Combinations Of Input Features Research Articles

Ensembling methods for protein-ligand binding affinity prediction

Protein-ligand binding affinity prediction is a key element of computer-aided drug discovery. Most of the existing deep learning methods for protein-ligand binding affinity prediction utilize single models and suffer from low accuracy and generalization capability. In this paper, we train 13 deep learning models from combinations of 5 input features. Then, we explore all possible ensembles of the trained models to find the best ensembles. Our deep learning models use cross-attention and self-attention layers to extract short and long-range interactions. Our method is named Ensemble Binding Affinity (EBA). EBA extracts information from various models using different combinations of input features, such as simple 1D sequential and structural features of the protein-ligand complexes rather than 3D complex features. EBA is implemented to accurately predict the binding affinity of a protein-ligand complex. One of our ensembles achieves the highest Pearson correlation coefficient (R) value of 0.914 and the lowest root mean square error (RMSE) value of 0.957 on the well-known benchmark test set CASF2016. Our ensembles show significant improvements of more than 15% in R-value and 19% in RMSE on both well-known benchmark CSAR-HiQ test sets over the second-best predictor named CAPLA. Furthermore, the superior performance of the ensembles across all metrics compared to existing state-of-the-art protein-ligand binding affinity prediction methods on all five benchmark test datasets demonstrates the effectiveness and robustness of our approach. Therefore, our approach to improving binding affinity prediction between proteins and ligands can contribute to improving the success rate of potential drugs and accelerate the drug development process.

Integrating solar-induced chlorophyll fluorescence with traditional remote sensing and environmental variables for enhanced rice yield prediction in Nepal using machine learning

Rice is a vital cereal crop providing essential food for Nepal. To mitigate climate risks and ensure food security, especially with climate change and increasing extreme weather events, it is imperative and a prerequisite to have timely and reliable rice yield prediction. While traditional vegetation indices (VIs) and environmental variables are commonly used in rice yield prediction, the potential of solar-induced chlorophyll fluorescence (SIF), despite its greater sensitivity, has been explored in limited studies. Furthermore, integrating SIF with other remote sensing and environmental data for regional crop yield modeling still needs to be explored. This study aims to address these gaps by incorporating SIF to enhance rice yield prediction in Nepal, offering a novel approach to improving model accuracy. This study uses multi-source data and machine learning algorithms to study 22 major rice-growing districts in the Terai region of Nepal. Random forest (RF), support vector machine (SVR), gradient boosting regressor (GBR), and least absolute shrinkage and selection operator (LASSO) algorithms were employed to predict crop yield using various combinations of input features. Machine learning models were trained and tested over four different time windows within the growing period of rice, using 15 years of rice yield data from 2003 to 2017. Our analysis reveals that GBR out-performed other machine learning algorithms when using input features from July to September, achieving root mean square error (RMSE) and the ratio of RMSE (RRMSE) to observed mean values of 250 kg/ha and 8.6%, respectively. The findings also reveal that incorporating SIF alongside traditional remote sensing and environmental variables significantly enhances the model's prediction accuracy. These findings are valuable for ensuring food security, optimizing agricultural resource management, and supporting decision-making for farmers, policymakers, and supply chain stakeholders.

Fusion of spectral and topographic features for land use mapping using a machine learning framework for a regional scale application.

