the influence of different dataset allocations on exchange rate prediction. Finally, the proposed LSTM-ELM weighted combination exchange rate prediction model is compared with SVM, Random Forest, ELM, LSTM, and LSTM-ELM average combination models. Experimental results demonstrate that the LSTM-ELM weighted combination exchange rate prediction model outperforms the others in both univariate and multivariate prediction settings, yielding higher predictive accuracy and superior fitting performance. Consequently, the LSTM-ELM weighted combination prediction model proves to be effective in exchange rate forecasting.

Understanding imbalanced data: XAI & interpretable ML framework

There is a gap between current methods that explain deep learning models that work on imbalanced image data and the needs of the imbalanced learning community. Existing methods that explain imbalanced data are geared toward binary classification, single layer machine learning models and low dimensional data. Current eXplainable Artificial Intelligence (XAI) techniques for vision data mainly focus on mapping predictions of specific instances to inputs, instead of examining global data properties and complexities of entire classes. Therefore, there is a need for a framework that is tailored to modern deep networks, that incorporates large, high dimensional, multi-class datasets, and uncovers data complexities commonly found in imbalanced data. We propose a set of techniques that can be used by both deep learning model users to identify, visualize and understand class prototypes, sub-concepts and outlier instances; and by imbalanced learning algorithm developers to detect features and class exemplars that are key to model performance. The components of our framework can be applied sequentially in their entirety or individually, making it fully flexible to the user’s specific needs (https://github.com/dd1github/XAI_for_Imbalanced_Learning).

Enhancing Solar Radiation Forecasting in Diverse Moroccan Climate Zones: A Comparative Study of Machine Learning Models with Sugeno Integral Aggregation

Hourly solar radiation (SR) forecasting is a vital stage in the efficient deployment of solar energy management systems. Single and hybrid machine learning (ML) models have been predominantly applied for precise hourly SR predictions based on the pattern recognition of historical heterogeneous weather data. However, the integration of ML models has not been fully investigated in terms of overcoming irregularities in weather data that may degrade the forecasting accuracy. This study investigated a strategy that highlights interactions that may exist between aggregated prediction values. In the first investigation stage, a comparative analysis was conducted utilizing three different ML models including support vector machine (SVM) regression, long short-term memory (LSTM), and multilayer artificial neural networks (MLANN) to provide insights into their relative strengths and weaknesses for SR forecasting. The comparison showed the proposed LSTM model had the greatest contribution to the overall prediction of six different SR profiles from numerous sites in Morocco. To validate the stability of the proposed LSTM, Taylor diagrams, violin plots, and Kruskal–Wallis (KW) tests were also utilized to determine the robustness of the model’s performance. Secondly, the analysis found coupling the models outputs with aggregation techniques can significantly improve the forecasting accuracy. Accordingly, a novel aggerated model that integrates the forecasting outputs of LSTM, SVM, MLANN with Sugeno λ-measure and Sugeno integral named (SLSM) was proposed. The proposed SLSM provides spatially and temporary interactions of information that are characterized by uncertainty, emphasizing the importance of the aggregation function in mitigating irregularities associated with SR data and achieving an hourly time scale forecasting accuracy with improvement of 11.7 W/m2.

A bearing fault diagnosis method with improved symplectic geometry mode decomposition and feature selection

A rolling bearing fault diagnosis method based on improved symplectic geometry mode decomposition (SGMD) and feature selection was proposed to solve the problem of low fault identification due to the influence of noise on early bearing fault features. First, the SGMD SGMD is improved to enhance its robustness in decomposing signals with noise, then the time domain, frequency domain, and time-frequency features of each symplectic geometric component are extracted as feature vectors. Second, a comprehensive feature selection strategy is proposed to select the optimal subset of features that are conducive to fault classification. Finally, considering the problem of low classification accuracy of a single machine learning model, the AdaBoost-WSO-SVM model is constructed for fault classification using the AdaBoost algorithm of integrated learning. Experimental decomposition of complex signals with noise indicates that the improved SGMD is more effective compared to traditional SGMD. Subsequently, multiple experiments were conducted using the bearing datasets from Case Western Reserve University (CWRU) and Jiangnan University (JNU). The experimental results reveal that, after comprehensive feature selection and ensemble learning pattern recognition experiments on the CWRU dataset, the average accuracy of fault diagnosis can reach 99.67%. On the JNU dataset, the proposed fault diagnosis method achieves an average accuracy of 95.03%. This suggests that, compared to other feature selection methods and classification models, the proposed approach in this paper exhibits higher accuracy and generalization capabilities.

