Twitter Sentiment Classification Using ESOA Based Feature Selection with MHAM-DMO Model

1. Introduction

Sentiment analysis has become a vital part of modern data analysis, helping organizations and researchers understand public opinions shared across digital platforms. With increasing use of social media, people regularly express their views about products, services, events, and other topics, making platforms like Twitter a valuable source of short, real-time textual data. However, analyzing these user-generated posts is challenging due to the informal writing style, noise, inconsistent grammar, emojis, abbreviations, and limited context typically found in tweets.

Many studies have explored sentiment analysis using machine learning and deep learning frameworks. Traditional approaches often rely on manually engineered features, which limits their ability to generalize across highly dynamic linguistic patterns in social media text. Recent work highlights the need for more robust and scalable systems capable of handling short and noisy content. For example, Carvalho and Plastino [1] demonstrated that the performance of Twitter sentiment models greatly depends on the quality of extracted features, while Kumar and Jaiswal [4] identified shortcomings in conventional soft computing techniques when applied to high-volume Twitter data. Similarly, Pota et al. [6] showed that deep contextual embeddings such as BERT significantly improve sentiment prediction but still face difficulties when handling sparse, high-dimensional input representations.

In addition to feature extraction challenges, optimization of model parameters plays a crucial role in achieving stable and accurate results. Many recent sentiment analysis studies emphasize the importance of utilizing attention mechanisms and neural architectures that can capture semantic nuances in short texts [9,18]. However, these models often struggle with redundant attention focus, suboptimal feature selection, and manually tuned hyperparameters, which introduce instability and reduce overall effectiveness.

To address these limitations, our work proposes a unified sentiment classification framework that integrates advanced natural language processing with hybrid optimization techniques. By incorporating TF-IDF based text representations, the Egret Swarm Optimization Algorithm (ESOA) for feature selection, a Multi-Head Attention Mechanism (MHAM) for semantic learning, and Dwarf Mongoose Optimization (DMO) for hyperparameter tuning, the system aims to overcome the challenges previously highlighted in the literature. The goal is to develop a more efficient and accurate method capable of handling the complexities of Twitter data while leveraging optimization-driven enhancements to improve classification performance.

2. Literature Review

Sentiment analysis on Twitter has been widely studied using various methods, including traditional machine learning techniques, deep learning, and models based on optimization. Earlier work mainly used classic classifiers like Support Vector Machines, Logistic Regression, and Decision Trees. These models had some success but had trouble handling the complex and noisy data found on Twitter. For example, Carvalho and Plastino [1] noted that the effectiveness of these classical models depends heavily on how well the chosen features can distinguish between different sentiments. Similarly, Mehta and Pandya [8] looked at different machine learning methods and found that traditional algorithms aren’t very strong when dealing with informal and quickly changing language on social media. Decision tree-based sentiment classifiers, like those developed by Kasthuri and Jebaseeli [2], provided models that were easier to understand, but they still faced issues with sparse text data and repeated features.

To overcome the problems with manually creating features, researchers started using soft computing and hybrid feature-based approaches. Kasthuri and Jebaseeli [5,7,10] looked into nature-inspired optimization, fuzzy logic, and mixed methods to improve feature selection and classification accuracy. Their findings showed better performance than traditional methods, but these techniques were still sensitive to noisy data and needed a lot of fine-tuning. Kumar and Jaiswal [4] and Kasthuri and Jebaseeli [2] also found that soft computing models worked well but were dependent on high-quality, manually crafted features, which made them hard to use in real-time Twitter settings. Multi-objective optimization models, like the LAN2FIS model introduced by Nagamanjula and Pethalakshmi [14], showed some improvements but came with higher computational costs.

