Feature selection for sentiment analysis using hybrid multiobjective evolutionary algorithm

1.1 Research objectives

The study seeks to address the challenges associated with data preparation for sentiment analysis, particularly in dealing with irrelevancy and redundancy in textual data. The study aims to combine filter based and wrapper based methods for feature selection. The research objectives of this study revolve around recognizing the importance of text classification, addressing data preparation challenges, exploring feature selection for text data, and advocating for more efficient sentiment analysis models.

1.2 Research contributions

The primary contributions of our research can be summarized as follows:

Pioneering the fusion of Information Gain (IG) with Binary Multi-Objective Grey Wolf Optimizer (BMOGWO) for sentiment analysis.

The remaining paper is organized as follows. Section 2 describes the recent studies on feature selection. Section 3 provides the proposed feature selection model. Section 4 discusses the experimental setup and results. Section 5 presents limitations of this study. Section 6 presents the conclusion and future work.

2.1 Mainstream methods of sentiment analysis

Sentiment analysis has three mainstream methods as discussed in the following:

Rule-Based Systems: Rule-based systems employ predefined linguistic rules to identify sentiment expressions, offering transparency but potential limitations in adaptability and flexibility [14, 15]. These methods are not popular as they cannot be scaled to large datasets.

Lexicon-based Approaches: These methods use predefined sentiment dictionaries to assess sentiment in text [16]. These methods lack the context information for analyzing sentiment of text.

Machine Learning Techniques: Machine learning models like Support Vector Machines, Naive Bayes, and deep learning models such as BERT are widely used for sentiment analysis [17, 18]. They excel at capturing text patterns but often require extensive labeled data.

In this study we have used machine learning based methods for sentiment analysis. Feature selection plays a pivotal role in improving performance of machine learning models. We have used feature selection in this study, it is discussed in next section.

2.2 Information Gain (IG) and BMOGWO

Feature selection has two mainstream methods, filter based approach and wrapper based approach. Filter based approach involved the steps of feature scoring, feature ranking and feature selection. Wrapper based methods generate different feature subsets, evaluate model based on different feature subsets, and then select feature subset based on best performance.

In this study we have used Information Gain (IG) [7] and Binary Multi-Objective Grey Wolf Optimizer (BMOGWO) [19, 20] for feature selection. IG is a important metric for feature selection and text classification. It is a filter based feature selection approach. Its key features include its ability to identify the most informative features for classification tasks, contributing to better model performance. We have chosen Information Gain (IG) for feature selection due to its proficiency in reducing uncertainty in the target variable, a key aspect in semantic contexts. IG’s capability to highlight the most informative features is crucial for effective binary classification, ensuring focus on pertinent data aspects. Furthermore, IG’s alignment with BMOGWO’s objectives enhances the optimization process, ensuring a balance between model complexity and performance.

It is defined in detail in Section 3.2. BMOGWO is an optimization algorithm. Its features include binary nature, use of a sigmoid transfer function, and competitive results feature selection problems. It is a wrapper based approach for feature selection. It is discussed in detail in Section 3.3.

2.3 Related work on feature selection for text classification

Recent studies, such as the work by Abiodun et al. [21], offers invaluable insights into feature selection methods for text classification. Their systematic review of over 200 articles underscores the promise of metaheuristics and hybridized versions to optimize feature selection, advocating for the use of hyper-heuristics to enhance classification accuracy. Their comprehensive assessment of feature selection methods serves to clarify the intricate details of accurate method implementation. Deng et al. [22] conducted a detailed review of feature selection techniques in the context of sentiment analysis. Their work delves into the advantages and drawbacks of these methods while clarifying the most suitable classifiers to accompany them. Additionally, they explore areas like multi-label feature selection, online feature selection, and stability measures for feature selection methods.