Fatigue life prediction of the FCC-based multi-principal element alloys via domain knowledge-based machine learning

We propose a machine learning (ML) model to predict the fatigue life of multi-principal element alloys (MPEAs) by extracting features from empirical formulas. The model is built on XGBoost and GBDT, and outperforms the single ML model, with almost all predictions lying in the ± 2 error bands and the relative error not exceeding 0.16 in the extrapolation test. Feature analysis shows that for the nine explored MPEAs systems, their S–N curves are more suitable to be fitted by logN=a+blogσmax. Interpretable analysis indicates that for the explored alloys, elongation > 47% benefits the increase of fatigue life; if their yield strength is less than 720 MPa, improving strength will favor improvement in lifetime, otherwise improving ductility will favor lengthening their lifetime. It provides a fast and low-cost method to predict the fatigue life of those FCC-based MPEAs, which guides designing alloys with longer fatigue life.

Machine-Learning-Based Solution Predicts Fluid Properties for Gas-Injection Data

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 211080, “Machine-Learning-Based Prediction of Pressure/Volume/Temperature Fluid Properties for Gas-Injection Laboratory Data,” by Kassem Ghorayeb, SPE, American University of Beirut and Schlumberger; Kristian Mogensen, SPE, ADNOC; and Nour El Droubi, Schlumberger, et al. The paper has not been peer reviewed. _ While machine learning (ML) is used extensively to predict black-oil properties, it is used less often for compositional reservoir properties, including those related to gas injection. Can typically extensive conventional laboratory data be used to help predict the necessary gas-injection parameters? This question is addressed in the complete paper. The authors present an ML-based solution that predicts pertinent gas-injection studies from known fluid properties such as fluid composition and black-oil properties—that is, learning from samples with gas-injection laboratory studies and predicting gas-injection fluid parameters for the remaining, much larger data set. Methodology The objective of the ML component is to predict the results of the swelling test using compositional and black-oil properties. Fig. 1 illustrates the three main steps that make up the work flow. The swelling-test data is processed and analyzed for use in the ML training as the first phase in the data-preparation and -analysis process. The following step is ML training, where features are engineered to examine correlations in the data, followed by the training of ML models using selected features and evaluation of the results. The swelling-test data for each fluid sample will be predicted using the trained ML models in the last stage, which creates a digital-twin database for swelling-test results. The three steps are detailed in the complete paper. Results and Discussion Several approaches were taken while training the ML model. One approach was to either cluster before the ML training, with one ML model trained for each cluster, or train a single ML model using all the data. The other approach was to use the chain method for the prediction of the outputs. In the chain approach, the outputs are given a specific order and, whenever one output is predicted, it is added to the inputs of the following output. The chain approach is particularly useful when the outputs are correlated.

Quantitative investigation and intelligent forecasting of thermal conductivity in lime-modified red clay.

This paper delves into the engineering applications of lime-stabilized red clay, a highly water-sensitive material, particularly in the context of the climatic conditions prevalent in the Dalian region. We systematically investigate the impact of water content, dry density, and freeze-thaw cycles (with a freezing temperature set at -10°C) on the thermal conductivity of stabilized soil, a crucial parameter for analyzing soil temperature fields that is influenced by numerous factors. By developing and validating both empirical and machine learning prediction models, we unravel the evolution of thermal conductivity in response to these factors: within the range of influencing variables, thermal conductivity exhibits an exponential or linear increase with rising water content and dry density, while it decreases exponentially with increasing freeze-thaw cycles. Furthermore, we quantitatively analyze the specific influence of water content and other factors on the thermal conductivity of stabilized soil and construct a comprehensive prediction model encompassing BP neural network, gradient boosting decision tree, and linear regression models. Comparative analysis highlights the significant enhancement in prediction accuracy achieved by the proposed ensemble model over single machine learning models, with root mean square error (RMSE) values below 0.05 and mean absolute percentage error (MAPE) values remaining under 2.5% in both frozen and unfrozen states. Additionally, a secondary validation using experimental data from other researchers confirms the model's good agreement with previous results, demonstrating its robust generalization ability. Our findings provide valuable insights for engineering studies in the Dalian region and red clay areas subjected to extreme climatic conditions.