Deep learning became a strong option that can learn meaning patterns directly from text. Models based on LSTM and CNN have shown big improvements, especially for short and casual tweets. For example, Kaur and Sharma [16] created a mixed deep learning model that uses different ways to extract features for analyzing consumer opinions. Krishna et al. [17] used independent component analysis to improve sentiment classification using deep learning. Yin et al. [18] introduced the DPG-LSTM model, which uses dependency parsing and graph convolution to understand both sentence structure and meaning in social media text. These studies showed that deep learning models are better than traditional methods, but they still face challenges like dealing with large amounts of data, messy information, and complicated settings.

Transformer-based models took the field even further by using contextual embeddings. Pota et al. [6] made a BERT-based system for analyzing sentiment on Italian tweets, and it worked well because BERT can understand relationships between words in context. Naseem et al. [9] used transformer-based embeddings that are smart and context-aware, which helped improve accuracy and make models more reliable for sentiment tasks. Wankhade and Rao [21] made a BERT-BiLSTM model that works well for analyzing sentiment about COVID-19, showing how combining transformer embeddings with recurrent models helps handle short and specialized tweets. Azzouza et al. [15] built on this by creating TwitterBERT, a system designed specifically for the types of text found in tweets.

Alongside progress in neural modeling, researchers have looked into using emojis, emoticons, and other language clues to better understand emotions in text. Ullah and others [23] suggested combining text with emoticons, showing how non-verbal expressions are key for predicting emotions. Rachman [24] showed that using TF-IDF along with logistic regression can work well if emotive symbols are kept. Studies by Manguri et al. [3] and Neogi et al. [11] also found that using language patterns specific to certain areas can greatly improve emotion analysis, especially in areas like tracking public health responses to COVID-19 and studying social movements.

There’s also been interest in optimization-based sentiment models, especially for choosing the best features and adjusting model settings. Kasthuri and Jebaseeli [2] introduced a mixed optimization method, showing how nature-inspired techniques can help pick the best features. In stock market prediction, Albahli et al. [19] used machine learning on real-time Twitter data, stressing the need to fine-tune model settings for consistent results. Tesfagergish et al. [20] developed a zero-shot learning method for detecting emotions using sentence transformers and group learning techniques, showing that optimization methods are becoming more important in analyzing text.

In this varied collection of work, a few common issues keep showing up. Many current models have trouble with messy or repeated features, lack of different meanings in their attention systems, and depend too much on manually adjusted settings. While deep learning and transformer models have better ways of representing data, they still don’t perform well if problems like repeated or unstable features aren’t fixed. These shortcomings show that we need a combined approach that includes better ways to pick features, improved attention mechanisms, and automatic tuning of settings—this is why the ESOA-MHAM-DMO model was developed in this study.

3. Problem Statement & Objecctives

3.1. Problem Statement

Sentiment analysis of Twitter data is essential for understanding public opinion across various domains such as product reviews, social events, and policy discussions. However, effective sentiment classification remains a challenging task due to several inherent characteristics of Twitter data. Tweets are often short, noisy, and written in informal language, frequently containing abbreviations, misspellings, emojis, hashtags, and inconsistent grammatical structures. These features make it difficult for traditional sentiment analysis models to accurately capture meaningful contextual and emotional information.

Existing machine learning and deep learning approaches commonly rely on high-dimensional text representations, such as TF-IDF, which often include redundant and irrelevant features. This leads to increased computational complexity and degraded classification performance. Additionally, many deep learning models employ attention mechanisms that tend to focus on overlapping semantic regions, limiting their ability to capture diverse sentiment cues present in short text sequences. Another major limitation observed in current approaches is the dependence on manually selected hyperparameters, which leads to unstable model performance and poor generalization across datasets.

Although recent studies have introduced optimization algorithms and advanced neural architectures, they often address feature selection, attention mechanisms, or hyperparameter tuning in isolation. A unified framework that simultaneously reduces feature redundancy, enhances semantic representation, and optimizes training parameters is still lacking. These challenges motivate the need for an integrated and optimized sentiment classification model capable of handling the complexity and variability of Twitter data more effectively.