This study investigated the dynamics of land use and land cover (LULC) modelling, mapping, and assessment in the Kegalle District ofSri Lanka, where policy decision-making is crucial in agricultural development where LULC temporal datasets are not readily available.Employing remotely sensed datasets and machine learning algorithms, the work presented here aims to compare the accuracy of three classification approaches in mapping LULC categories across the time in the study area primarily using the Google Earth Engine (GEE). Three classifiers namely random forest (RF), support vector machines (SVM), and classification and regression trees (CART) were used in LULC modelling, mapping, and change analysis. Different combinations of input features were investigated to improve classification performance. Developed models were optimised using the grid search cross-validation (CV) hyperparameter optimisation approach. It was revealed that the RF classifier constantly outstrips SVM and CART in terms of accuracy measures, highlighting its reliability in classifying the LULC. Land cover changes were examined for two periods, from 2001 to 2013 and 2013 to 2022, implying major alterations such as the conversion of rubber and coconut areas to built-up areas and barren lands. For suitable classification with higher accuracy, the study suggests utilising high spatial resolution satellite data, advanced feature selection approaches, and a combination of several spatial and spatial-temporal data sources. The study demonstrated practical applications of derived temporal LULC datasets for land management practices in agricultural development activities in developing nations.

Explainable ensemble learning graphical user interface for predicting rebar bond strength and failure mode in recycled coarse aggregate concrete

Novel study deploys robust machine learning algorithms using newly built comprehensive dataset to predict reinforcing rebar-to-recycled coarse aggregate concrete (RCA) bond strength and failure mode. Prior investigations have solely concentrated on bond strength, resulting in a limited comprehension of the bond failure pattern. Considering the increasing significance of sustainable construction methods, it is crucial to examine both the failure pattern and bond strength to expand the versatility of RCA in various reinforced concrete structures. Accordingly, XGBoost, CatBoost, Random Forest, and LightGBM were trained for this purpose. Model performance was appraised using various statistical metrics, while failure classification performance was assessed using accuracy, recall, and precision indicators. Model performance was ranked using Copeland's algorithm. Feature importance was quantified using SHAP. Coefficient of determination of 0.91 was achieved by XGBoost in predicting bond strength, outperforming other nine analytical models in literature. Failure mode was predicted with accuracy of 94% by CatBoost, XGBoost, and LightGBM. Embedment length and compressive strength features had greatest influence on bond strength and failure mode, respectively. User-friendly graphical interface was developed to harvest ML models in real-world engineering practice. Online free access accurately assigns to any given combination of input features corresponding accurate rebar bond strength and failure mode.

Improving the experience of machine learning in compressive strength prediction of industrial concrete considering mixing proportions, engineered ratios and atmospheric features

In previous literature on predicting compressive strength (CS) using machine learning (ML), the focus has primarily been on algorithm-specific improvements, with less emphasis on improving the data perspective of CS prediction. Departing from these conventional investigations, this study presents data-centric enhancements. A systematic and comprehensive methodology is adopted, involving series of steps. Firstly, a large dataset (> 24,214 datapoints) comprising 26 input features relating to mixture proportions, engineered ratios and atmospheric parameters is prepared and processed. Feature engineering is then performed for selection and processing of data features. Subsequently, six scenarios are designed and investigated for CS prediction, focusing on distinct combinations of input features. Finally, CS predictions for all six scenarios are compared using five well-known ML model, and results are validated using statistical hypothesis testing. The results show significance improvements in CS prediction, with a 12.8 % increase in coefficient of determination (R2) and a 61 % reduction in root mean square error (RMSE) achieved by incorporating engineered and atmospheric parameters in conjunction with mixture proportion as inputs in CS prediction. Based on CS prediction in six scenarios, the study presents useful discussions focusing on concrete data perspective, role of engineering ratio and atmospheric features, and interaction and influence of input features in CS prediction. In conclusion, this work emphasizes the importance of enhancing data management practices for concrete performance assessment to have a more efficient, realistic, and sustainable outcomes in the concrete industry.