A Parallel-Cascaded Ensemble of Machine Learning Models for Crop Type Classification in Google Earth Engine Using Multi-Temporal Sentinel-1/2 and Landsat-8/9 Remote Sensing Data

The accurate mapping of crop types is crucial for ensuring food security. Remote Sensing (RS) satellite data have emerged as a promising tool in this field, offering broad spatial coverage and high temporal frequency. However, there is still a growing need for accurate crop type classification methods using RS data due to the high intra- and inter-class variability of crops. In this vein, the current study proposed a novel Parallel-Cascaded ensemble structure (Pa-PCA-Ca) with seven target classes in Google Earth Engine (GEE). The Pa section consisted of five parallel branches, each generating Probability Maps (PMs) for different target classes using multi-temporal Sentinel-1/2 and Landsat-8/9 satellite images, along with Machine Learning (ML) models. The PMs exhibited high correlation within each target class, necessitating the use of the most relevant information to reduce the input dimensionality in the Ca part. Thereby, Principal Component Analysis (PCA) was employed to extract the top uncorrelated components. These components were then utilized in the Ca structure, and the final classification was performed using another ML model referred to as the Meta-model. The Pa-PCA-Ca model was evaluated using in-situ data collected from extensive field surveys in the northwest part of Iran. The results demonstrated the superior performance of the proposed structure, achieving an Overall Accuracy (OA) of 96.25% and a Kappa coefficient of 0.955. The incorporation of PCA led to an OA improvement of over 6%. Furthermore, the proposed model significantly outperformed conventional classification approaches, which simply stack RS data sources and feed them to a single ML model, resulting in a 10% increase in OA.

A kinetic-assisted growth curve prediction method for Chlamydomonas reinhardtii incorporating transfer learning

Traditional predictions of microalgal growth states rely on empirical or easily implementable kinetic models, leading to significant biases and elevated cost. This study proposes a kinetic-assisted machine learning method for predicting the growth curve of microalgal biomass under small sample conditions. Firstly, a microalgae growth kinetic model is constructed based on the logistic model. A two-stage kinetic fitting strategy is specified to account for the light–dark ratio. The Box-Behnken method is employed for experimental design. Then, using Two-stage TrAdaboost.R2 algorithm, the kinetic model is utilized as the source domain, and the experimental design data serves as the target domain for training machine learning models. The results indicate that the proposed method outperforms a single machine learning model in terms of prediction and has the potential to rapidly estimate microalgal growth trends under different conditions and accurately predict harvested biomass, potentially reducing the need for laborious, expensive, and time-consuming laboratory trials.

Stokes shift prediction of fluorescent organic dyes using machine learning based hybrid cascade models

Fluorescent organic dyes are widely used in various fields, including science and technology, research and development, medicine, and drug delivery. Multitudinous attempts have been made by experimentalists to develop such fluorescent organic dyes with the desired Stokes shift property at negligible cost and time. For quickly and accurately predicting the Stokes shift property of fluorescent organic dye, we proposed eight hybrid models based on the combination of nine single machine-learning models. To fulfill the objective, we considered a dataset of 3066 fluorescent organic materials and evaluated the performance of each model using three evaluation parameters: mean absolute error (MAE), root mean squared error (RMSE), and the coefficient of determination (R2). The hybrid cascade model of Extreme Gradient Boosting Regression and Light Gradient Boosting Machine Regression (XGBR + LGBMR) performed best for Stokes shift prediction, with MAE of 13.83 nm, RMSE of 19.95 nm, and R2 of 86.18 %. The prediction performance of all the undertaken models was validated by the experimental data of four xanthene dyes (Rh-19, Rh–B, Rh-6G, and Rh-110). In this regard, the XGBR + DTR (Extreme Gradient Boosting Regression + Decision Tree Regression) model was the best performer, with errors ranging from 5 to 13 nm for four dyes. The resultant errors are much smaller than the recently reported synthesized material with an error of 30 nm. The proposed models allow for rapid and cost-effective screening of a wide range of fluorescent organic dyes, which assists the researchers in gaining prior knowledge of materials and accelerates the discovery of new materials.