Gokalp et al. [23] contributed an algorithm based on an iterated greedy approach for feature selection. Utilizing pre-calculated scores from filter methods and the Multinomial Naïve Bayes classifier, they conducted experiments on various datasets, including public sentiment and Amazon product reviews, demonstrating the superiority of their algorithm over existing alternatives. Deniz et al. [11] introduced a groundbreaking hybrid feature selection method that addressed the problem with a two-objective approach. They employed Information Gain Filtering (IGF) as a filter method, followed by NSGA-II as a wrapper method. Their experiments, employing text representation techniques like Bag-of-Words (BoW) and Glove, showcased the remarkable performance of the hybrid method, IGF+NSGA-II, outshining its individual components.

Banka H. Dara [24] proposed a binarized particle swarm optimization (HDBPSO) algorithm, leveraging hamming distance to reduce dimensionality in feature selection. Their comparative experiments underscored the algorithm’s effectiveness in discovering feature subsets with competitive classification accuracy. Shang et al. [25] introduced a binary particle swarm optimization (FS-BPSO) algorithm that employed fitness sum proportionate selection. By enhancing velocity updating and evaluating individual features, it addressed the limitations of standard BPSO. Experiments on consumer review datasets validated its ability to select high-quality feature subsets.

Shunmugapriya P and Kanmani S [26] combined ant and bee colony optimization to create the AC-ABC Hybrid for feature selection, mitigating weaknesses in both algorithms. Their novel approach demonstrated improved global search efficiency and avoidance of stagnation issues. Liu et al. [27] tackled imbalanced and high-dimensional datasets using multi-objective ant colony optimization. Employing the Bootstrap method, they compared their model with six state-of-the-art algorithms, revealing its superior performance. Tawhid and Dsouza [28] proposed a hybrid algorithm called HBBEPSO, leveraging the BAT algorithm for exploration and PSO for exploitation. Experiments on 20 UCI repository datasets demonstrated its efficacy in identifying optimal feature subsets. Mafarja and Mirjalili [29] presented a hybrid method combining Whale Optimization Algorithm (WOA) with simulated annealing to enhance feature selection. Their method improved classification performance and was competitive with three wrapper-based methods on 18 UCI repository datasets. Al-Tashi et al. [12] introduced BMOGWO, a binary version of the Multi-Objective Grey Wolf Optimizer (MOGWO) using a sigmoid transfer function for feature selection. Comparative experiments against MOGWO with a tanh transfer function, NSGA-II, and MOPSO demonstrated BMOGWO’s competitive results with reduced computational cost.

The collective body of research in feature selection methods and hybrid algorithms provides a rich foundation for advancing the field of sentiment analysis and text classification, bridging the gap between theoretical frameworks and practical implementation.

2.4 Existing research gaps and motivations for this research

Despite advances, there are persistent research gaps. Notably, the need for more interpretable and efficient sentiment analysis models that optimize feature selection. This manuscript aims to bridge these gaps by proposing a novel approach integrating IG and BMOGWO in sentiment analysis. This research is motivated by the dual objective of addressing these existing gaps and enhancing the practicality of sentiment analysis methods. Information Gain (IG) and Multiobjective Grey Wolf Optimizer (MOGWO) offer distinct advantages for feature selection in sentiment analysis. IG is computationally efficient, model-independent, interpretable, reduces dimensionality, and doesn’t require labeled data. MOGWO is flexible for optimization, handles multiple objectives, combines global and local search, adapts to problem spaces, and enhances model performance through optimal feature subset selection. When used together, IG and MOGWO provide a well-rounded approach for efficient and effective feature selection in sentiment analysis, addressing various challenges and objectives. The ultimate goal is to facilitate informed decision-making across diverse applications, from marketing to customer feedback analysis, and social media monitoring. This research contributes to the evolution of sentiment analysis methods and their real-world utility.

