Page 2428
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
A Dual-Phase Hyperparameter Tuning Approach for Emotion
Detection Using Boosting-Based Machine Learning Algorithms
Dennis S. Nava
1
*, Roman B. Villones
2
, John Joshua E. Mendoza
3
, Rhayz Steven Kyle P. Bautista
4
, Alfred
Brian C. Bautista
5
Graduate School, La Consolacion University Philippines, Malolos, Bulacan, Philippines
*Corresponding Author
DOI:
https://doi.org/10.51583/IJLTEMAS.2026.150500194
Received: 10 May 2026; Accepted: 15 May 2026; Published: 12 June 2026
ABSTRACT
The purpose of this study is to develop and evaluate a dual-phase hyperparameter tuning approach for enhancing
the performance of emotion detection systems using boosting-based machine learning algorithms. The
methodology involves data collection and preprocessing, feature engineering, model definition, training, and
evaluation. Specifically, the study applies a two-stage optimization process in initial coarse tuning with
RandomizedSearchCV followed by fine-tuning with GridSearchCV on models including XGBoost, LightGBM,
CatBoost and GradientBoosting. The results showed that LightGBM achieved the highest overall accuracy of
92.20%, followed by XGBoost with 91.47%, GradientBoosting with 91.19%, and CatBoost with 88.23%.
Confusion matrix analysis revealed that LightGBM and XGBoost produced more balanced and accurate
classifications across the six emotion classes, while CatBoost exhibited higher misclassification rates in
challenging classes. In terms of computational efficiency, LightGBM provided the best balance between
accuracy and training speed, whereas XGBoost demonstrated the lowest memory usage. GradientBoosting
achieved competitive performance but required significantly higher computational resources, while CatBoost
achieved the fastest prediction time. Based on the findings, LightGBM was identified as the most suitable
boosting algorithm for emotion classification due to its superior balance of predictive performance, efficiency,
and reliability. Future studies are recommended to explore hybrid and deep learning approaches, larger datasets,
and real-time implementation strategies to further improve emotion classification systems.
Keywords: Emotion Detection, Hyperparameter Tuning, Boosting Algorithms, Text Classification, Natural
Language Processing.
INTRODUCTION
In an increasingly digitized and interconnected world, the ability to detect human emotions through
computational methods has become critically important across a wide range of applications. Emotion detection
also known as affective computing, allows machines to recognize and respond to human emotional states,
enhancing user experience, and personalization. As modern societies grapple with rising mental health
challenges, detecting emotions through text or social media expressions has also gained prominence in early
diagnosis and intervention strategies.
Recent advancements in machine learning (ML) and natural language processing (NLP) have significantly
improved the ability of systems to understand emotional cues from textual activity. Alswaidan and Menai (2020)
emphasize that emotion recognition from textual data has gained momentum, particularly with the rise of social
media where users frequently express emotions in short and unstructured formats. In healthcare, emotion
detection is increasingly used for monitoring mental health conditions such as depression or anxiety, especially
through analyzing patients’ speech or online activity as explained by Prabhudesai et al. (2021). In education,
emotion-aware tutoring systems can adapt to students’ affective states by improving engagement and learning
outcomes based on Gayed et al. (2022).
Page 2429
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
The primary objective of this study is to develop and evaluate a dual-phase hyperparameter tuning approach to
enhance the performance of emotion detection systems using boosting-based machine learning algorithms. The
approach involves a twostage optimization process: coarse tuning using RandomizedSearchCV, followed by
finetuning using GridSearchCV, then applied to multiple boosting algorithms such as XGBoost, LightGBM,
CatBoost, and GradientBoosting. The goal is to identify the most effective tuning strategy that optimizes
predictive accuracy and computational efficiency in emotion classification tasks.