Implementation of deep learning models in predicting ESG index volatility

The consideration of environmental, social, and governance (ESG) aspects has become an integral part of investment decisions for individual and institutional investors. Most recently, corporate leaders recognized the core value of the ESG framework in fulfilling their environmental and social responsibility efforts. While stock market prediction is a complex and challenging task, several factors associated with developing an ESG framework further increase the complexity and volatility of ESG portfolios compared with broad market indices. To address this challenge, we propose an integrated computational framework to implement deep learning model architectures, specifically long short-term memory (LSTM), gated recurrent unit, and convolutional neural network, to predict the volatility of the ESG index in an identical environment. A comprehensive analysis was performed to identify a balanced combination of input features from fundamental data, technical indicators, and macroeconomic factors to delineate the cone of uncertainty in market volatility prediction. The performance of the constructed models was evaluated using standard assessment metrics. Rigorous hyperparameter tuning and model-selection strategies were implemented to identify the best model. Furthermore, a series of statistical analyses was conducted to validate the robustness and reliability of the model. Experimental results showed that a single-layer LSTM model with a relatively small number of neurons provides a superior fit with high prediction accuracy relative to more complex models.

Quantitative detection of moisture content of corn by olfactory visualization technology

In this study, a novel method for quantitatively detecting the moisture content of corn by olfactory visualization technology is proposed. First of all, an olfactory visualization system was developed to obtain odor information of corn samples with varying moisture contents using sensor arrays. The sensor array was constructed with twelve chemical dyes, which helped assemble the system. Then, the RreliefF algorithm was utilized to sort the characteristic variables of the sensor image by weight, and the analytical model based on different combinations of input features was built using support vector regression (SVR). Moreover, the ability of three optimization algorithms, namely grid search (GS), particle swarm optimization (PSO) and dung beetle optimizer (DBO), was compared to determine their effectiveness in optimizing the parameters of the SVR model. The results reveal that the RreliefF-GS-SVR model demonstrates an enhanced correlation coefficient of prediction (RP), with an increase from 0.97 to 0.98, as well as the root mean square error of prediction (RMSEP), is reduced from 0.66% to 0.53% when compared to the GS-SVR model. Similarly, in comparison to the RreliefF-GS-SVR model and RreliefF-PSO-SVR model, the RreliefF-DBO-SVR model shows better predictive performance, with an increase in RP to 0.99 and a decrease in RMSEP to 0.25%. The comprehensive findings indicate that integrating olfactory visualization technology and multivariate analysis is feasible for conducting quantitative analysis of grain moisture content.

Extraction of eutrophic and green ponds from segmentation of high-resolution imagery based on the EAF-Unet algorithm

Inland ponds exhibit remarkable ubiquity across the globe, playing a vital role in the sustainability of global continental freshwater resources and contributing significantly to their biodiversity. Numerous ponds are eutrophic and experience recurrent seasonal or year-round algal blooms or persistent duckweed cover, conferring a characteristic green hue. Here, we denote these eutrophic and green ponds as EGPs. The excessive proliferation of algal blooms and duckweed within these EGPs poses a significant threat to the ecological functioning of these aquatic systems, which can lead to hypoxia or the release of microcystins. To identify these EGPs automatically, we constructed an Efficient Attention Fusion Unet (EAF-Unet) algorithm using Gaofen-2 (GF2) panchromatic and multispectral imagery. The attention mechanism was incorporated in Unet to help better detect EGPs. Using the first EGP labeled dataset, we determined the best input feature combination (RGB, NIR, NDVI, and Bright) and the most effective encoding (Rasnet50) for EAF-Unet for distinguishing EGPs from other ground cover types. The evaluation indices – Precision (0.81), Recall (0.79), F1-Score (0.80), and Intersection over Union (IoU, 0.67) – indicate that EAF-Unet can accurately and robustly extract EGPs from GF2 images without relying on pond water masks. Remote-sensing EGP products can assist in identifying ponds with severe eutrophication. Moreover, these products can serve as references for identifying high-risk areas prone to improper sewage discharge or inadequate sewer construction.

