Page 2389
www.rsisinternational.org
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025
A Comparative Review of Attribute Selection Techniques for PM2.5
Prediction Using Machine Learning Models
Dr. Sachin Arun Thanekar
Associate Professor, Computer Science and Engineering Department, MITADT University, Loni
Kalbhor, Pune, India
DOI: https://dx.doi.org/10.51244/IJRSI.2025.1210000212
Received: 07 October 2025; Accepted: 14 October 2025; Published: 15 November 2025
ABSTRACT
Accurate prediction of fine particulate matter (PM2.5) is vital for understanding and mitigating air pollution
impacts on public health. With the rise of machine learning (ML) in environmental forecasting, selecting the
most influential features remains a critical preprocessing step. This review paper evaluates the effectiveness of
various attribute selection techniques applied to PM2.5 prediction, including filter, wrapper, and embedded
methods. We compare the results from Random Forest, LASSO regression, Recursive Feature Elimination
(RFE), and correlation analysis. Our comparative analysis reveals that Random Forest consistently highlights
meteorological variables such as temperature and wind speed as top contributors, whereas LASSO reduces
model complexity by focusing on core pollutants. The paper provides insights for researchers aiming to develop
robust and computationally efficient models for real-time PM2.5 forecasting.
Keywords: PM2.5, Feature Selection, Random Forest, LASSO, Machine Learning, Air Quality Prediction,
Attribute Selection
INTRODUCTION
A large set of features can be considered for use by machine learning models for PM2.5 prediction. These
typically include meteorological parameters (temperature, relative humidity, wind speed, atmospheric pressure),
concentration of other pollutants (NO₂, SO₂, CO, O₃), time-related variables (season, month, hour), location, and
perhaps socio-economic indicators. However, using all available features does not guarantee better model
performance. Irrelevant or redundant features may only add noise, increase the time cost due to overfitting, or
even require additional computing resources. That is where attribute (feature) selection approaches become
relevant. These methods try to select a smaller subset of features that are considered most informative and
relevant, which improves interpretability, reduces dimensionality, and increases the prediction performance of
the models. Proper feature selection not only streamlines the modeling process but also leads to the identification
of drivers responsible for PM2.5 changes, which may be helpful information from a scientific and policy
perspective. Chen, J., Li, S., Huang, G., & Liu, X. (2023), Wang, Y., Liu, Y., Wu, Y., & Zhang, L. (2022).
REVIEW METHODOLOGY
This review followed a structured literature review approach inspired by the PRISMA guidelines for systematic
reviews, adapted to the scope of a comparative methodological review. The steps were as follows:
Database Selection: We searched major academic databases, including IEEE Xplore, ScienceDirect,
SpringerLink, and Google Scholar.
Search Strategy: Keywords such as “PM2.5 prediction,” “feature selection,” “attribute selection,” “machine
learning,” and “air quality forecasting” were used in various Boolean combinations.
Inclusion Criteria: Studies were included if they (a) focused on PM2.5 or air quality prediction, (b) applied or
compared feature/attribute selection techniques, and (c) reported quantitative performance metrics.
Page 2390
www.rsisinternational.org
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025
Exclusion Criteria: Papers were excluded if they (a) did not involve machine learning methods, (b) were purely
theoretical without experiments, or (c) duplicated findings already covered in earlier reviews.
Screening Process: An initial set of papers was identified. After title/abstract screening and full-text evaluation,
some papers were retained for final analysis.
Data Extraction and Analysis: For each selected study, details such as feature selection technique, dataset used,
machine learning models applied, and performance metrics (R², RMSE, MAE, etc.) were extracted and
compared.
Importance of Attribute Selection
Feature selection serves as an important preprocessing procedure of a machine learning pipeline that endeavors
to identify and keep only the variables considered most useful and informative for the betterment of predictive
models. The purpose of feature selection is to cast away irrelevant, redundant, or noisy attributes that do not
contribute meaningfully to the target variable, here PM2.5 concentration levels, thus improving model efficiency,
interpretability, and predictive accuracy. Sharma, A., & Saini, L. M. (2021).
Potential input variables considered for PM2.5 forecasting are indeed rather broad and often very high
dimensional. Common inputs may include concentrations of pollutants such as NO₂, SO₂, CO, O₃, and PM10,
meteorological parameters such as temperature, humidity, wind speed, wind direction, rainfall, and atmospheric
pressure, temporal indicators such as time of day, day of the week, season, and month, as well as spatial
information including geographic coordinates, altitude, land use patterns, and proximity to pollution sources.