This study contributes to the growing field of affective computing by introducing a dual-phase hyperparameter
tuning approach that systematically enhances the performance of boosting-based machine learning algorithms
for emotion detection. The proposed methodology offers a balanced and efficient framework for model
optimization. The study further provides a comparative analysis of six popular boosting algorithms and
demonstrating how tuning strategies impact their classification performance in emotion detection tasks. The
approach offers scalability and reproducibility, making it adaptable to other NLP and text classification problems
beyond emotion detection.
LITERATURE REVIEW
Review of Related Studies
In the study conducted by Alswaidan and Menai (2020) emotion recognition from textual data has gained
significant attention because social media platforms allow users to express their emotions through posts and
online interactions. The researchers explained that affective computing improves human-computer interaction
by enabling systems to identify and respond to emotional states. Similarly, Khalil et al. (2019) discussed the
effectiveness of deep learning techniques in speech emotion recognition using facial expressions, speech, and
audiovisual inputs, particularly in healthcare surveillance applications. Furthermore, Yan et al. (2022) explored
emotion classification using physiological signals such as EEG and ECG, emphasizing the importance of
preprocessing, feature extraction, and multimod al fusion techniques in improving classification accuracy.
The study of Bischl et al. (2023) provided a comprehensive review of hyperparameter optimization methods and
compared the effectiveness of RandomizedSearchCV and GridSearchCV in improving machine learning model
performance. In addition, Khansa et al. (2025) applied both optimization methods to the XGBoost algorithm for
Parkinson’s disease classification and found that GridSearchCV achieved better predictive accuracy. Malhotra
and Cherukuri (2024) also explained that hyperparameter tuning techniques significantly contribute to improving
software quality prediction models by optimizing model generalization and predictive capability. These studies
highlighted the importance of parameter optimization in machine learning systems.
According to Sadaf (2023) CatBoost and XGBoost algorithms achieved strong performance in phishing website
detection systems by accurately classifying malicious URLs. Similarly, Kumar and Guleria (2024) utilized
CatBoost and LightGBM algorithms in AI-enhanced cybersecurity datasets and reported high accuracy in
detecting anomalous network traffic. The study of Athanasiou and Maragoudakis (2017) also demonstrated that
GradientBoosting is effective in sentiment analysis tasks even in low-resource language environments. These
studies collectively support the effectiveness of boosting algorithms in classification and sentiment analysis
tasks.
Review of Related Systems
In the study conducted by Al-Zakhali et al. (2024) an XGBoost-based text classification system was developed
to improve document classification performance through parallel processing and boosting tree techniques. The
researchers highlighted that the system achieved efficient prediction speed and high classification accuracy when
handling textual datasets. Similarly, Lokker et al. (2024) implemented a LightGBM-based classification system
for biomedical research abstracts and demonstrated that the system could rapidly process high-dimensional
textual data while maintaining strong predictive performance.
According to Narasamma and Sreedevi (2021) a CatBoost-based Twitter sentiment analysis system was designed
to improve sentiment classification by efficiently handling textual and categorical data from social media
Page 2430
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
platforms. In addition, Prabhakar et al. (2019) utilized AdaBoost in a sentiment analysis system for airline
Twitter data, where the algorithm improved classification performance by iteratively correcting weak
learnererrors. Moreover, Maftoun et al. (2024) developed a HistGradientBoosting system for malicious URL and
fake news detection, emphasizing its capability to process sparse and large-scale textual features efficiently.
These systems demonstrate the practical applications of boosting algorithms in natural language processing and
emotion detection environments.
The study of Gayed et al. (2022) introduced an AI-based emotion-aware tutoring system capable of adapting
educational content according to students’ emotional states to improve learning engagement and performance.
Similarly, Prabhudesai et al. (2021) developed a depression detection and analysis system using deep learning
techniques to identify mental health conditions through speech and online activities. Furthermore, Yan et al.
(2022) proposed an emotion recognition system using multichannel physiological signals combined with
adaptive decision fusion methods to accurately classify emotional responses. These systems highlight the
growing importance of emotion detection technologies in healthcare, education, and intelligent human-computer
interaction systems.