Neural network energy management strategy with optimal input features for plug-in hybrid electric vehicles

The neural network energy management strategy can be implemented online and effectively save the energy for the plug-in hybrid electric vehicles (PHEVs). However, the selection of input features and application to untrained driving cycles are two issues faced by the strategy. This paper proposes an energy management strategy with the optimal input features for the PHEV. The global optimal datasets are first obtained based on the dynamic programming (DP) algorithm. Then the random forest classification models are trained with different combinations of input features to select the input feature combination with the highest classification accuracy. A neural network with the selected input features is finally trained using the optimal datasets for online control. With the optimal input features, the strategy can be adopted to both trained driving cycles and untrained driving cycles. Results demonstrate that the proposed strategy can save 4.44% to 7.75% energy compared to the charge depleting and charge sustaining strategy on trained cycles and 3.80% to 7.70% on untrained cycles respectively. The efficiency of the proposed strategy is only less than 2.41% worse than the DP algorithm.

Lidar-based virtual load sensors for mooring lines using artificial neural networks

Floating offshore wind turbines are equipped with a variety of sensors, which are measuring data, valuable for the control and monitoring of the turbine. However, reliable measurements are difficult or costly for some physical quantities. This includes load estimates for mooring lines and fairleads. In this study, we investigate an approach using wind speed measurements from a forward-looking nacelle-based lidar as inputs to long short-term memory networks to estimate fairlead tensions. Nacelle-based lidar wind speed measurements on floating offshore wind turbines are influenced by platform motions, in particular by the rotational pitch displacement and the surge displacement of the floater. Therefore, the lidar wind speed measurement contains information about the dynamic behavior of the floater. In turn, the floater’s dynamics determine the fair lead loads. Thus in this study, we directly use the lidar-measured line of sight (LOS) wind speeds to estimate mooring line tensions. The model training data is obtained using the aero-elastic wind turbine simulation tool openFAST in combination with the numerical lidar simulation framework ViConDAR. Results show, that lidar-based virtual load sensors can reproduce mean fairlead tension as well as low-frequency fluctuations, with varying accuracy dependant on the combination of input features. For the model which is only using LOS wind speed measurements as input a normalized root mean squared error of 0.55 was obtained.

Deep-learning-enhanced digital twinning of complex composite structures and real-time mechanical interaction

Digital twins are undergoing growth that enables highly informative and scaleable interaction between physical objects and virtual twins, which is of great significance to the life cycle analysis of composites. Real-time fine evolution is challenging due to vast combinations of input features and high-resolution calculated variables. Here, we systematically demonstrate an AI-based methodology for digital twinning of complex composite structures. First, three types of deep neural networks are created with optionally used autoencoders as surrogate models, with architectures and data processing inspired by the rule-of-mixture of composites. Second, the prediction accuracy and efficiency are evaluated quantitatively and qualitatively, demonstrating the feasibility of predicting 3D displacement and stress fields directly from sensing data of temperature, pressure and loading displacement, and the optimal architecture is selected to be the evolving digital twin. Finally, the real-time interactive experiments are relayed to the digital twin and demonstrated that it can interact with physical objects and evolve online with high accuracy. These results indicate that the computational time can be reduced by 3∼6 orders of magnitude with high information intensity and scalability compared with conventional numerical and experimental methods, which opens up the avenues for the cost-effective and efficient development of digital twin services for composites.

Nonlinear input feature reduction for data-based physical modeling

This work introduces a novel methodology to derive physical scalings for input features from data. The approach developed in this article relies on the maximization of mutual information to derive optimal nonlinear combinations of input features. These combinations are both adapted to physics-related models and interpretable (in a symbolic way). The algorithm is presented in detail, then tested on a synthetic toy model. The results show that our approach can effectively construct relevant combinations by analyzing a strongly noisy nonlinear dataset. These results are promising and may significantly help training data-driven models. Finally, the last part of the paper introduces a way to account for the physical dimension of data. The test case is a synthetic dataset inspired by the Law of the Wall from turbulent boundary layer theory. Once again, the algorithm shows that it can recover relevant nondimensional variables for data-base modeling.