3.2. Objectives of the Study

The goal of this research is to develop a strong and efficient sentiment analysis system for Twitter data by combining feature selection, deep learning, and optimization techniques. The specific aims of this study are as follows:

• To create an effective preprocessing and feature extraction process that transforms raw Twitter text into meaningful numerical data while retaining sentiment information.
• To apply the Egret Swarm Optimization Algorithm (ESOA) for selecting relevant and non-redundant features from high-dimensional TF-IDF representations, thereby reducing computational complexity and enhancing classification accuracy.
• To build a sentiment classification model utilizing a Multi-Head Attention Mechanism (MHAM) that captures diverse semantic patterns and improves the contextual understanding of short Twitter messages.
• To optimize the hyperparameters of the deep learning model using Dwarf Mongoose Optimization (DMO) in order to improve training stability and overall performance.
• To evaluate the effectiveness of the proposed framework using the Sentiment140 dataset and compare its performance with existing sentiment analysis approaches based on standard evaluation metrics.

4. Proposed System

The system suggested uses a complete method for analyzing emotions in tweets. It aims to correctly sort the tweets by using natural language processing, better ways to pick important features, and deep learning to classify the sentiment. The process begins with getting the data ready, then cleaning and preparing the text. After that, features are taken out using TF-IDF. The most helpful features are then picked using the Egret Swarm Optimization Algorithm (ESOA). Finally, the sentiment is determined using a Seq2Seq model with a Multi-Head Attention Mechanism (MHAM). The best settings for the model are fine-tuned using the Dwarf Mongoose Optimization (DMO) algorithm.

4.1. Dataset Description for Sentiment140

The Sentiment140 dataset is created by collecting tweets from twitter.com and is commonly used for large-scale sentiment analysis research. Each tweet is automatically labeled based on the emoticons found within the text, eliminating the need for manual labeling. Positive sentiment is identified using emoticons like :), while negative sentiment is identified using :(.

The dataset consists of six attributes:

• Tweet polarity
• Tweet ID
• Date
• Query
• Username
• Tweet text

In this study, only the tweet ID, tweet text, and polarity attributes are used. Sentiment labels are encoded as 0 for negative sentiment and 4 for positive sentiment. The dataset includes 1,600,000 labeled tweets, split evenly into 800,000 positive and 800,000 negative examples. When looking at how the tweets are structured, the longest one has 110 words, but the average length is around 14 words.

Table 1. Sample Instances from the Sentiment140 Dataset.

Tweet ID	Tweet Text	Polarity
1467810672	feels cheated out of the ability to text Facebook status updates	0
1467810917	@Kenichan Numerous times, I dove for the ball	0
1467824967	No jobs = no income. How the hell is the minimum wage	0
2065732947	@KateHope thanks for having my back Haha	4
2065733289	Finally home and getting stuff done before an awesome birthday dinner...	4
2065782488	@DerrenLitten Oh, fantastic, I checked it out while napping	4

Table 1. Sample Instances from the Sentiment140 Dataset.

Tweet ID	Tweet Text	Polarity
1467810672	feels cheated out of the ability to text Facebook status updates	0
1467810917	@Kenichan Numerous times, I dove for the ball	0
1467824967	No jobs = no income. How the hell is the minimum wage	0
2065732947	@KateHope thanks for having my back Haha	4
2065733289	Finally home and getting stuff done before an awesome birthday dinner...	4
2065782488	@DerrenLitten Oh, fantastic, I checked it out while napping	4

All the labeled data is in one file called training.1600000.processed.noemoticon.csv.

To make sure results are consistent, a fixed random seed of 2019 was used. The dataset was split into 80% for training and 20% for testing.

The training part was then divided again, with 90% used for actual training and 10% set aside for validation, without using cross-validation.

All the experiments were run on a system that has 24 CPU cores, 252 GB of RAM, and an NVIDIA GTX 1080Ti graphics card.

4.2. Tweets Pre-Processing

Twitter text tends to be messy because people use casual language, emojis, shortened words, hashtags, and sometimes incorrect grammar. Because of this, we use a detailed process to clean the text before we analyze it for sentiment.