This heterogeneity of features represents the complexity and multifactorial nature of air pollution that gets
influenced by both anthropogenic and natural causes. Ahmed, R. M., & Badr, A. (2021), Zhang, Y., Wang, J., &
Wang, S. (2020).
However, not all these features contribute equally to the PM2.5 levels. Some carry redundant information (for
example: PM10 and PM2.5 tend to be strongly correlated to each other), whereas some would introduce
irrelevant noise or spurious relationships that would hamper the performance of a model. Greater inclusion of
such features can be detrimental as the model learns to fit the noise in the training data, gives lower performance
on new data, and is thus, overfitting in nature. Conversely, optimal feature selection has been known to help
improve the generalizability of models, reduce computational cost, as well as the time necessary for training and
inference-speed-two great issues in real-time air quality monitoring systems.
Considering how multidimensional and often sputter-co-the dimension does air quality data sets, feature
selection could wait not only be beneficial but necessary. Such high-dimensional data could almost serve to mask
the very patterns one would want to detect, complicating analysis and validation strategies. That conundrum
worsens, as different regions and climates might demand different subsets of features for their best performance,
which highlights the necessity for dynamic, context-aware feature selection mechanisms.
According to Guyon and Elisseeff (2003), feature selection is essential to improve learning algorithms, especially
when datasets are such that the number of features exceeds the number of observations-many scenarios typical
of environmental datasets that cover many seasons or multiple locations. They classify feature selection methods
into three broad categories: Filter methods (independent of any learning algorithm), Wrapper methods (which
use a model performance measure to evaluate the quality of a feature subset), and Embedded methods (where
feature selection is performed as part of model construction).
Chandrashekar and Sahin (2014) went further to say that feature selection helps in reducing the dimensionality
of data, thereby simplifying models and improving their interpretability, which is an exigency in policy-related
contexts like the management of air quality. Their studies indicate that through the right selection of features,
domain specialists are capable of better understanding the relationships linking environmental factors and
pollution levels, leading to more informed decision-making.
A sound PM2.5 prediction model will weigh its feature selection with care. Feature selection creates a
compromise between complexity and performance, making sure that the model is workable and prolific. With a
Page 2391
www.rsisinternational.org
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025
growing number of machine learning techniques becoming an integral part of environmental monitoring
systems, the significance of an intelligent and well-justified assessment of feature selection shall multiply,
resulting in more intelligent, faster, and transparent forecasts of air quality.
Commonly Used Attribute Selection Techniques
Filter Methods
Filter methods assess the relevance of features by examining their statistical relationship with the target variable.
Correlation Analysis: Correlation Analysis is a statistical technique that considers the strength and direction of
a linear relationship between two continuous variables. In air quality prediction, the Pearson correlation is mostly
applied in determining how climate variables and pollutants are varied in relation to certain target variables, such
as PM2.5 concentrations. This classification method is very powerful in the initial stage of feature selection
because it eliminates those variables that are irrelevant or weakly correlated and thus simplifies the modeling
process. For instance, one study utilized the Pearson correlation analysis technique to identify meaningful
positive and negative correlations between PM2.5 levels and other variables such as NO₂, temperature, and wind
speed. Consequently, these environmental factors should be deemed important for particulate matter
concentration modeling (Hu et al., 2013).
Chi-square Test: The Chi-square Test is a statistical test for assessing the associations between two categorical
variables by comparing the observed and expected frequency distribution in a contingency table and is thus
useful in feature selection for testing a given feature's independence from the target variable, e.g., to evaluate
whether a certain feature is informative for a classification assignment. However, with most features in air quality
datasets being continuous variables, such as pollutant concentrations and meteorological indices (e.g., PM2.5
concentration, temperature, humidity), the Chi-square Test becomes less applicable. In order to make use of this
methodology for AQI-related studies, all continuous variables in the study would require conversion into
categorical bins, which will pose threats to the retention of useful information and further jeopardize predictive
performance. Thus, this technique has little relevance in the case of continuous AQI data (Hu et al., 2013).
Wrapper Methods
Wrapper techniques evaluate subsets of features based on model performance.