This is the first step in any sentiment analysis algorithm. The preprocessing step can make a model or break a model. It is a crucial step in a classification problem [30]. This step removes the features that do not contribute to the classification problem, providing meaningful subset of features. This, as a result, reduces computation time of the training process by reducing the high dimensionality of data. We present a four-fold processing phase:

The preprocessing step serves as the foundational stage in any sentiment analysis algorithm. Its significance cannot be overstated; it has the potential to either elevate or undermine the model’s performance. This step involves the removal of features that do not contribute to the classification problem, resulting in a more meaningful subset of features [30]. Consequently, this reduction in feature dimensionality has the dual benefit of enhancing the computational efficiency of the training process and simplifying the data structure. The preprocessing phase can be comprehensively broken down into four distinct steps:

3.2 Filter-based feature selection

This section delves into the initial feature selection methods applied in this study. Our first approach involves the use of Information Gain (IG) as a filter-based method. IG assesses each feature’s information content, quantifying how much valuable information each feature carries. For a single feature, denoted as F, Information Gain is calculated as follows:

IG ( N , F ) = Entropy ( N ) ∑ u ϵ U | N u | N Entropy ( N u )(1) where N is a variable contains all features of the data, F is a single feature, U represents a set of unique values for the particular feature F. Nu represents a subset of N which contains the values where F is u. We calculate a subset S’s entropy as: Entropy ( S ) = ∑ c ϵ C p c log 2 p c(2) where C represents the set of all classes and pc gives the ratio of the instances belonging to the cth class over all instances in S.

After calculating Information Gain of all features, we calculate a threshold value and remove the features whose IG values are less than that threshold. The threshold value is the median value of the IG values of all features. We sorted Information Gain (IG) values into quartiles and set the feature selection threshold at the median, corresponding to the second quartile (50th percentile). This approach ensured we selected only the top 50 percent of features by IG, balancing model efficiency and effectiveness. By avoiding a lower threshold, like the first quartile, we prevented the exclusion of crucial discriminative features, and by not setting it too high, like the third quartile, we avoided cluttering the model with low-impact features. This step is to keep features having higher predictive power. This process is named as Information Gain Filtering (IGF).

This section introduces the second feature selection method, which leverages the Binary Multi-Objective Grey Wolf Optimizer (MOGWO) algorithm. This approach is categorized as a wrapper-based feature selection technique. Before delving into the algorithm, it’s important to establish a clear understanding of what constitutes a multi-objective optimization problem. Additionally, we’ll explore how to assess the results using Pareto fronts as a key evaluation metric.

3.3.1 Multi-objective optimization and Pareto front

Feature selection can be formulated as a multi-objective optimization task aiming to optimize two objectives: reducing the features set size and maximizing classification accuracy. We can define this multi-objective optimization problem as: minimize f 1 ( x ) = | M | , M ⊆ N(3) maximize f 2 ( x ) = accuracy ( M )(4) where N denotes the whole features set and M is the subset of N i.e., the selected feature subset. Accuracy gives us the percentage of correctly predicted data points from all the data points.

In an optimization problem involving multiple objectives, we can get a solution set containing all the possible solutions to the problem. Within this set, one solution may excel in terms of one objective, while another surpasses it in a different objective. To visually illustrate this concept within the context of our specific problem, we present sample solutions in Fig. 2.

OPEN IN VIEWER

Fig. 2

Sample solutions for the multi-objective optimization problem.

The objective in our problem is to simultaneously minimize the number of features and maximize classification accuracy. Figure 2 serves as an illustrative example of solutions for this multi-objective problem, specifically feature selection. The green solutions (S1 to S4) in the figure represent the Pareto front, consisting of non-dominated solutions. These solutions outperform others in both objectives.