The steps in this process are:

• First, we split the tweets into individual words, making sure to remove extra spaces and punctuation.
• Next, we take out any numbers since they don’t help in determining the sentiment.
• We also remove common words like ’a’, ’the’, ’to’, and ’at’ because they don’t carry much meaning in this context.
• Then, we get rid of punctuation marks that don’t add any real meaning to the text.
• After that, we use a process called stemming, which reduces words to their basic form using the Porter stemming method.
• We also pay special attention to words like ’not’, ’never’, and ’do not’, as they can change the meaning and sentiment of a sentence.
• We use a part-of-speech tagging tool called Apache OpenNLP to identify and keep words that are important for sentiment, especially adjectives.
• Finally, we remove mentions of users, URLs, and retweet symbols, but we keep hashtags by removing the ’#’ symbol so the actual words are still useful.

4.2.1. SentiWordNet-Based Polarity Computation

To quantify sentiment strength, SentiWordNet is used to compute positivity and negativity scores for opinion words.

let: Let:

• P be the set of positive opinion words,
• N be the set of negative opinion words,
• $p_i$ be the positivity score of the $i^{\mathrm{th}}$ word,
• $n_i$ be the negativity score of the $i^{\mathrm{th}}$ word.

$$Pos\left(T\right)={\displaystyle \frac{1}{\left|P\right|}}\sum _{i=1}^{\left|P\right|}{p}_i$$

(1)

The Negative sccore is caluculated as:

$$Neg\left(T\right)=-{\displaystyle \frac{1}{\left|N\right|}}\sum _{i=1}^{\left|N\right|}{n}_i$$

(2)

where $p_i\in [0,1]$ and $n_i\in [-1,0]$. The negativity score is multiplied by $-1$ to map its value into the range $[0,1]$.

4.3. Feature Extraction Using Term Weighting TF-IDF

Feature extraction is a key step in sentiment classification, where text is turned into numbers that machine learning models can use. In this study, the Term Frequency–Inverse Document Frequency (TF-IDF) method is used to represent tweets.

Term weighting gives more importance to words that appear often in a document but not too much in other documents. TF-IDF is based on three things: how often a word appears in a single document, how common it is across all documents, and a normalization factor.

Let:

${\mathrm{TF}}_{i,j}$ denote the frequency of term j in document i,

${\mathrm{DF}}_j$ denote the number of documents containing termj

N denote the total number of documents.

The TF-IDF weight is computed as:

$$TF-ID{F}_{i,j}=T{F}_{i,j}\times log\left({\displaystyle \frac{N}{D{F}_j}}\right)$$

(3)

This approach focuses on highlighting words that carry emotional or meaningful information while removing common words that don’t add much value. However, TF-IDF creates a large number of features, which is why people often use feature selection methods.

4.4. Feature Selection Algorithm: Egret Swarm Optimization Algorithm (ESOA)

The Egret Swarm Optimization Algorithm (ESOA) is a problem-solving method inspired by the behavior of Snowy Egrets and Great Egrets when they look for food. Snowy Egrets usually stay still and wait for prey, while Great Egrets move around actively to find their food. The ESOA operates through three main stages:

• Sit-and-Wait Strategy
• Aggressive Strategy
• Discriminant Condition

In this algorithm, each group of egrets has three members:

• Egret A, which uses a method similar to finding the best path
• Egret B, which randomly explores different areas
• Egret C, which helps bring the group together to focus on the best solution

4.4.1. Sit & Wait Strategy

Let the position of the th i th Egret be:

$$x_i\in {\mathbb{R}}^n$$

(4)

The observation function is defined as:

$$y_i=A\left(x_i\right)$$

(5)

The estimation equation is:

$${\widehat{y}}_i=w_i^T{x}_i$$

(6)

The estimation error is calculated as:

$$e_i=y_i-{\widehat{y}}_i$$

(7)

The descent direction is obtained as:

$$d_i=-\nabla e_i$$

(8)

Egrets update weights adaptively as:

$$w_i^{t+1}=\beta w_i^t+(1-\beta )e_i{x}_i$$

(9)

where = $\mathsf{\beta }$=0.9.