Using NDVI, climate data and machine learning to estimate yield in the Douro wine region

Estimating vineyard yield in advance is essential for planning and regulatory purposes at the regional level, with growing importance in a long-term scenario of perceived climate change. With few tools available, the current study aimed to develop a yield estimation model based on remote sensing and climate data with a machine-learning approach. Using a satellite-based time-series of Normalized Difference Vegetation Index (NDVI) calculated from Sentinel 2 images and climate data acquired by local automatic weather stations, a system for yield prediction based on a Long Short-Term Memory (LSTM) neural network was implemented. The study was conducted in the Douro Demarcated Region in Portugal over the period 2016–2021 using yield data from 169 administrative areas that cover 250,000 ha, in which 43,000 ha of the vineyard are in production. The optimal combination of input features, with an Mean Absolute Error (MAE) of 672.55 kg/ha and an Mean Squared Error (MSE) of 81.30 kg/ha, included the NDVI, Temperature, Relative Humidity, Precipitation, and Wind Intensity. The model was tested for each year, using it as the test set, while all other years were used as input to train the model. Two different moments in time, corresponding to FLO (flowering) and VER (veraison), were considered to estimate in advance wine grape yield. The best prediction was made for 2020 at VER, with the model overestimating the yield per hectare by 8 %, with the average absolute error for the entire period being 17 %. The results show that with this approach, it is possible to estimate wine grape yield accurately in advance at different scales.

Prediction of significant wave height in hurricane area of the Atlantic Ocean using the Bi-LSTM with attention model

The Bi-LSTM with attention (BLA) model is proposed to predict wave height in the hurricane area of the Atlantic Ocean. Four data features (wave height, wind speed, wind direction, and wave direction) collected at five buoy stations are selected as model inputs, while future wave height is specified as model output. Predictions are obtained for 1-h, 3-h, 6-h, and 12-h lead times. Two evaluation metrics are introduced to evaluate the accuracy and stability of the model. Influencing factors such as input–output ratio and input feature combination are studied. For the input–output ratio, it is found that the shorter lead time requires a larger ratio, with a ratio of 9:1 required for 1-h lead time prediction, with 3:1 for 3-h prediction, 2:1 for 6-h prediction, and 1:1 for 12-h prediction. For the input feature combination, the combination of the wave height, the wind speed, and the coupling of the wind direction and wave direction via the cosine function holds the best prediction performance in 1-h and 3-h lead time prediction. And for 6-h and 12-h forecasts, a combination of wave height and wind speed is best. By comparing overall evaluation metrics and extreme wave height prediction results of the BLA model with those of the Bi-LSTM, LSTM, and LSTM with attention models, we notice that the BLA model has the best and most stable prediction performance.

What Does Your Bio Say? Inferring Twitter Users’ Depression Status From Multimodal Profile Information Using Deep Learning

People suffering from stress and various mental health problems find it easier to express and share their feelings on online platforms, such as Twitter. However, the imposed character limit (280 characters) by Twitter and infrequent online activities of a section of users poses a serious setback in using computational methods for mental health analysis or emotion research. Twitter provides rich metadata information about its users (such as user’s description, geolocation, and profile image URL), which can provide valuable information regarding the mental state of the users. We hypothesize that Twitter’s rich metadata information about their users can provide some valuable depression cues, which may help in an early low-profile evaluation. In this article, we investigate this hypothesis by developing an end-to-end multimodal multitask (MT) system for depression detection (primary task) and emotion recognition (auxiliary task), where the variation of emotion information based on different user descriptions assists the learning of the primary task. The proposed system attains 70% accuracy on the depression detection task outperforming several single-task (ST) baselines built on the various combination of input features. Our findings indicate that Twitters’s rich metadata information can be leveraged to detect depression among users with significant confidence.