METHODOLOGY
This study employed a structured machine learning pipeline to improve the performance of emotion detection
using boosting-based algorithms. The methodology consisted of five main stages: data collection and
preprocessing, feature engineering, model definition, hyperparameter optimization, and model training and
evaluation.
Data Collection & Preprocessing
The dataset used in this study was sourced from Kaggle, a reputable platform for data science competitions and
dataset sharing. It initially contained a total of 422,746 rows, each comprising two columns: one for the textual
content “sentence” and another for the associated emotional label “emotion”. A thorough data validation process
confirmed that there were no missing values, ensuring the dataset’s completeness and reliability for subsequent
analysis.
Figure 1. Emotion Distribution based on the dataset.
Page 2431
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
As illustrated in Figure 1, the dataset exhibited a significant imbalance in the distribution of emotional categories:
'fear', 'sad', 'love', 'joy', 'surprise', and 'anger'. Most of the instances were labeled as 'joy' (143,067) and 'sad'
(121,187), followed by 'anger' (59,317), 'fear' (49,649), and 'love' (34,554). The category with the fewest samples
was 'surprise' (14,972). However, an imbalance was observed in the distribution of samples across these classes,
which posed a risk of bias during model training.
Figure 2. Emotion Distribution after balancing.
Figure 2 illustrates the transformation of the original imbalanced dataset into a balanced form by applying a
down sampling technique. Initially, the dataset exhibited highly unequal class distributions, with the number of
instances ranging from 14,972 ('surprise') to 143,067 ('joy'). To mitigate this imbalance and prevent model bias,
the dataset was down sampled to ensure equal representation across all six emotion categories with each class
containing 14,972 instances.
Feature Engineering
Feature engineering plays a critical role in transforming raw data into formats suitable for machine learning
algorithms. In this phase, feature engineering involved two main steps:
1. Target Encoding - The categorical target variable representing discrete emotion labels was numerically
encoded using label mapping. Specifically, the six emotion classes were converted into integer labels as
follows: 'anger': 0, 'fear': 1, 'joy': 2, 'love': 3, 'sad': 4, 'surprise': 5. This transformation was implemented using
the ‘fit_transform’ method of a label encoder which ensures that the categorical labels are represented in a
format compatible with classification models.
2. Text Vectorization - The input feature is consisting of textual data and transformed into numerical
representations using Term Frequency–Inverse Document Frequency (TF-IDF) Vectorization. According to
Triana et al. (2025) the TF-IDF is a well-established technique in natural language processing that converts
text into feature vectors while reducing the impact of commonly used words and highlighting more
informative terms. The TF-IDF was configured with a `max_features` parameter set to 5000, thereby limiting
the feature space to the top 5000 most informative terms across the sentenses.
Model Definition
This study explores the effectiveness of advanced boosting algorithms in enhancing the accuracy of emotion
detection systems and occurs the following:
Page 2432
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
1. XGBoost - The study of Al-Zakhali et al. (2024) applied the TF-IDF-transformed text to efficiently classify
documents and leveraging its parallel boosting trees to improve prediction speed and accuracy.
2. LightGBM - As explained by Lokker et al. (2024) LighthGBM was employed to classifying the biomedical
research abstract, demonstrating rapid training and strong performance, and high-dimensional textual data.
3. CatBoost – Integrated with textual features from Twitter posts to improve sentiment classification and
benefiting from its native handling of categorical and textual data as mentioned by Narasamma and Sreedevi
(2021).
4. GradientBoosting – Based on Athanasiou and Maragoudakis (2017) GradientBoosting applied to TF-IDF
features for sentiment analysis in Greek, demonstrating that boosting methods remain effective even in low-
resource NLP settings.