Recursive Feature Elimination (RFE): Recursive Feature Elimination is a wrapper-type feature selection tool
that recursively fits a model and removes the least important features until the desired number of features to
select is finally reached, thereby stripping away irrelevant or redundant inputs. Because these approaches can
easily rank the features by their importance, RFE is very useful for linear modeling (SVM, logistic regression)
and tree-based models (decision trees, random forests). By gradually pruning out the weaker predictors, RFE
reduces overfitting and improves generalization and accuracy of a model by removing noise from the dataset. In
the air quality prediction framework, RFE filters out some meteorological and pollutant variables with higher
influence and performance tied to the predictive model (Guyon et al., 2002).
Sequential Feature Selection: Sequential Feature Selection builds an optimal feature list by incrementally
adding or removing one feature at a time and evaluating the resulting feature subset using a learning algorithm.
Forward selection would start with no features and keep adding whichever feature most improves the model at
each step, whereas backward elimination would proceed in the opposite manner by iteratively throwing away
the least impactful features. It is a model-dependent technique; thus, it discriminates between feature subsets by
way of a given learning algorithm and allows optimization suited to the model's reaction to various combinations.
Given that feature numbers are not too large,Sequential Feature Selection should prove fruitful in recognizing
those variables most relevant to augmenting accuracy, robustness, and interpretability for regression modeling.
Such an exhaustive feature study reduces model complexity such that predictive strength is either maintained or
enhanced.
Page 2392
www.rsisinternational.org
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025
Embedded Methods
Embedded methods integrate feature selection as part of the model training process.
Random Forest Feature Importance: Random Forest Feature Importance is considered an embedded method
since it is executed during the Random Forest training stage for ranking the relevance of input features based on
how they contribute to predicting accuracy. The Random Forest, a tree ensemble model, builds several decision
trees whose outputs are amalgamated to yield a more robust result and guard against overfitting. During training,
the algorithm sums each feature's capacity to reduce impurity (Gini impurity or entropy, for example) across all
trees, assigning greater importance to features that help shrink the prediction error more significantly. For air
quality modeling purposes, Random Forest has been especially efficient in distinguishing between key
meteorological variables like temperature and humidity, which substantially affect PM2.5 concentration levels.
This user-friendly feature importance ranking built into Random Forest makes it a great medium for both feature
selection and environmental data prediction modeling (Liaw & Wiener, 2002).
LASSO Regression: The name LASSO stands for Least Absolute Shrinkage and Selection Operator and is
embedded feature selection with an intervention introducing L1 regularization into the regression model that
penalizes the absolute magnitude of the regression coefficient. Through the regularization, the coefficients are
forced to become zero for variables deemed relatively unimportant or redundant, resulting in a selection of
features during the very training of the model. With high-dimensional data sets and many variables, the output
generated by LASSO can be very helpful, deciding to keep only those predictors which are most relevant, thus
making interpretation easier and alleviating the problem of overfitting. LASSO is used in air quality prediction
to discard less important pollutant variables while preserving the model's performance, hence making it the
appropriate means for developing very sparse but still acceptably accurate predictions (Tibshirani, 1996).
Table 1: Classification Chart: Attribute Selection Techniques
Method
Type
Technique Approach Data Type Suitability Example Usage
Filter
Methods
Pearson
Correlation
Statistical
correlation with
target
Continuous Identifying linear relationships
(e.g., PM2.5 vs. NO₂,
temperature, wind speed)
Chi-square Test Chi-square score
for categorical
relevance
Categorical Feature relevance in
classification (less relevant for
AQI)
Wrapper
Methods
Recursive
Feature
Elimination
(RFE)
Iteratively
removes features
and evaluates
Continuous/Categorical Enhancing model accuracy
using SVM, Random Forest
Sequential
Feature
Selection
Adds/removes
features step-by-
step
Continuous/Categorical Forward/Backward selection
using custom scoring metrics
Embedded
Methods
Random Forest
Feature
Importance
Built-in
importance
scoring during
training
Mostly continuous Ranking meteorological
variables like humidity,
temperature
LASSO
Regression
L1 regularization
penalizes weak
features
Continuous Eliminating redundant
variables while maintaining
model performance
Page 2393
www.rsisinternational.org
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025
Comparative Evaluation
The ability of feature selection methods was assessed in terms of being able to interpret the model, predictive
abilities measured by R2-RMSE, and computational costs. Results indicate:
Random Forest generates increased robustness with regard to feature importance, exploiting the ensemble
learning structure that sets up a plethora of decision trees and amalgamates their outputs to be able to predict
with a higher degree of stability. This process excels in managing the complexities of nonlinear interactions
among features; hence, it would be a top choice for datasets where relationships between variables are seldom
linear, such as the cases of air quality prediction. By looking at the role each feature plays in reducing impurity
on all trees and ranking them on the basis of this, Random Forest can distinguish between meteorological and
pollutant factors that have the greatest effect on the outcome and, thus, assist with selections to improve the
model and interpretability. Its analysis of high-dimensional data and uncovering of hidden patterns makes
Random Forest a distinct contender among other feature selection methods of environmental modeling criteria
(Liaw & Wiener, 2002).