In contrast, the blue solutions are classified as dominated solutions. This means that at least one other solution within the Pareto front dominates them in both objectives. For instance, S3 surpasses S5 in both criteria, as it exhibits a reduced feature count and higher accuracy. This can be mathematically expressed as:

f 1 ( S 3 ) < f 1 ( S 5 )(5) f 2 ( S 3 ) > f 2 ( S 5 )(6)

In these equations, S3 dominates solution S5, demonstrating its superiority in both feature reduction and classification accuracy.

he Pareto front solutions hold a unique position in that they are not dominated by any other solution, including those within the Pareto front itself. Let’s illustrate this with solutions S1 and S2 as examples. S1 excels over S2 in terms of feature count, boasting a smaller number of features. Meanwhile, S2 outperforms S1 in accuracy, displaying a higher level of precision. As they each surpass the other in different objectives, they are deemed non-dominated by one another. Thus, the solutions on the Pareto front represent the best possible outcomes in the context of the defined objectives. Our primary aim in a multi-objective problem is to identify and characterize this Pareto front in accordance with the specified objectives.

3.3.2 Binary Multi-Objective Grey Wolf Optimizer (BMOGWO)

For wrapper based feature selection, we apply the Binary Multi-Objective Grey Wolf Optimizer (BMOGWO). Each solution is represented by a wolf as [f₁, f₂, . . . , f _k ] where k denotes the overall count of features and f _k is the dataset’s feature number k. The total number of features gives us the length of a solution. This is a binary optimization problem, implying that a feature can be selected (1) or not (0). A sample solution is shown in Fig. 3. In the figure, we can see that the solution has selected features 2, 4, 5 and 7. So, the number of features (first objective value) of this solution is 4. The other features i.e., 1, 3 and 6 are removed and classifier is trained on the selected features, which will the accuracy (second objective value) for this solution.

OPEN IN VIEWER

Fig. 3

A sample solution.

In our BMOGWO implementation, the algorithm proceeds as follows: Initially, we generate a random population. Then each search agent is evaluated based on the objective functions. We calculated fitness values by focusing on two main criteria: the number of selected features and the classification accuracy using the Multinomial Naive Bayes (MNB) classifier. We determined the number of selected features by counting the features included in each model variant. For classification accuracy, we assessed the effectiveness of each model variant specifically with the MNB classifier. This approach helped us effectively balance a lean model with fewer features and optimal predictive performance using MNB, which is essential in multi-objective optimization tasks. The non-dominated solutions are retrieved from the wolves to initialize the archive. Then, three best solutions are selected from the archive namely alpha, beta and delta. The selection of these leaders and the determination of the archive being full are done by a controller. The leaders are selected by a leader selection strategy. The three leaders guide other wolves to globally optimal search areas, in hopes to find global best or near global best solutions. The least crowded regions are selected by leader selection technique and one of its non-dominated solutions is chosen as alpha, beta or delta. Each hypercube has a probability in the roulette-wheel selection method to select such solutions. This probability is defined below: P i = b S i(7) where b is a constant having value greater than 1 and S is the number of solutions in the pareto front in the ith segment.

Following the initialization phase, the iterations commence. In each iteration, the positions of the wolves are updated. Subsequently, the continuous values obtained during this update are converted into binary form. This binarization process is facilitated through the use of a sigmoid activation function, implemented as follows: x d t + 1 = { 1 if sigmoid ( x 1 + x 2 + x 3 3 ) ≥ rand 0 otherwise(8) where x d t + 1 is the modified binary grey wolf position in d dimension at t iteration, rand indicates a random number taken from a uniformly distributed range ϵ [0, 1], and sigmoid(a) is calculated as follows: sigmoid ( a ) = 1 1 + e - 10 ( x - 0.5 )(9)x₁,x₂ and x₃ are defined by the following equations: x 1 d = { 1 if ( x α d + bstep α d ) ≥ 1 0 otherwise(10) x 1 d = { 1 if ( x β d + bstep β d ) ≥ 1 0 otherwise(11) x 1 d = { 1 if ( x δ d + bstep δ d ) ≥ 1 0 otherwise(12) where bstep for alpha, beta and delta wolves is defined as: bstep α , β , δ d = { 1 if cstep α , β , δ d ≥ rand 0 otherwise(13) where rand shows a random number taken from a uniformly distributed range ϵ [0, 1], and cstep α , β , δ d is a step size for d dimension having a continuous value, calculated using a sigmoid function as shown below: cstep α , β , δ d = 1 1 + e - 10 ( A 1 d D α , β , δ d - 0.5 )(14) where A is a coefficient vector, defined as: A → = 2 a → . r 1 → - a →(15) where a → is a vector that linearly reduces over the course of iterations from 2 to 0 and can be formulated as: a → = 2 - t . 2 maxIter(16) where maxIter is the maximum number of iterations. r 1 → is a random vector taken uniformly from