Design optimization of a multi-layer porous wave absorber using an artificial neural network model

The design optimization of a multi-layer porous wave absorber has been achieved through the supervised artificial neural network (ANN) and analytical model validated by CFD and experiments. The analytical model is established by means of a matched eigenfunction expansion method (MEEM) by applying the boundary condition with a quadratic relationship between the pressure drop and traversing fluid velocity at the porous plates. Along with the reflection coefficient as a target feature, input key features of the multi-layer porous wave absorber are selected through the parametric study. A dataset with 200 combinations of input design features is generated by the developed analytical model. Using the dataset, the ANN model is trained with a R2 score of 0.97. Predictions are generated for a large sample set with the trained ANN model. Top 1% combinations of input features that yield a minimum value of the reflection coefficient are used for the optimal design of the wave absorber. Based on the interquartile range (IQR) value of this 1% data, it is found that the most important design features are the porosity and submergence depth of the upmost plate and their optimal ranges for a double-layer wave absorber are 0.055≤d1/h≤0.067 and 0.117≤P1≤0.173. The optimal range of design features can be used as a guideline for the design of an effective wave absorber.

Prediction of growth in grower-finisher pigs using recurrent neural networks

Accurate prediction of pig growth is an important tool for supporting sustainable global pork production, due to the potential benefits to farm management in monitoring growth according to target, promoting efficient use of resources and planning sale of finished animals. In this paper, a long short-term memory (LSTM) recurrent neural network architecture was designed to predict how long it would take for grower-finisher pigs housed in a climate-controlled building reflecting commercial practice to reach a target liveweight of 90 kg. To make these predictions, data on pig liveweight and food consumption were integrated, along with various combinations of environmental temperature recorded by the building control system and, to represent seasonality, encodings of day of the year. Also, ensembles were created and evaluated by averaging the predictions of multiple models, each trained with different combinations of input features, and comparisons were made between the performance of the LSTM models and other machine learning algorithms. Overall, the LSTM model achieved the best prediction performance, with a RMSE of 2.470 days and a Pearson's correlation coefficient of 0.962. Integrating both environmental temperature and seasonality data into the models generally resulted in an improvement in predictive performance. The results demonstrate the potential of LSTM models as a precision livestock farming tool for the early prediction of growth in commercial finisher pigs. • A machine learning pipeline for the forecasting of pig growth was designed. • Predicting the time to reach a target weight can improve data availability. • Specialised neural network architectures are very effective at predicting growth. • Utilising environmental temperatures can improve growth prediction accuracy. • Specialised algorithms can be used to remove outliers from non-cyclical time series.

RECHOmmend: An ECG-Based Machine Learning Approach for Identifying Patients at Increased Risk of Undiagnosed Structural Heart Disease Detectable by Echocardiography.

Background:Timely diagnosis of structural heart disease improves patient outcomes, yet many remain underdiagnosed. While population screening with echocardiography is impractical, ECG-based prediction models can help target high-risk patients. We developed a novel ECG-based machine learning approach to predict multiple structural heart conditions, hypothesizing that a composite model would yield higher prevalence and positive predictive values to facilitate meaningful recommendations for echocardiography.Methods:Using 2 232 130 ECGs linked to electronic health records and echocardiography reports from 484 765 adults between 1984 to 2021, we trained machine learning models to predict the presence or absence of any of 7 echocardiography-confirmed diseases within 1 year. This composite label included the following: moderate or severe valvular disease (aortic/mitral stenosis or regurgitation, tricuspid regurgitation), reduced ejection fraction <50%, or interventricular septal thickness >15 mm. We tested various combinations of input features (demographics, laboratory values, structured ECG data, ECG traces) and evaluated model performance using 5-fold cross-validation, multisite validation trained on 1 site and tested on 10 independent sites, and simulated retrospective deployment trained on pre-2010 data and deployed in 2010.Results:Our composite rECHOmmend model used age, sex, and ECG traces and had a 0.91 area under the receiver operating characteristic curve and a 42% positive predictive value at 90% sensitivity, with a composite label prevalence of 17.9%. Individual disease models had area under the receiver operating characteristic curves from 0.86 to 0.93 and lower positive predictive values from 1% to 31%. Area under the receiver operating characteristic curves for models using different input features ranged from 0.80 to 0.93, increasing with additional features. Multisite validation showed similar results to cross-validation, with an aggregate area under the receiver operating characteristic curve of 0.91 across our independent test set of 10 clinical sites after training on a separate site. Our simulated retrospective deployment showed that for ECGs acquired in patients without preexisting structural heart disease in the year 2010, 11% were classified as high risk and 41% (4.5% of total patients) developed true echocardiography-confirmed disease within 1 year.Conclusions:An ECG-based machine learning model using a composite end point can identify a high-risk population for having undiagnosed, clinically significant structural heart disease while outperforming single-disease models and improving practical utility with higher positive predictive values. This approach can facilitate targeted screening with echocardiography to improve underdiagnosis of structural heart disease.