LASSO regression simplified the models by accomplishing retardation of coefficients through its L1
regularization technique that penalizes the real-valued present form of regression coefficients. Actual
penalization forcibly drives the values of regression coefficients of trivial and highly correlated variables to zero,
producing a sparse model that keeps really relevant features. In doing so, it excludes all inputs that may
redundantly explain the variance in a dependant variable, PM2.5 concentration being one of those in surface-air
quality forecasting. This dimensionality reduction technique enhances model interpretability and computational
efficiency and lowers the chances of overfitting, especially for high-dimensional datasets. More importantly,
LASSO does this without notably detracting from the predictive performance; hence the two most suitable
properties for lightweight and robust modeling approaches to be used in real-time or resource-limited
environments (Tibshirani, 1996).
The Recursive Feature Elimination technique increased model predictability moderately, systematically
selecting features that rank least important in an iterative manner, wherein one iteration involves training the
model, evaluating those features for importance, and braking the least supportive contributor for it to possibly
improve the accuracy while decreasing overfitting due to the quality of being highly informative. Its
computational intensity rendered it paradoxical to use for real-time or large-scale applications since it required
a higher number of rounds for training and evaluation to settle on an optimum feature list, whereas RFE proved
to be useful in pruning the final feature sets, therefore increasing the generalization of the models (Guyon et al.,
2002).
Quantitative Synthesis of Comparative Studies
To substantiate the comparative analysis, we compiled results from representative prior studies that applied
feature selection techniques for PM2.5 prediction. The synthesis focuses on (a) datasets used, (b) feature
selection methods applied, (c) machine learning models tested, and (d) performance metrics such as R², RMSE,
and MAE.
Table 2: Datasets and Feature Selection Techniques in Prior PM2.5 Prediction Studies
Study Dataset
(Region/Size)
Feature Selection
Method
Model(s) Used Key Features Selected
Chen et al. (2023) China, multi-city
air quality dataset
Hybrid FS
(Correlation +
RFE)
Random Forest,
XGBoost
Temperature, humidity,
NO₂, SO₂
Wang et al. (2022) Beijing AQI (5
years, hourly)
Random Forest
Importance
RF, LSTM Meteorological +
pollutant mix
Page 2394
www.rsisinternational.org
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025
Sharma & Saini
(2021)
Delhi AQI (3
years)
LASSO
Regression
SVR, ANN NO₂, CO, humidity
Ahmed & Badr
(2021)
Cairo (hourly, 2
years)
Wrapper
(Sequential FS)
RF, SVM PM10, wind speed, O₃
Zhang et al.
(2020)
China (multi-city,
daily)
Hybrid Feature
Selection (SSA +
FS)
BiGRU, RF Pollutant concentrations
+ temperature
Table 3: Performance Comparison of Feature Selection Techniques
Study Method Model R² RMSE MAE Key Findings
Chen et al.
(2023)
Correlation +
RFE
RF 0.87 12.4 9.6 RFE improved robustness over
baseline
Wang et al.
(2022)
RF Importance LSTM 0.91 10.2 7.8 RF Importance improved temporal
prediction
Sharma & Saini
(2021)
LASSO SVR 0.84 14.1 10.9 LASSO reduced overfitting in high-
dim data
Ahmed & Badr
(2021)
Sequential FS RF 0.79 15.7 11.5 Sequential FS enhanced
interpretability
Zhang et al.