D → = | C → . X → p ( t ) - X → ( t ) |(17) where C → is a coefficient vector defined as: C → = 2 . r 2 →(18) where r 2 → is a random vector taken uniformly from the distribution [0, 1]. X → p ( t ) is vector representing position of the prey, and X → ( t ) is vector representing position of the wolf.

The wolves are again evaluated based on the objective functions and non-dominated solutions are extracted. We update the archive with the obtained non-dominated solutions. We use grid mechanism to remove the extra solution(s) if there is no more space for more solutions. New solutions are added. The three best solutions are determined again. If the stopping criteria is met, we return the archive. The flowchart of the proposed model is shown in Fig. 1 and the pseudo-code of the feature selection algorithm is shown in Algorithm 1.

We utilized the Multinomial Naïve Bayes (MNB) classifier for our experiments, which uses word frequency information for its algorithm. MNB has been proven to be very effective in text classification tasks. MNB works by assuming that the probabilities of features do not depend on the class c (conditional independence). MNB is a naïve classification method that uses the Bayes rule. According to Bayes rule, the probability of a document d belonging to a class c is calculated by: P ( c | d ) = P ( d | c ) P ( c ) P ( d )(19)

4.1 Datasets

We performed our experiments on 6 datasets. The first dataset for the experiments is the Stanford Sentiment Treebank (SST). The next 4 are Amazon product reviews datasets (Watches, Mobile Electronics, Gift card and Digital Video Games. The sixth is the Sentiment140 dataset. We provide characteristics of all the six datasets in Table 1. These datasets are discussed in the following sections.

4.2 Parameter settings

In this work, we set training and test sets of SST as their given labels i.e., 6920 training samples and 1821 test samples. The other five datasets i.e., the four Amazon reviews datasets and the Sentiment140 dataset were randomly partitioned into 75 percent training and 25 percent test sets. We set the population size/number of wolves as 100 and the number of iterations as 200, recognizing that a larger population size disproportionately increases computation time compared to a higher number of iterations. Each additional individual in the population demands more computational resources for evaluation, leading to a significant escalation in overall processing time. Consequently, we balanced the population size to ensure computational efficiency. We determined the number of iterations as 200 based on convergence behavior observed in preliminary runs. This specific count ensured that our BMOGWO had ample scope to effectively explore and exploit the solution space. With 200 iterations, the algorithm was able to converge towards optimal solutions, striking a balance between thorough search and optimization processes and computational feasibility. It provided a sufficient number of cycles for the algorithm to mature in its search without unduly extending the computation time. The detailed parameter settings are shown in Table 3.

We computed fitness values by focusing on two primary criteria: the count of selected features and the classification accuracy using the Multinomial Naive Bayes (MNB) classifier. The number of selected features was determined by tallying the features included in each model variant. For assessing classification accuracy, we gauged the effectiveness of each model variant specifically with the MNB classifier. This approach enabled us to effectively strike a balance and train a model with fewer features while maintaining optimal predictive performance using MNB, a crucial aspect in multi-objective optimization tasks. To establish the feature selection threshold, we utilized the median of Information Gain (IG) values. We organized IG values into quartiles and positioned the feature selection threshold at the median, corresponding to the second quartile (50th percentile). This strategy ensured that we selected only the top 50% of features based on IG, achieving a balance between model efficiency and effectiveness. By avoiding a lower threshold, such as the first quartile, we mitigated the risk of excluding essential discriminative features. Simultaneously, by not setting the threshold too high, like the third quartile, we avoided cluttering the model with low-impact features.