RESULTS
Hyperparameter Optimization
Phase I – Coarse Hyperparameter Optimization with RandomizedSearchCV
This study adopts a coarse-to-fine tuning strategy, where RandomizedSearchCV is initially employed as a coarse
hyperparameter optimization technique. Five boosting-based algorithms were subjected to coarse
hyperparameter tuning: LightGBM, XGBoost, GradientBoosting, and CatBoost. Each model was independently
tuned with a defined parameter distribution and 5-fold cross-validation (CV).
Boosting Model
Best Parameters
XGBoost
subsample = 0.8, n_estimators = 200, max_depth = 3, learning_rate = 0.2, colsample_bytree = 1.0,
random_state = 42;
LightGBM
num_leaves = 31, n_estimators = 100, max_depth = -1, learning_rate = 0.05, random_state=42;
CatBoost
learning_rate = 0.1, l2_leaf_reg = 1, iterations = 200, depth = 7, random_state=42;
GradientBoosting
subsample = 0.8, n_estimators = 200, max_depth = 5, learning_rate = 0.1, random_state=42;
Table 1. RandomSearchCV best parameters results.
Phase II – Fine-Tuning of Parameters with GridSearchCV
GridSearchCV is employed in the second phase for fine hyperparameter optimization. After the initial coarse
search, GridSearchCV systematically evaluates combinations of hyperparameters within narrowed ranges to fine-
tune model performance. Four boosting-based algorithms were subjected to fine hyperparameter tuning:
LightGBM, XGBoost, GradientBoosting, and CatBoost. The tuning process was executed using 5-fold cross-
validation (CV) to ensure robustness and generalization of the optimized models.
Boosting Model
Best Parameters
XGBoost
colsample_bytree
=
0.9, learning_rate
=
0.25, max_depth = 4,n_estimators = 220,
subsample = 0.8, random_state = 42
LightGBM
n_estimators = 100, max_depth = -1, learning_rate = 0.07,num_leaves = 31,
random_state=42
CatBoost
iterations = 180, depth = 7, learning_rate = 0.12, l2_leaf_reg = 0.5,random_state=42
GradientBoosting
n_estimators = 200, max_depth = 6, learning_rate = 0.08, subsample = 0.8,
random_state=42
Table 2. GridSearchCV best parameters results.
Page 2433
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Model Training and Evaluation
XGBoost
Precision
Recall
F1-score
Support
0
0.93
0.91
0.92
4491
1
0.91
0.86
0.88
4491
2
0.9
0.87
0.89
4492
3
0.9
0.97
0.94
4492
4
0.96
0.88
0.92
4492
5
0.89
0.98
0.94
4492
accuracy
0.91
26950
macro avg
0.92
0.91
0.91
26950
weighted avg
0.92
0.91
0.91
26950
Table 3. Classification Results for XGBoost.
The XGBoost model achieved an overall accuracy of 91% after applying two-phase hyperparameter tuning using
Random Search and Grid Search. The model obtained a macro average precision of 92%, recall of 91%, and F1-
score of 91%, indicating strong and balanced classification performance across the six emotion categories. The
results demonstrate that the model is effective and reliable in predicting emotions from the dataset.
LightGBM
Precision
Recall
F1-score
Support
0
0.91
0.94
0.93
4491
1
0.91
0.87
0.89
4491
2
0.96
0.87
0.91
4492
3
0.9
0.97
0.94
4492
4
0.96
0.89
0.93
4492
5
0.89
0.98
0.94
4492
accuracy
0.92
26950
macro avg
0.92
0.92
0.92
26950
weighted avg
0.92
0.92
0.92
26950
Table 4. Classification Results for LightGBM.
The LightGBM model achieved an overall accuracy of 92% after hyperparameter tuning. It obtained a macro
average precision, recall, and F1-score of 92%, with a weighted average of 92% across all metrics. These results
indicate strong and well-balanced performance across all six emotion classes, demonstrating the model’s
effectiveness and reliability in emotion classification.