Monitoring of drought stress and transpiration rate using proximal thermal and hyperspectral imaging in an indoor automated plant phenotyping platform

BackgroundThermography is a popular tool to assess plant water-use behavior, as plant temperature is influenced by transpiration rate, and is commonly used in field experiments to detect plant water deficit. Its application in indoor automated phenotyping platforms is still limited and mainly focuses on differences in plant temperature between genotypes or treatments, instead of estimating stomatal conductance or transpiration rate. In this study, the transferability of commonly used thermography analysis protocols from the field to greenhouse phenotyping platforms was evaluated. In addition, the added value of combining thermal infrared (TIR) with hyperspectral imaging to monitor drought effects on plant transpiration rate (E) was evaluated.ResultsThe sensitivity of commonly used TIR indices to detect drought-induced and genotypic differences in water status was investigated in eight maize inbred lines in the automated phenotyping platform PHENOVISION. Indices that normalized plant temperature for vapor pressure deficit and/or air temperature at the time of imaging were most sensitive to drought and could detect genotypic differences in the plants’ water-use behavior. However, these indices were not strongly correlated to stomatal conductance and E. The canopy temperature depression index, the crop water stress index and the simplified stomatal conductance index were more suitable to monitor these traits, and were consequently used to develop empirical E prediction models by combining them with hyperspectral indices and/or environmental variables. Different modeling strategies were evaluated, including single index-based, machine learning and mechanistic models. Model comparison showed that combining multiple TIR indices in a random forest model can improve E prediction accuracy, and that the contribution of the hyperspectral data is limited when multiple indices are used. However, the empirical models trained on one genotype were not transferable to all eight inbred lines.ConclusionOverall, this study demonstrates that existing TIR indices can be used to monitor drought stress and develop E prediction models in an indoor setup, as long as the indices normalize plant temperature for ambient air temperature or relative humidity.

River water temperature prediction using hybrid machine learning coupled signal decomposition: EWT versus MODWT

Accurate prediction of water temperature (Tw) will greatly help in improving our understanding of the overall thermal regime fluctuation, and it can help in making sound decisions. While great efforts have been devoted to the development of Tw models, further improvement in the prediction accuracy is challenging. Here, we propose a new hybrid machine learning (ML) models for predicting Tw from air temperature (Ta). First, two signal decomposition algorithms, i.e., the empirical wavelet transform (EWT) and maximum overlap discrete wavelet transform (MODWT) are used for decomposing the Ta into several subsequences. Second, the obtained subsequences are used as input variables for four ML models, i.e., the feedforward neural network (FFNN), the optimally pruned extreme learning machine (OPELM), the adaptive boosting (AdaBoost), and the bootstrap aggregating (Bagging). The development of the hybrid models is based on measured data form five measurement stations distributed over Poland. The experimental results show that, the new hybrid ML models achieved high predictive accuracies with Pearson correlation coefficient (R), Nash-Sutcliffe efficiency (NSE), root mean square error (RMSE) and mean absolute error (MAE) of approximately ≈0.99, ≈0.98, ≈0.718 °C, and ≈0.599 °C, respectively. Finally, the experimental results have demonstrated that, the MAE and RMSE of the single models were reduced by 66.12% and 65.30%, while the R and NSE values were improved by 4.85% and 10.02%, respectively, showing a very promising prediction performance. Our new modelling approach proposed in the present study clearly highlight the high contribution of the EWT and MODWT in improving the Tw estimation that is most strongly influenced by Ta. The significant difference between single and hybrid models can be explained by the ability of the EWT and MODWT to capture the high nonlinearity between Ta and Tw. Our new approach is of interest for water resources management and for assessing the variability of Tw over time and space. Finally yet importantly, this is the first study in the literature for which EWT and MODWT were used for modelling Tw, through which we have demonstrated that, an important amount of information is available and cannot be captured using single ML models, and signal decomposition have helped to overcome this challenge.

Predicting the risk of in-hospital Asian NSTEMI patients using stacked ensemble learning