4.4 Results

We compare our model with IG and the very popular evolutionary algorithm NSGA-II [11]. The pareto fronts obtained by IGF+NSGA-II and IGF+BMOGWO on all datasets are shown in Fig. 4. The dataset, the original number of features and the accuracy of MNB on those features without feature selection are specified above each graph. The number of features are represented on the x-axis and the accuracy is shown on the y-axis. We can clearly see from Fig. 4 that BMOGWO performs significantly well than NSGA-II. BMOGWO reduces features size to substantially low and achieves greater accuracy than NSGA-II on all datasets.

OPEN IN VIEWER

Fig. 4

Comparison of Pareto fronts obtained by NSGA-II and BMOGWO on all datasets. The title of the figures show dataset, the original number of features and the accuracy without feature selection.

We conducted a comprehensive comparison between our model, and the widely recognized evolutionary algorithm NSGA-II combine with IG, as described in reference [11]. The Pareto fronts obtained through the IGF+NSGA-II and IGF+BMOGWO approaches across all datasets are visually depicted in Fig. 4. Each graph in the figure is accompanied by information about the dataset, the original number of features, and the accuracy of Multinomial Naive Bayes (MNB) when applied to the features without feature selection.

The x-axis represents the number of features, while the y-axis displays accuracy. Upon closer examination of Fig. 4, it becomes evident that BMOGWO outperforms NSGA-II significantly. BMOGWO achieves a substantial reduction in feature size while simultaneously attaining higher accuracy levels compared to NSGA-II across all datasets.

Figure 5 shows percentage reduction in number of features by both algorithms on all six datasets. It is evident that the proposed model outperforms NSGA-II [11]. Our feature selection model achieves reduction to 90 percent, whereas NSGA-II is able to achieve only 76 percent reduction in number of features.

OPEN IN VIEWER

Fig. 5

Feature reduction on all datasets by NSGA-II [11] and BMOGWO.

Figure 6 shows percentage improvement in accuracy by both algorithms on all six datasets. The figure shows the superiority of the proposed model over NSGA-II. Our feature selection model improves classification accuracy by a maximum of nine percent, whereas NSGA-II is able to achieve a maximum of around six percent boost in accuracy.

OPEN IN VIEWER

Fig. 6

Accuracy improvement on all datasets by NSGA-II [11] and BMOGWO.

Table 4 shows the computation time for both algorithms on all datasets. The results show that the proposed algorithm is fairly efficient in terms of computation time.

To measure the performance of both algorithms statistically, we performed Wilcoxon Rank Sum Test (Mann-Whitney U Test). This non-parametric test was chosen because it doesn’t assume data normality or equal variances, making it suitable for comparing our algorithms’ non-normally distributed accuracy values. The p-values for all datasets are shown in Table 5. All six datasets showed statistically significant differences between the two algorithms, with all p-values <0.05.

These extracted features demonstrate the model’s effectiveness in discerning various emotional tones and intensities in text. By focusing on a select yet impactful set of features, the model achieved a balance between accuracy in sentiment classification and computational efficiency. This targeted approach in feature extraction underlines the importance of selecting quality features over quantity, ensuring that the model remains both accurate and agile.

In our pursuit of feature selection for classification tasks, our developed model has emerged as a noteworthy contender. When compared with NSGA-II, our model consistently outperforms it across all datasets.