CatBoost
Precision
Recall
F1-score
Support
0
0.92
0.88
0.9
4491
1
0.91
0.84
0.87
4491
2
0.76
0.87
0.81
4492
3
0.9
0.95
0.92
4492
4
0.95
0.77
0.85
4492
5
0.89
0.98
0.93
4492
accuracy
0.88
26950
Page 2434
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
macro avg
0.89
0.88
0.88
26950
weighted avg
0.89
0.88
0.88
26950
Table 5. Classification Results for CatBoost.
The CatBoost model achieved an overall accuracy of 88% after hyperparameter tuning. It obtained a macro
average precision of 89%, recall of 88%, and F1-score of 88%, with similar weighted averages across all metrics.
While the model shows generally good performance across the six emotion classes, its results are slightly lower
and less balanced compared to XGBoost and LightGBM, particularly in some emotion categories.
GradientBoosting
Precision
Recall
F1-score
Support
0
0.92
0.91
0.92
4491
1
0.9
0.85
0.88
4491
2
0.9
0.87
0.89
4492
3
0.9
0.98
0.94
4492
4
0.96
0.88
0.92
4492
5
0.89
0.98
0.93
4492
accuracy
0.91
26950
macro avg
0.91
0.91
0.91
26950
weighted avg
0.91
0.91
0.91
26950
Table 6. Classification Results for GradientBoosting.
The GradientBoosting model achieved an overall accuracy of 91% after hyperparameter tuning. It obtained a
macro average precision, recall, and F1-score of 91%, with consistent weighted averages across all metrics. The
results indicate strong and stable performance across all six emotion classes, showing that the model is effective
and well-balanced in emotion classification.
Page 2436
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Compared to the XGBoost and LightGBM results, the CatBoost and GradientBoosting models also demonstrate
strong overall classification performance, with predictions largely concentrated along the diagonal of the
confusion matrices. However, the performance difference between CatBoost and GradientBoosting appears more
noticeable, particularly in handling difficult class separations such as the confusion between Classes 4 and 2.
While CatBoost exhibits stronger misclassification patterns in these challenging classes, GradientBoosting
significantly reduces these errors and achieves more balanced predictions across multiple classes. In contrast,
the comparison between XGBoost and LightGBM shows more subtle differences, where both models maintain
highly similar error distributions and only slight advantages in specific emotion categories. Overall,
GradientBoosting shows a clearer improvement over CatBoost, whereas XGBoost and LightGBM demonstrate
more closely matched performance with only marginal differences in overall accuracy and class discrimination.
Boosting Model
Accuracy
Training Time (s)
Prediction Time (s)
Peak Memory Usage (MB)
LightGBM
0.922004
20.62865
0.561455
7.824903
XGBoost
0.914731
58.55378
0.206033
2.599951
GradientBoosting
0.911911
1307.367
1.482315
26.79842
CatBoost
0.882338
296.5523
0.069385
11.14843
Table 7. Ranking of Model by Performance Metrics.
Based on the overall performance metrics, LightGBM achieved the highest accuracy among the four boosting
models at 92.20%, followed by XGBoost at 91.47%, GradientBoosting at 91.19%, and CatBoost at 88.23%. In
terms of computational efficiency, LightGBM maintained a strong balance between accuracy and training speed,
requiring only 20.63 seconds for training, while XGBoost required a longer training time of 58.55 seconds but
achieved the lowest memory usage at 2.60 MB. GradientBoosting, although competitive in accuracy, showed the
slowest performance with a significantly higher training time of 1307.37 seconds and the highest memory
consumption at 26.80 MB, making it less efficient for large-scale applications. Meanwhile, CatBoost recorded
the fastest prediction time at 0.069 seconds, indicating strong inference efficiency despite having the lowest
accuracy among the models. Overall, LightGBM demonstrates the best trade-off between accuracy, speed, and
resource utilization, while XGBoost provides a memory-efficient alternative with competitive predictive
performance.