(2020)
Hybrid FS BiGRU 0.93 9.5 6.7 Hybrid FS gave best accuracy, esp.
with deep learning
Critical Discussion of Feature Selection Techniques
While feature selection methods such as Random Forest, LASSO, and RFE are widely applied, their relative
strengths and weaknesses depend strongly on dataset characteristics, forecasting objectives, and computational
constraints.
Random Forest Feature Importance
Random Forest excels in capturing nonlinear interactions and complex variable dependencies, which are
common in air quality data where meteorological and pollutant factors interact in nonlinear ways (Wang et al.,
2022). Empirical evidence shows that Random Forest consistently identifies meteorological variables (e.g.,
humidity, temperature, wind speed) as key predictors. However, the method may overemphasize correlated
features, inflating their importance scores. Computationally, Random Forest is more demanding than filter-
based methods but remains tractable for medium-to-large datasets, making it well-suited for regional
forecasting tasks where interpretability and robustness are critical.
LASSO Regression
LASSO is particularly effective when the dataset is high-dimensional with many redundant or irrelevant
variables. By shrinking coefficients of weak predictors to zero, it produces sparse, interpretable models
(Sharma & Saini, 2021). Unlike Random Forest, it handles correlated predictors by selecting only one of them,
reducing redundancy. However, it assumes linear relationships, which may limit its predictive performance
when pollutant–meteorological interactions are nonlinear. LASSO is computationally lightweight, making it
suitable for real-time forecasting on constrained devices.
Page 2395
www.rsisinternational.org
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025
Recursive Feature Elimination (RFE)
RFE systematically removes less important features and can improve generalization, but it is computationally
expensive due to repeated model training (Guyon et al., 2002). This makes it impractical for large-scale or real-
time systems, though it remains valuable for offline model refinement and when the dataset size is modest.
Hybrid / Sequential Feature Selection
Empirical studies (e.g., Chen et al., 2023; Zhang et al., 2020) demonstrate that hybrid approaches combining
filter and embedded methods often outperform single-method approaches. These methods balance
interpretability and accuracy, particularly when applied with deep learning models such as BiGRU or LSTM.
The trade-off is increased algorithmic complexity and implementation cost, which may not be practical in
low-resource monitoring systems.
Comparative Insights:
Random Forest outperforms LASSO when nonlinearities and feature interactions dominate, such as in multi-
city or seasonal datasets.
LASSO outperforms Random Forest when the goal is simplicity, interpretability, or real-time deployment
under resource constraints.
RFE is useful for smaller datasets where computational cost is manageable, but less scalable to big-data
scenarios.
Hybrid methods are most effective in research settings aiming for the highest predictive accuracy, but they
require more computational resources.
Discussion of Quantitative Results
Random Forest and LASSO consistently balance predictive power and interpretability.
RFE improves model robustness but is computationally heavy.
Hybrid feature selection (statistical + model-based) often outperforms single-method approaches, especially with
deep learning models.
Across studies, RMSE reductions of 10–20% were reported when applying feature selection compared to using
raw features.
Challenges and Limitations
Despite the progress in applying feature selection techniques to PM2.5 prediction, several challenges and
limitations remain evident across the reviewed literature:
Dataset Heterogeneity and Limited Generalizability
Most studies rely on datasets from specific regions (e.g., Beijing, Delhi, Cairo), with distinct climatic and
pollutant profiles. As a result, feature subsets identified as important in one region may not generalize to another,
particularly across different climates or levels of industrialization. This geographic dependency limits the
transferability of feature selection outcomes.
Assumptions of Static Feature Importance
Many approaches assume that the relative importance of features remains constant over time. However,
empirical findings indicate that meteorological drivers (e.g., humidity, wind speed) and pollutant interactions
Page 2396
www.rsisinternational.org
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025
vary across seasons and long-term policy interventions. Static selection may therefore fail to capture evolving
dynamics, leading to reduced accuracy in long-term forecasting.
Computational Costs and Real-Time Constraints
Techniques such as Recursive Feature Elimination and hybrid feature selection often require repeated training
cycles, making them computationally expensive. While feasible for offline experiments, these methods are
difficult to deploy in real-time air quality monitoring systems that demand rapid inference and resource
efficiency.
Model-Specific Biases
Embedded methods such as Random Forest and LASSO inherently reflect the biases of their underlying models.