The position update for Egret A is:

$$x_i^{t+1}=x_i^t+ste{p}_a·d_i$$

(10)

4.4.2. Aggressive Strategy

Egret B explores randomly:

$$x_b^{t+1}=x_b^t+r_1$$

(11)

Egret C exploits promising regions using encircling behavior:

$$x_c^{t+1}=x_{best}-r_2·|x_{best}-x_c|$$

(12)

where $r_1,r_2\in [0,0.5]$.

4.4.3. Discriminant Condition

A discriminant rule selects the best solution:

$$x_{best}=\left\{\begin{array}{cc}{x}_{new},\hfill & \mathrm{if}f\left(x_{new}\right)<f\left(x_i\right)\hfill \\ x_i,\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(13)

With a probability of 0.3, inferior solutions may be accepted to avoid premature convergence.

4.4.4. ESOA Pseudocode

Algorithm 1 Egret Swarm Optimization Algorithm (ESOA)

Initialize population X Evaluate fitness of each egret while termination condition not met do     Apply Sit-and-Wait strategy     Apply Aggressive search strategy     Apply Discriminant condition     Update best solution end while Return $x_{best}$

4.4.5. Computational Complexity

The sit-and-wait strategy needs 11n + 9 floating-point operations, while aggressive strategies need 2n + 2. The discriminant condition needs n operations. Therefore, ESOA has a computational complexity of:

$$O(k\times n)$$

(14)

where k is population size and n is feature dimension.

4.5. Classification Using Deep Learning Model

4.5.1. Seq2Seq Model and Attention Mechanism

Given an input tweet sequence:

$$X=(x_1,x_2,\dots ,x_n)$$

(15)

The encoder hidden state is computed as:

$$h_t=f(h_{t-1},x_t)$$

(16)

The decoder probability is:

$$P\left(y_t\right|y_{<t},X)=\mathrm{softmax}\left(W{s}_t\right)$$

(17)

The attention score is computed as:

$${\alpha }_{t,i}={\displaystyle \frac{exp\left(g(s_t,h_i)\right)}{{\sum }_{j}exp\left(g(s_t,h_j)\right)}}$$

(18)

4.5.2. Multi-Head Attention Mechanism

Encoder states are projected into K semantic spaces:

$$h_i^{\left(k\right)}=W_p^{\left(k\right)}{h}_i$$

(19)

Context vectors are computed as:

$$c_t^{\left(k\right)}=\sum _i{\alpha }_{t,i}^{\left(k\right)}{h}_i^{\left(k\right)}$$

(20)

Final context vector:

$$c_t=\sum _{k=1}^K{r}_k{c}_t^{\left(k\right)}$$

(21)

4.5.3. Penalty Term

To avoid redundancy among attention heads, a penalty term is introduced:

$$L_{penalty}={∥A{A}^T-I∥}_F^2$$

(22)

The overall loss function is:

$$L=L_{seq}+\lambda L_{penalty}$$

(23)

4.5.4. Hyperparameter Optimization

There are a lot of hyperparameters set up for this neural network that need tweaking, and there are some choices that could be debated, like how the data was scaled, if shortcut connections were included, if LSTM layers were used, if two attention layers were added, and if any hidden decisions were missed. Hyperparameter optimization, also called DMO, is done to better understand the model and possibly make it work better. The runtime was improved to 0.01170 seconds using just 41 epochs of training.

4.5.5. Dwarf Mongoose Optimization (DMO) Procedure

Dwarf Mongoose Optimization (DMO) is a stochastic population-based metaheuristic optimization algorithm inspired by the social organization and foraging behavior of the dwarf mongoose (Helogale). In nature, dwarf mongooses exhibit a unique combination of individual food searching and group-based foraging strategies, which enables them to locate optimal food resources efficiently. These animals are semi-nomadic in nature and tend to construct temporary sleeping mounds close to resource-rich regions. The behavioral patterns of dwarf mongooses are statistically modeled in DMO to solve complex optimization problems effectively.