DISCUSSION
The findings of this study demonstrate that boosting-based machine learning algorithms are highly effective for
multiclass emotion classification, particularly when combined with a two-phase hyperparameter optimization
strategy using RandomizedSearchCV and GridSearchCV. The adopted coarse-to-fine tuning approach enabled
the models to first explore a broad hyperparameter space and subsequently refine the most promising parameter
combinations, leading to improved predictive performance and stronger model generalization. The use of 5-fold
cross-validation throughout both tuning phases further enhanced the reliability of the results by minimizing
overfitting and ensuring consistent evaluation across different data partitions.
Among the evaluated models, LightGBM achieved the highest overall performance with an accuracy of 92.20%,
together with balanced macro-average precision, recall, and F1-score values of 92%. This indicates that the
model was able to classify the six emotion categories consistently and effectively without favoring specific
classes. The superior performance of LightGBM can be attributed to its leaf-wise tree growth strategy and
efficient gradient-based optimization, which allow it to capture complex feature interactions while maintaining
computational efficiency. In addition, LightGBM achieved a relatively short training time of only 20.63 seconds,
demonstrating its capability to deliver both high predictive accuracy and fast model training. These
characteristics make LightGBM highly suitable for real-time emotion classification systems and large-scale
applications where both speed and accuracy are critical.
XGBoost also demonstrated strong and competitive performance, achieving an accuracy of 91.47% with
balanced precision, recall, and F1-score metrics. Although its predictive performance was slightly lower than
Page 2437
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
LightGBM, the model showed strong capability in handling minority misclassification cases and distinguishing
closely related emotion categories. The confusion matrix results revealed that XGBoost maintained stable
classification behavior across challenging classes, indicating strong generalization capability. Furthermore,
XGBoost recorded the lowest memory consumption among all models at 2.60 MB, highlighting its efficiency in
resource utilization. This suggests that XGBoost remains a practical choice for deployment in environments with
limited computational resources or memory constraints.
GradientBoosting achieved an overall accuracy of 91.19%, producing performance results that were close to
XGBoost. The confusion matrix analysis showed that GradientBoosting significantly reduced several major
misclassification cases, particularly between difficult class pairs, resulting in more balanced class predictions
compared to CatBoost. However, despite its strong predictive capability, GradientBoosting exhibited the slowest
computational performance, requiring more than 1300 seconds for training and consuming the highest amount
of memory among the evaluated models. These findings indicate that while GradientBoosting can provide
competitive classification performance, its high computational cost limits its practicality for large datasets and
real-time implementations.
In comparison, CatBoost achieved the lowest overall accuracy at 88.23%, although it still demonstrated generally
good classification performance across the six emotion categories. The confusion matrix revealed higher
misclassification rates in certain challenging classes, suggesting that the model had more difficulty distinguishing
emotions with overlapping characteristics. Nevertheless, CatBoost recorded the fastest prediction time at only
0.069 seconds, indicating excellent inference efficiency during deployment. This implies that while CatBoost
may not provide the highest predictive accuracy in this study, it may still be beneficial in applications where rapid
prediction speed is prioritized over maximum classification performance.
CONCLUSION AND RECOMMENDATIONS
This study successfully developed and evaluated a dual-phase hyperparameter tuning approach for emotion
detection using boosting-based machine learning algorithms. The findings revealed that hyperparameter tuning
has a significant but model-dependent impact on classification performance. Among the evaluated algorithms,
LightGBM achieved the highest accuracy, followed by XGBoost and GradientBoosting, demonstrating their
effectiveness in handling text-based emotion classification tasks. The study also confirmed that combining
RandomizedSearchCV and GridSearchCV provides a systematic optimization process that improves predictive
capability and model generalization. However, the results also showed that improper or excessive tuning may
negatively affect performance, as observed in the CatBoost model. Overall, the study highlights the importance
of selecting appropriate optimization strategies and validation techniques when developing emotion detection
systems.