Abstract Background NSTEMI affects 60–75% of patients with MI and has a variety of clinical manifestations and prognoses. Ensemble learning (EL) utilizes the machine learning (ML) models on a classification task to make more accurate predictions than any single ML model. It has also been demonstrated to be superior to conventional risk scores in predicting mortality risk. However, no studies on the application of EL to ASIAN NSTEMI patients have been published. Purpose To identify factors associated with Asian in-hospital NSTEMI patients using ML and to develop a stacking EL-based NSTEMI risk score that is specifically tailored to the Asian population. Methods From 2006 to 2016, 10464 data were collected from the Malaysian National Cardiovascular Disease Database (NCVD-ACS) registry on hospitalized multiethnic Asian patients admitted with NSTEMI. This study utilized 39 variables, including demographic, cardiovascular risk, medication, and clinical variables. For stacked EL model development, four algorithms, Support Vector Machine (SVM), Random Forest (RF), eXtreme Gradient Boosting (XGB), and Naive Bayes (NB), are used as base learners, and Generalized Linear Model (GLM) is used as the meta learner. Significant variables were chosen and ranked using ML variable importance with backward elimination. As a predictive performance metric, the area under the curve (AUC) was used. Using a validation dataset, algorithms were validated against the TIMI for NSTEMI, and the net reclassification index (NRI) was calculated. Results The stacked EL models developed using SVM feature selection (13 predictors, AUC = 0.823, CI: 0.779 - 0.867) outperforms the conventional risk score, TIMI (AUC = 0.563, CI: 0.515 - 0.610). RF contributes the most to the EL model predictions, followed by NB, XGB and SVM (Figure 1). Common predictors between SVM feature selection and TIMI are; age, hypertension and ECG-type depression. SVM feature selection also identified gender, MI history, chronic lung disease, heart rate, systolic blood pressure, Killip class, fasting blood glucose, LMWH, beta-blockers and angiotensin II receptor blocker as predictors that improve mortality prediction in ASIAN NSTEMI patients. Our findings indicate that the TIMI score underestimates the mortality risk of NSTEMI patients. Stacked EL model using selected predictors classified 47.4% of nonsurvivors as high risk (risk probabilities > 50%), whereas TIMI (score > 5) classified only 3.3% of nonsurvivors as high risk (Figure 2). The NRI of NSTEMI patients using the EL model (SVM selected var) resulted in an NRI of 79.8% with a p-value < 0.00001 when compared to the TIMI for NSTEMI risk score. Conclusions The stacked EL algorithm classifies ASIAN NSTEMI patients more accurately than the single ML model and TIMI score. A precise prognosis is useful for selecting the appropriate level of care and pharmacological or invasive treatment for patients with Asian NSTEMI.Relative influence of each ML modelsStacked EL performance on validation set

A Comparative Study of Genetic Algorithm-Based Ensemble Models and Knowledge-Based Models for Wildfire Susceptibility Mapping

Wildfire susceptibility mapping (WSM) plays a crucial role in identifying areas with heightened vulnerability to forest fires, allowing for proactive measures in fire prevention, management, and resource allocation, ultimately leading to more effective fire control and mitigation strategies. This paper describes our undertaking to develop and compare the performance of two knowledge-based models, namely the analytic hierarchy process (AHP) and the technique for order performance by similarity to ideal solution (TOPSIS), as well as two novel genetic algorithm (GA)-based ensemble data-driven models: boosting and random subspace. The objective was to map susceptibility to forest fires in the Northern Mazar District in Jordan. The ensemble models were constructed using four well-known classifiers: decision tree (DT), support vector machine (SVM), k-nearest neighbors (kNN), and naive Bayes (NB) algorithms. This study utilized seventy forest fire locations and twelve influential factors to build and evaluate the models. To identify the optimal features for constructing the data-driven models, a GA-based wrapper method and four machine learning models were applied. During the validation phase, the area under the receiver operating characteristic curve (AUROCC) values for the single SVM, single NB, single DT, single kNN, GA-based boosting, GA-based random subspace, FR-AHP, and AHP-TOPSIS models were found to be 85.3%, 85.9%, 73.8%, 88.7%, 95.0%, 95.0%, 74.0%, and 65.4% respectively. The results indicated that the GA-based ensemble models outperformed both the single machine learning models and the knowledge-based techniques in terms of performance. The developed models in this study can be effectively utilized in various management and decision-making processes aimed at mitigating forest fire risks and enhancing fire control strategies.

Predictive modeling of dose-volume parameters of carcinoma tongue cases using machine learning models

The aim of this study is to create a single institution-based machine learning model for a dose prediction generation tool for post-operative carcinoma of the tongue cases prospectively. Intensity-modulated radiotherapy (IMRT) plans for 20 patients with carcinoma of the tongue were generated using the Eclipse treatment planning system. A machine learning model was generated using a Python 3.10 computer language in a Jupyter notebook using Anaconda software. The PTVs and OARs doses obtained from the clinical treatment plans were used as a primary dataset. Machine learning models are built with two different datasets (10 and 20) for each selected volume. Volumes from 10 new sets of patients were fed into the software for predicting the corresponding dose values. Through the input given, the plan generated dose values of 10 patients were compared with the predicted outcomes of the 10 and 20 dataset models. The model created using the PTVs volume data predicted the dose values with increased accuracy. By verifying the model prediction with the TPS generated value, both the 10 and 20 dataset models predict all the 10 PTVs data within an error bound of 3% and most of the OARs data within an error bound of 5%. The dosimetric features implemented in the machine learning models reasonably predict both the PTVs dose parameter and OARs constraints and give confidence in decision-making during the clinical planning process.