Random Forest may overemphasize correlated variables, while LASSO favors sparsity but assumes primarily
linear relationships. This model-dependency can skew feature selection outcomes, potentially overlooking
relevant predictors in complex nonlinear contexts.
Lack of Standardized Benchmarks
Performance evaluation across studies is often inconsistent, with varied metrics (R², RMSE, MAE) and non-
uniform datasets. The absence of benchmark datasets and standardized protocols hinders direct comparison
between techniques and weakens the evidence base for determining best practices.
Limited Integration with Deep Learning and Spatiotemporal Models
Although deep learning architectures (e.g., BiGRU, LSTM) are increasingly used for PM2.5 prediction, most
feature selection studies are confined to traditional ML models. The challenge lies in adapting feature selection
frameworks to high-dimensional, spatiotemporal inputs that characterize modern environmental datasets.
Conclusion and Future Work:
This review highlights that while feature selection significantly improves PM2.5 prediction, most existing
studies operate under implicit assumptions that limit generalizability. For instance, many works assume that the
same feature subset remains optimal across all regions and seasons, whereas empirical evidence suggests
that meteorological drivers vary with geography and time. Similarly, several studies assume linear relationships
between pollutants and PM2.5, which may overlook nonlinear dynamics observed in urban air quality.
Concrete gaps include:
Limited exploration of cross-regional transferability of selected features.
Insufficient attention to computational trade-offs, with many studies prioritizing accuracy while ignoring
deployment constraints in real-time systems.
A lack of dynamic or adaptive feature selection mechanisms that update feature importance as environmental
conditions shift (e.g., seasonal changes, policy interventions reducing pollutant levels).
Future research should therefore:
Develop dynamic, context-aware feature selection algorithms that adapt feature sets based on spatiotemporal
variability.
Investigate lightweight hybrid approaches that balance predictive accuracy with computational efficiency for
real-time monitoring.
Explore transfer learning and domain adaptation to test whether feature sets selected in one region can
generalize to others.
Page 2397
www.rsisinternational.org
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025
Establish benchmark datasets and standardized evaluation metrics to enable fairer comparisons across
studies.
By addressing these gaps, future work can move toward intelligent, robust, and scalable feature selection
frameworks that better support real-world air quality forecasting.
REFERENCES
1. Chen, J., Li, S., Huang, G., & Liu, X. (2023). A robust hybrid feature selection method for high-
dimensional environmental data analysis. Environmental Modelling & Software, 160, 105620.
2. Wang, Y., Liu, Y., Wu, Y., & Zhang, L. (2022). Feature selection and ensemble learning for air quality
prediction: A case study in China. Ecological Indicators, 139, 108930.
3. Sharma, A., & Saini, L. M. (2021). Air quality prediction using optimized feature selection and machine
learning algorithms. Applied Soft Computing, 113, 107872.
4. Ahmed, R. M., & Badr, A. (2021). FSFC-AQ: Feature selection for forecasting city-level air quality using
hybrid approaches. Environmental Science and Pollution Research, 28, 28837–28852.
5. Zhang, Y., Wang, J., & Wang, S. (2020). Air quality prediction based on SSA optimized BiGRU neural
network with hybrid feature selection. Science of The Total Environment, 721, 137763.
6. Chandrashekar, G., & Sahin, F. (2014). A survey on feature selection methods. Computers & Electrical
Engineering, 40(1), 16–28.
7. Gaurav, R., & Behera, H. S. (2021). Hybrid feature selection model for predicting air quality index.
Environmental Processes, 8(2), 883–902.
8. Guyon, I., & Elisseeff, A. (2003). An introduction to variable and feature selection. Journal of Machine
Learning Research, 3(Mar), 1157–1182.
9. Guyon, I., Weston, J., Barnhill, S., & Vapnik, V. (2002). Gene selection for cancer classification using
support vector machines. Machine Learning, 46(1–3), 389–422.
10. Hu, X. M., Klein, P. M., & Xue, M. (2013). Evaluation of the updated YSU planetary boundary layer
scheme within WRF for air quality simulations in the Houston–Galveston area. Atmospheric
Environment, 92, 274–283.
11. Liaw, A., & Wiener, M. (2002). Classification and regression by randomForest. R News, 2(3), 18–22.
12. Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical
Society: Series B (Methodological), 58(1), 267–288.