Based on the findings, future researchers are encouraged to apply model-specific tuning strategies and explore
advanced approaches such as deep learning and transformer-based architectures to further improve emotion
classification accuracy. Educational institutions, healthcare organizations, and developers may also adopt
boosting-based emotion detection systems for applications in sentiment analysis, mental health monitoring, and
intelligent human-computer interaction. Furthermore, future studies should utilize larger and multilingual
datasets, improve preprocessing techniques, and conduct continuous testing and evaluation to ensure system
reliability, scalability, and real-world effectiveness.
ACKNOWLEDGEMENT
We would like to thank Kaggle.com for sharing the publicly available datasets for this study. Their platform
made it possible for me to carry out the analysis and evaluation of the machine learning algorithms presented
here. We are also grateful for the support provided by the Graduate School Department in helping with their
support throughout this research.
Ethical Approval
This study did not get funding from any institution.
Page 2438
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Conflict of Interest
The author declared that there is no conflict of interest.
Data Availability
The data that support the findings of this study are available from the authors upon reasonable request.
Authors Biography
Dennis S. Nava is pursuing his Doctorate degree in Information Technology in La Consolacion University
Philippines, Malolos, Bulacan, Philippines. He also a college instructor in College of Engineering and
Information Sciences, Trinity University of Asia, Quezon City, Philippines.
Roman B. Villones is pursuing his Doctorate degree in Information Technology in La Consolacion University
Philippines, Malolos, Bulacan, Philippines. He also a college instructor and researcher in College of Informatics,
Philippine Christian University, Manila, Philippines. He has authored several peer-reviewed publications on
software engineering and now focuses on classification algorithms, model optimization, and educational data
mining.
John Joshua E. Mendoza is pursuing his Doctorate degree in Information Technology in La Consolacion
University Philippines, Malolos, Bulacan, Philippines. He also a college instructor and department head in
College of Informatics, Philippine Christian University, Manila, Philippines.
Rhayz Steven Kyle P. Bautista is pursuing his Doctorate degree in Information Technology in La Consolacion
University Philippines, Malolos, Bulacan, Philippines. He also a college instructor and information assurance
security expert in College of Informatics, Philippine Christian University, Manila, Philippines.
Alfred Brian C. Bautista is pursuing his Doctorate degree in Information Technology in La Consolacion
University Philippines, Malolos, Bulacan, Philippines. He also a college instructor at Manila Central University,
Metro Manila, Philippines
REFERENCES
1. Al-Zakhali, O. A., Zeebaree, S., & Askar, S. (2024). Comparative analysis of XGBoost performance for
text classification with CPU parallel and non-parallel processing. The Indonesian Journal of Computer
Science, 13(2). https://doi.org/10.33022/ijcs.v13i2.3798
2. Alswaidan, N., & Menai, M. E. B. (2020). A survey of state-of-the-art approaches for emotion
recognition in text. Knowledge and Information Systems, 62(8), 2937–2987.
https://doi.org/10.1007/s10115-020-01449-0
3. Athanasiou, V., & Maragoudakis, M. (2017). A novel, gradient boosting framework for sentiment
analysis in languages where NLP resources are not plentiful: A case study for modern Greek. Algorithms,
10(1), 34. https://doi.org/10.3390/a10010034
4. Bischl, B., Binder, M., Lang, M., Pielok, T., Richter, J., Coors, S., Thomas, J., Ullmann, T., Becker, M.,
Boulesteix, A., Deng, D., & Lindauer, M. (2023). Hyperparameter optimization: Foundations,
algorithms, best practices, and open challenges. Wiley Interdisciplinary Reviews: Data Mining and
Knowledge Discovery, 13(2), e1484. https://doi.org/10.48550/arXiv.2107.05847
5. Gayed, J. M., Carlon, M. K. J., Oriola, A. M., & Cross, J. S. (2022). Exploring an AI-based writing
assistant's impact on English language learners. Computers and Education: Artificial Intelligence, 3,
100055.