Carbon price prediction based on decomposition technique and extreme gradient boosting optimized by the grey wolf optimizer algorithm

It is essential to predict carbon prices precisely in order to reduce CO2 emissions and mitigate global warming. As a solution to the limitations of a single machine learning model that has insufficient forecasting capability in the carbon price prediction problem, a carbon price prediction model (GWO–XGBOOST–CEEMDAN) based on the combination of grey wolf optimizer (GWO), extreme gradient boosting (XGBOOST), and complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is put forward in this paper. First, a random forest (RF) method is employed to screen the primary carbon price indicators and determine the main influencing factors. Second, the GWO–XGBOOST model is established, and the GWO algorithm is utilized to optimize the XGBOOST model parameters. Finally, the residual series of the GWO–XGBOOST model are decomposed and corrected using the CEEMDAN method to produce the GWO–XGBOOST–CEEMDAN model. Three carbon emission trading markets, Guangdong, Hubei, and Fujian, were experimentally predicted to verify the model’s validity. Based on the experimental results, it has been demonstrated that the proposed hybrid model has enhanced prediction precision compared to the comparison model, providing an effective experimental method for the prediction of future carbon prices.

Combining physical-based model and machine learning to forecast chlorophyll-a concentration in freshwater lakes

Increasing algal blooms in freshwater lakes have become a serious challenge facing the world. Short-term forecast of chlorophyll-a concentration (Chla) is essential for providing early warnings and taking action to mitigate the risks of algal blooms in freshwater lakes. At present, a variety of data-driven models and physical-based models have been developed for Chla forecast, yet how to effectively combine multiple models for improving the forecast accuracy remains largely unknown. Here we developed an effective model by combining a physical-based model and machine learning algorithms (long short-term memory, LSTM; random forest, RF; support vector machine, SVM) to forecast the Chla in a freshwater lake, and a Bayesian model averaging (BMA) ensemble forecasting method was further proposed to improve the accuracy and reliability of the forecast results. We found that, with the increase of time steps of advance forecast from 1-day to 7-day, the forecast accuracy as measured by R2 of the machine learning algorithms is decreased from 0.95 to 0.68. The combination of physical-based modeling with LSTM had great capability in short-term forecast of Chla, owing to the fact that the physical-based model can provide high-frequency Chla data and LSTM is skilled at forecasting in the sequence. This is also evidenced by the weights in the BMA method. The proposed BMA short-term ensemble forecasting results had the robust performance when compared to each individual machine learning forecast model for the 7-day advance forecast, with the largest R2 (0.834) and the smallest RMSE (0.267 μg/L). In particular, the uncertainty of a single machine learning model can be effectively reduced by the BMA method.

Performance analysis of machine learning algorithms: Single Model VS Ensemble Model

Machine Learning is a branch of Artificial Intelligence that predicts several naturally occurring events by training a model with some data and then using unseen data to test it. This paper seeks to analyze the performances of single and ensemble machine learning algorithms on the Cleveland Heart disease data set. Experimental study proves that the accuracy score and area under the ROC curve in the ensemble machine learning model is higher than the single machine learning model in predicting non-CVD and CVD patients.

Transfer Learning for P300 Brain-Computer Interfaces by Joint Alignment of Feature Vectors.

This article presents a new transfer learning method named group learning, that jointly aligns multiple domains (many-to-many) and an extension named fast alignment that aligns any further domain to previously aligned group of domains (many-to-one). The proposed group alignment algorithm (GALIA) is evaluated on brain-computer interface (BCI) data and optimal hyper-parameter values of the algorithm are studied for classification performance and computational cost. Six publicly available P300 databases comprising 333 sessions from 177 subjects are used. As compared to the conventional subject-specific train/test pipeline, both group learning and fast alignment significantly improve the classification accuracy except for the database with clinical subjects (average improvement: 2.12±1.88%). GALIA utilizes cyclic approximate joint diagonalization (AJD) to find a set of linear transformations, one for each domain, jointly aligning the feature vectors of all domains. Group learning achieves a many-to-many transfer learning without compromising the classification performance on non-clinical BCI data. Fast alignment further extends the group learning for any unseen domains, allowing a many-to-one transfer learning with the same properties. The former method creates a single machine learning model using data from previous subjects and/or sessions, whereas the latter exploits the trained model for an unseen domain requiring no further training of the classifier.