COVID-19 Epidemic Analysis in India with Multi-Source State-Level Datasets.

The COVID-19 pandemic has been a global crisis affecting billions of people and causing countless economic losses. Different approaches have been proposed for combating this crisis, including both medical measures and technical innovations, e.g., artificial intelligence technologies to diagnose and predict COVID-19 cases. While there is much attention being paid to the USA and China, little research attention has been drawn to less developed countries, e.g., India. In this study, I conduct an analysis of the COVID-19 epidemic in India, with datasets collected from different sources. Several machine learning models have been built to predict the COVID-19 spread, with different combinations of input features, in which the Transformer is proven as the most precise one. I also find that the Facebook mobility dataset is the most useful for predicting the number of confirmed cases. However, I find that the datasets from different sources are not very effective when predicting the number of deaths caused by the COVID-19 infection.

An adaptive feature selection schema using improved technical indicators for predicting stock price movements

Accurate stock market forecasts can bring high returns for investors. There have been a growing number of studies employing machine learning technology to perform stock prediction tasks with the development of machine learning and artificial intelligence technologies. However, accurately predicting stock price trends still is an elusive goal, not only because the stock market is affected by policies, market environment, market sentiment, etc., but also because stock price data is inherently complex, noisy, and nonlinear. Many technical indicators have been used as input features to stock prediction models, but the quality of technical indicators has always been a neglected issue, thus the application of feature engineering in stock prediction tasks needs to be further expanded. Using 18 technical indicators as the original features, this paper presents improved technical indicators based on wavelet denoising and a novel two-stage adaptive feature selection method. Finally, the random forest model is used as the stock prediction model. Experiments show that in contrast to the original technical indicators, the improved technical indicators significantly enhance the performance of the model (e.g., F1 scores increased by 34.48% on the SSE Composite Index (SSEC) data set, 41.56% on the Hang Seng Index (HSI) data set, 34.48% on the Dow Jones Industrial Average (DJI) data set, 32.75% on the Standard & Poor's 500 Index (S&P 500) data set). The experimental results verify the importance of the quality of technical indicators in the task of stock prediction. Meanwhile, the results also demonstrate the effectiveness of the feature selection method, which can achieve higher prediction accuracy with fewer features. In addition, we established multiple data sets according to the size-varied time windows to study the influence of the size-varied time windows. The results show that properly increasing the size of the time window can exert a positive impact on the model. Finally, by utilizing our two-stage adaptive feature selection method, we remove redundant features, and achieve excellent results on data sets from four different stock markets (e.g., F1 scores reached 0.754 on the SSEC data set, 0.794 on the HSI data set, 0.789 on the DJI data set, 0.821 on the S&P 500 data set). Overall, this study experimentally verifies that improving feature quality can positively impact model performance, and that choosing an appropriate combination of input features can not only improve model performance, but reduces the negative impact of the curse of dimensionality as well.