The results demonstrate the superior performance of our algorithm. Notably, our model achieved a substantial reduction in the number of features. For instance, in the SST dataset, the feature count was reduced from 15,163 to 2,638. Similarly, for the Amazon product reviews datasets, the feature counts were dramatically reduced from the range of 100,000 to 200,000 down to approximately 2,500 to 5,000. On the Sentiment140 dataset, the feature count was curtailed from 17,852 to 1,870.

One of the standout achievements of our model is its selectivity, retaining as few as only ten percent of the available features. In contrast, NSGA-II conserves a minimum of 23 percent of the features. Furthermore, our model exhibits an accuracy advantage over NSGA-II, with a noteworthy difference of nearly 3 percent. Our model also realized an impressive nine percent increase in accuracy, while the best performance by NSGA-II was an approximate six percent increase in accuracy.

BMOGWO distinguishes itself through its adeptness at maintaining diversity when selecting solutions and employs various techniques for this purpose. Additionally, it effectively manages the solutions within the archive via an archive controller. The leader selection strategy plays a pivotal role in preventing BMOGWO from becoming ensnared in local optima, promoting its exploration of the solution space.

BMOGWO outperforms NSGA-II on both objectives by excelling in two critical aspects inherent to evolutionary algorithms: exploration and exploitation. An added advantage of BMOGWO is its streamlined parameter requirements, enhancing its overall efficiency.

The grid mechanism integrated into BMOGWO surpasses the non-dominated sorting method employed by NSGA-II. BMOGWO’s utilization of the three best solutions from the non-dominated set fosters greater diversity when compared to NSGA-II. Notably, in NSGA-II, the best solutions are updated only after a full iteration is completed, whereas BMOGWO updates its best solutions whenever the position of a search agent is altered. This feature significantly bolsters the exploration aspect of the algorithm, further contributing to its effectiveness.

The superior diversity achieved by BMOGWO, in comparison to NSGA-II, can be attributed to the leader selection strategy embedded within BMOGWO. This strategy involves the selection of three leaders for each search agent during every iteration, and it plays a pivotal role in elevating the algorithm’s performance over NSGA-II.

Furthermore, BMOGWO surpasses NSGA-II in terms of both convergence and diversity when seeking solutions. This high convergence is primarily attributable to the leader selection method, which consistently identifies the three best leaders for each search agent. The algorithm’s ability to maintain diversity in solutions can be traced back to various techniques employed in maintaining the archive and selecting leaders. In particular, BMOGWO’s preference for the least crowded regions when choosing leaders significantly contributes to the preservation of diversity within the solutions. An additional noteworthy advantage of our model is its absence of parameter tuning requirements for BMOGWO, streamlining its utilization and making it an efficient choice for feature selection.

These results show the effectiveness of our approach in achieving feature reduction and enhancing classification accuracy, firmly establishing it as a superior alternative to NSGA-II in feature selection.

The primary implication of the study is the potential for improved sentiment analysis performance. By selecting a subset of relevant features and optimizing classification accuracy, the proposed hybrid algorithm can enhance the overall effectiveness of sentiment analysis models. This improvement can be beneficial for various applications, including recommendation systems, reviews analysis, and public opinion observation. The study highlights the efficiency of the feature selection process by significantly reducing the number of features. This reduction leads to reduced computational requirements. The study suggests new research directions for multi-objective feature selection and hybrid algorithms. The success of this approach may inspire researchers to explore similar techniques in various optimization problems, contributing to the advancement of evolutionary algorithms.

Despite excellent results on improved accuracy and feature set reduction, the proposed method has its limitations. The study primarily uses Multinomial Naive Bayes (MNB) for classification. Evaluating the proposed model with a variety of classifiers would provide a more comprehensive understanding of its generalizability and effectiveness. The study does not explicitly address the scalability of the proposed model. It is essential to assess its performance and efficiency as the size of the dataset and the number of features increase significantly. Imbalanced datasets can pose challenges for sentiment analysis. The study does not explicitly address methods for handling data imbalances, which can impact the model’s effectiveness.