Like other population-based optimization techniques, DMO begins with a random initialization phase, followed by iterative refinement through exploration (diversification) and exploitation (intensification) phases. The convergence towards a global optimum is driven by the collective interactions among the mongooses.

Initialization Phase

The population of candidate solutions is defined as:

$$X=\left[\begin{array}{cccc}{x}_{1,1}& x_{1,2}& \dots & x_{1,d}\\ x_{2,1}& x_{2,2}& \dots & x_{2,d}\\ ⋮& ⋮& \ddots & ⋮\\ x_{n,1}& x_{n,2}& \dots & x_{n,d}\end{array}\right]$$

(24)

where n represents the population size, and each candidate solution $x_i$ is a vector of dimension d.

Each candidate solution is initialized randomly within predefined lower and upper bounds using:

$$x_{i,j}=Va{r}_{min}+rand\left(\right)\times (Va{r}_{max}-Va{r}_{min})$$

(25)

where $x_{i,j}$ represents the position of the $i^{\mathrm{th}}$ mongoose in the $j^{\mathrm{th}}$ dimension, ${\mathrm{Var}}_{min}$ and ${\mathrm{Var}}_{max}$ denote the minimum and maximum permissible values, and $\mathrm{rand}\left(\right)$ is a uniformly distributed random number in the interval $[0,1]$.

At each step, the solution with the highest fitness value is chosen as the best solution so far.

DMO has two main stages. The first is called Exploitation, where individual mongooses search for better solutions in areas they are already familiar with. The second is called Exploration, where they search randomly to find new food sources or new resting spots. These two stages are handled by three groups within the DMO: the alpha group, the scout group, and the babysitter group.

Unifrnd is a random number that is spread out evenly within a specific range.

VarMin and VarMax are the minimum and maximum values allowed for the problem, and VarSize is the number of data points in the dataset.

In each round, the best solution found so far is retained as the best one.

The DMO, like other metaheuristic algorithms, has two main parts: exploitation, where each individual mongoose focuses on improving its current solution, also known as intensification, and exploration, where a random search is done to find a new food source or a new resting spot, also referred to as diversification.

These two processes are managed by the DMO’s three main social groups: the alpha group, the scout group, and the babysitters.

Choosing the leading female ($\mathsf{\alpha }$) in a family unit uses Equation (26):

$$\alpha ={\displaystyle \frac{f_i}{{\sum }_{i=1}^n{f}_i}}$$

(26)

The value represented by n minus b corresponds to the number of mongooses in the top tier. You can think of b as the number of nannies, and peep as the female alpha who calls out directions to the rest of the group. The amount of food available, shown in Equation (27), is what finally determines how the sleeping mound looks.

$$X_{t+1}=X_t+\varphi ·peep$$

(27)

After each iteration, the position update is calculated using Equation (28), where $\varphi$ is a random number that is evenly spread out between $[-1,1]$ and $[1,1]$:

$$s{m}_t={\displaystyle \frac{f_{t+1}-f_t}{max\left(\left|f_{t+1}\right|,\left|f_t\right|\right)}}$$

(28)

When a new sleeping mound is discovered, the average is calculated with the help of Eq. (29):

$$\phi ={\displaystyle \frac{{\sum }_{i=1}^{n}s{m}_i}{n}}$$

(29)

Once the conditions for changing babysitters are met, the search for the next sleeping spot starts. During this search, they look for a new place that has a different food source. Since mongooses never go back to the same sleeping spot, the group has to keep looking all the time for new ones. In DMO, mongooses often search for food and check out new spots at the same time. This is because the more they travel to look for food, the better chance they have of finding a good new place to sleep.