https://doi.org/10.1016/j.caeai.2022.100055
6. Khalil, R. A., Jones, E., Babar, M. I., Jan, T., Zafar, M. H., & Alhussain, T. (2019). Speech emotion
recognition using deep learning techniques: A review. IEEE Access, 7, 117327–117345.
https://doi.org/10.1109/ACCESS.2019.2936124
7. Khansa, S. F. A., Ulinnuha, N., & Utami, W. D. (2025). Grid search and random search hyperparameter
tuning optimization in XGBoost algorithm for Parkinson’s disease classification. BAREKENG: Jurnal
Page 2439
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Ilmu Matematika dan Terapan, 19(3), 1609–1624.
https://doi.org/10.30598/barekengvol19iss3pp16091624
8. Kumar, A., & Guleria, K. (2024, November). Leveraging machine learning algorithms for threat
detection using AI-enhanced cybersecurity datasets. In 2024 4th International Conference on
Technological Advancements in Computational Sciences (ICTACS) (pp. 483–488). IEEE.
https://doi.org/10.1109/ICTACS62700.2024.10840478
9. Lokker, C., Abdelkader, W., Bagheri, E., Parrish, R., Cotoi, C., Navarro, T., Germini, F., Linkins, L.,
Haynes, R., Chu, L., Afzal, M., & Iorio, A. (2024). Boosting efficiency in a clinical literature surveillance
system with LightGBM. PLOS Digital Health, 3(9), e0000299.
https://doi.org/10.1371/journal.pdig.0000299
10. Malhotra, R., & Cherukuri, M. (2024). A systematic review of hyperparameter tuning techniques for
software quality prediction models. Intelligent Data Analysis, 28(5), 1131–1149.
https://doi.org/10.3233/IDA-230653
11. Narasamma, V. L., & Sreedevi, M. (2021). Twitter based data analysis in natural language processing
using a novel CatBoost recurrent neural framework. International Journal of Advanced Computer Science
and Applications, 12(5). https://dx.doi.org/10.14569/IJACSA.2021.0120555
12. Prabhudesai, S., Mhaske, A., Parmar, M., & Bhagwat, S. (2021, June). Depression detection and analysis
using deep learning: Study and comparative analysis. In 2021 10th IEEE International Conference on
Communication Systems and Network Technologies (CSNT) (pp. 570–574). IEEE.
https://doi.org/10.1109/CSNT51715.2021.9509707
13. Sadaf, K. (2023, February). Phishing website detection using XGBoost and CatBoost classifiers. In 2023
International Conference on Smart Computing and Application (ICSCA) (pp. 1–6). IEEE.
https://doi.org/10.1109/ICSCA57840.2023.10087829
14. Salvador, E. L. (2024). Use of boosting algorithms in household-level poverty measurement: A machine
learning approach to predict and classify household wealth quintiles in the Philippines. arXiv Preprint
arXiv:2407.13061.
https://doi.org/10.48550/arXiv.2407.13061
15. Sengar, S., & Liu, X. (2020). Ensemble approach for short term load forecasting in wind energy system
using hybrid algorithm. Journal of Ambient Intelligence and Humanized Computing, 11(11), 5297–5314.
https://doi.org/10.1007/s12652-020-01866-7
16. Triana, E., Purnamasari, A. I., Bahtiar, A., & Tohidi, E. (2025). Improved spam email detection
performance based on Naïve Bayes approach TF-IDF vectorizer with multi-metric optimization. Journal
of Artificial Intelligence and Engineering Applications (JAIEA), 4(3), 1667–1672.
https://doi.org/10.59934/jaiea.v4i3.981
17. Yan, M., Deng, Z., He, B., Zou, C., Wu, J., & Zhu, Z. (2022). Emotion classification with multichannel
physiological signals using hybrid feature and adaptive decision fusion. Biomedical Signal Processing
and Control, 71, 103235.
https://doi.org/10.1016/j.bspc.2021.103235