The scouting update rule is defined by:

$$X_{t+1}=\left\{\begin{array}{cc}{X}_t-CF·\varphi ·rand·(X_t-\overline{M}),\hfill & \mathrm{if}{\phi }_{t+1}>{\phi }_t\hfill \\ X_t+CF·\varphi ·rand·(X_t-\overline{M}),\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(30)

Where $\mathrm{rand}\in [0,1]$, and

$$CF={\left(1-{\displaystyle \frac{iter}{\mathrm{MaxIter}}}\right)}^{{\displaystyle \frac{iter}{\mathrm{MaxIter}}}}$$

is a parameter that decreases linearly and controls how much the group of mongooses moves around together.

The vector

$$\overline{M}=\sum _{i=1}^n{X}_i{s}_i{m}_i$$

represents the force that makes the mongooses move to a new nesting place for the night.

The group responsible for taking care of the young stays with them, while the other mongooses search for a sleeping spot and food. They wait to start foraging or scouting until the babysitting condition in Equation (30) is met, so these members are not counted as part of the total group searching.

4.6. Performance Analysis of the Proposed Model Without Feature Selection

This section evaluates how effectively the proposed model performs without using the ESOA-based feature selection approach. To ensure a fair comparison, several other models such as LSTM, ICSD, DPG-LSTM, ELM-RNN, BERT-BiLSTM, and SBERT are tested on the same dataset. The average results from these models are then presented. Additionally, traditional machine learning models and standard deep learning models are also included in the testing. The complete comparison results are summarized in Table 3. The results in Table 3 and shown in Figure 4 indicate that the MHAM model works well even without using the ESOA-based feature selection method. Several traditional machine learning models, such as SVM and RF, and different types of neural networks like Auto-encoder, CNN, and LSTM, along with more advanced deep learning models like ICSD, DPG-LSTM, ELM-RNN, BERT-BiLSTM, and SBERT, were tested for comparison.

Among the models tested, SVM and RF achieved accuracy scores of 89.02% and 89.27%, which is quite good. The CNN and LSTM models performed better, with accuracy levels of 92.82% and 90.21% respectively. Transformer-based models, such as BERT-BiLSTM and SBERT, performed even better, reaching accuracy levels of 94.80% and 94.92% respectively.

However, the MHAM model performed the best, achieving an accuracy of 95.20%. It also had better recall, precision, and F-score values compared to all other models. These results show that the MHAM model is very effective for classifying the sentiment of tweets, even without using feature selection, which means it has strong discrimination abilities.

Table 3. Comparative Performance Analysis of the Proposed Model without ESOA.

Methodology	Accuracy (%)	Recall (%)	Precision (%)	F-Score (%)
SVM	89.02	90.91	88.34	79.82
RF	89.27	89.27	87.72	82.20
Auto-encoder	90.38	88.90	90.86	74.02
CNN	92.82	90.67	93.09	88.88
LSTM	90.21	92.91	93.20	83.49
ICSD	91.20	94.04	95.90	89.90
DPG-LSTM	92.03	95.92	95.55	91.91
ELM-RNN	94.44	95.64	95.30	92.04
BERT-BiLSTM	94.80	93.22	94.02	93.30
SBERT	94.92	94.30	96.90	94.43
Proposed MHAM	95.20	95.78	97.32	95.65

Table 3. Comparative Performance Analysis of the Proposed Model without ESOA.

Methodology	Accuracy (%)	Recall (%)	Precision (%)	F-Score (%)
SVM	89.02	90.91	88.34	79.82
RF	89.27	89.27	87.72	82.20
Auto-encoder	90.38	88.90	90.86	74.02
CNN	92.82	90.67	93.09	88.88
LSTM	90.21	92.91	93.20	83.49
ICSD	91.20	94.04	95.90	89.90
DPG-LSTM	92.03	95.92	95.55	91.91
ELM-RNN	94.44	95.64	95.30	92.04
BERT-BiLSTM	94.80	93.22	94.02	93.30
SBERT	94.92	94.30	96.90	94.43
Proposed MHAM	95.20	95.78	97.32	95.65

4.7. Performance Analysis of the Proposed Model with Feature Selection

This part looks at how well the suggested model works when it uses the ESOA-based feature selection method. Like before, the new approach is tested against several traditional machine learning models, deep learning structures, and transformer-based techniques. The results, which include accuracy, recall, precision, and F-score, are shown in Table 4.

Table 4. Comparative Performance Analysis of the Proposed Model with ESOA.

Methodology	Accuracy (%)	Recall (%)	Precision (%)	F-Score (%)
SVM	86.12	89.94	87.38	81.85
RF	87.29	88.28	88.78	86.27
Auto-encoder	91.88	87.95	91.87	84.42
CNN	93.42	92.68	92.12	86.89
LSTM	88.93	91.95	92.23	89.99
ICSD	89.23	92.72	92.45	89.95
DPG-LSTM	91.90	94.78	93.60	90.87
ELM-RNN	94.45	94.82	94.15	91.97
BERT-BiLSTM	95.46	95.55	95.80	92.54
SBERT	95.88	95.27	95.92	93.80
Proposed MHAM-DMO	97.28	97.98	98.77	96.29

Table 4. Comparative Performance Analysis of the Proposed Model with ESOA.

Methodology	Accuracy (%)	Recall (%)	Precision (%)	F-Score (%)
SVM	86.12	89.94	87.38	81.85
RF	87.29	88.28	88.78	86.27
Auto-encoder	91.88	87.95	91.87	84.42
CNN	93.42	92.68	92.12	86.89
LSTM	88.93	91.95	92.23	89.99
ICSD	89.23	92.72	92.45	89.95
DPG-LSTM	91.90	94.78	93.60	90.87
ELM-RNN	94.45	94.82	94.15	91.97
BERT-BiLSTM	95.46	95.55	95.80	92.54
SBERT	95.88	95.27	95.92	93.80
Proposed MHAM-DMO	97.28	97.98	98.77	96.29

The results in Table 4 and shown in Figure 5 show how using the ESOA-based feature selection method improves the classification results. The test includes traditional machine learning models like SVM and RF, neural networks such as Auto-encoder, CNN, and LSTM, hybrid deep learning models like ICSD, DPG-LSTM, and ELM-RNN, and transformer-based models like BERT-BiLSTM and SBERT.

SVM and RF, which are baseline models, have accuracy scores of 86.12% and 87.29% respectively.

Deep learning models perform better, with CNN reaching 93.42% accuracy. Transformer-based models do even better, with BERT-BiLSTM and SBERT achieving 95.46% and 95.88% accuracy respectively.

The MHAM-DMO model we propose performs the best among all tested methods, with an accuracy of 97.28%, and also has better recall, precision, and F-score.

These results show that combining ESOA-based feature selection greatly improves the model’s ability to distinguish between different classes, surpassing all other methods tested.

5. Conclusion

Sentiment analysis is a way to figure out what people are feeling about a piece of text. It helps us understand their opinions, emotions, and beliefs about something. Both training and testing the models were done using the Twitter dataset. Tweets are short messages with a limit of 140 characters and often include emoticons and emojis. In this study, SentiWordNet is used to group tweets into three categories: positive, negative, or neutral. Using TF-IDF in the feature extraction process led to a big improvement in the accuracy of sentiment analysis. In this research, we present the Egret Swarm Optimization Method, a new meta-heuristic algorithm for feature selection that is inspired by the foraging behaviors of two types of egrets: the Great Egret and the Snowy Egret. The three main ideas behind ESOA are the Snowy Egret’s passive method, the Great Egret’s aggressive approach, and a discriminant illness. To help the decoder give a clear response to an input question, we introduce a DMO with a multi-head attention mechanism. The experimental results show that the proposed model performs better than previous models in terms of classification accuracy. For example, while earlier models had accuracy between 89% and 95% and precision between 88% and 95%, the new model achieved 97.28% accuracy and 98.77% precision. However, there is still a need for a hybrid deep learning model to further improve the study. References