Sentiment analysis has become a crucial tool for understanding customer opinions and enhancing e-commerce platforms. Shathik et al. [11] provide a comprehensive review of sentiment analysis applications using machine learning techniques, discussing key challenges and improvements. A significant contribution is the categorization of machine learning based sentimental analysis methods, emphasizing their applications, strengths, and limitations. Chaturvedi et al. [12] examine sentiment analysis in business intelligence, emphasizing its role in decision-making. The paper introduces sentiment analysis as a key tool to process unstructured data, offering practical and computational insights into its application for improving decision-making processes in competitive business environments. More recently, Huang et al. [13] reviewed current machine learning techniques for sentiment analysis in e-commerce, highlighting their effectiveness. The study systematically reviewed 54 experimental papers on sentiment analysis in e-commerce, results show that Support Vector Machine (SVM) and Naïve Bayes (NB) were the most widely used machine learning models consistently shown better results than others, whereas Recurrent Neural Networks (RNN) and its variation like Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) were dominant in deep learning models. The study also finds that deep learning techniques like BERT, Convolutional Neural Networks (CNN), RNN, and LSTM classification are frequently used in recent studies. Alonso et al. [14] explore sentiment analysis for fake news detection, comparing models like Random Forest (RF), NB, SVM, LSTM, and CNN. It discusses how fake news often contains emotionally charged language to manipulate readers and increase engagement. Various sentiment analysis techniques are evaluated, and the result shows that hybrid models outperform standalone classifiers. Kawade et al. [15] apply sentiment analysis to Uri attack using R for textual data processing. It highlights how sentiment analysis can mine public emotions from social media platforms like Twitter and process them for better understanding of public opinions on events.

As shown in Figure 1 , the sentiment analysis process generally involves stages such as data collection, preprocessing, feature extraction, model training, and evaluation.

Building on the foundational overview of sentiment analysis techniques, several studies have implemented traditional machine learning techniques that have been widely applied for sentiment analysis in e-commerce. Yi et al. [16] apply multinomial NB, K-Nearest Neighbors (KNN), and SVM combined with collaborative filtering to recommend products and shops based on customer reviews, achieving improved recommendation accuracy measured by Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). By integrating collaborative filtering and product-product similarity techniques, the study provides a framework for identifying customer preferences and enhancing e-commerce experiences. Deniz et al. [17] introduce a multi-label classification approach using Binary Relevance, Multi-label KNN, and One-vs-Rest classifiers with Logistic Regression (LR), Stochastic Gradient Descent (SGD), and XGBoost. The best performing model was One-vs-Rest-XGBoost with 94.2% accuracy and 0.89 F1-score, indicating a low misclassification rate. Among feature extraction techniques, Term Frequency-Inverse Document Frequency (TF-IDF) performed better than word embeddings and transformers. Saura et al. [18] utilize SVM and Latent Dirichlet Allocation (LDA) to extract key indicators for startup business success, evaluated through topic coherence scores and classification accuracy. LDA identifies words and separates each word into a different document, randomly identifies distribution of the topics in a sample and then selects the main topics found in the sample. The identified topics are further analyzed using NVivo software to extract meaningful insight. Alnasrawi et al. [19] enhance sentiment classification by integrating text network features with various machine learning models, including SVM, RF, LR, NB, Decision Trees (DT), Neural Networks (NN), Gradient Boosting (GB), and AdaBoost. Among all methods, NN outperformed other models with 88.7% accuracy and 0.85 F1-score, and incorporating network-based features improved classification performance by 5% - 8% across precision, recall and F1-score metrics, particularly for NN and ensemble methods.

Despite the effectiveness of traditional machine learning models, their limitations in handling complex linguistic patterns have led researchers to explore deep learning techniques for more nuanced sentiment classification. Deep learning techniques have been widely adopted for sentiment analysis due to their ability to capture complex language patterns. Mehta et al. [20] propose a CNN-based sentiment classification model enhanced with Word2Vec and TF-IDF to analyze customer reviews in the apparel sector using deep learning models, achieving 92.3% accuracy and 0.91 F1-score. The study does not evaluate other deep learning models, which could improve sentiment classification. Lin [21] employs Natural Language Processing (NLP) methods to analyze correlation between e-commerce customer review features and product recommendations. The researcher examined several different models, like Lasso Regression, Ridge Regression, Linear Kernel SVM, Radial Basis Function (RBF) Kernel SVM, RF, XGBoost, and LightGBM. The best result was achieved by LightGBM with 0.847 R-squared value and the correlation between review ratings and product recommendations was found to be strong. Kundra et al. [22] compare traditional machine learning models such as NB, SVM, RF, and AdaBoost with deep learning based RNN and GRU. The research results show that the best-performing model is RNN-GRU with 89.5% accuracy and 0.87 F1-score for unstructured data, but SVM achieved 91.2% accuracy for structured data analysis and performed best in both structured and unstructured data analysis, demonstrating that RNN captures the nuances of unstructured and context-rich information while traditional models excel with structured inputs.

To capitalize on the strength of both traditional and deep learning approaches, recent studies have proposed hybrid models that aim to achieve higher accuracy and adaptability in sentiment analysis. Hybrid models have been developed to enhance sentiment analysis accuracy by combining traditional machine learning and deep learning methods. Zhao et al. [23] introduce a new optimized machine learning algorithm called the Local Search Improvised Bat Algorithm-based Elman Neural Network (LSIBA-ENN) for sentiment analysis, a new term weighting technique called Log Term Frequency-based Modified Inverse Class Frequency (LTF-MICF) for sentiment classification, and a Hybrid Mutation-based Earthworm Algorithm (HMEWA) for feature selection. The LSIBA-ENN model achieved 95.2% accuracy and 0.94 F1-score, outperforming traditional classifiers, and the LTF-MICF weighting scheme improved performance by 3% - 5% compared to TF-IDF and Word2Vec-based approaches. There are several studies combining the CNN and LSTM methods to improve the accuracy of sentiment analysis. Vatambeti et al. [24] employ a ConvBiLSTM model, integrating CNN and (Bi-LSTM) with an Elephant Herd Optimization (EHO) algorithm. EHO is a swarm-based optimization method used to optimize the weight of the Bi-LSTM, deep learning model parameters. The Optimized model achieved 93.8% accuracy with 0.92 precision and 0.91 recall on Twitter food service data. Behera et al. [25] present Co-LSTM, a combination of CNN and LSTM, for sentiment classification in social media data. The proposed model achieved 94.5% accuracy and 0.93 F1-score, outperforming standalone CNN with 87.2% accuracy and LSTM with 89.7% accuracy. The model is highly adaptable in examining big social data and is free from any particular domain, unlike conventional machine learning approaches. Co-LSTM performed particularly well on long and complex texts by effectively capturing dependencies between words. Ray et al. [26] also integrated CNN with rule-based methods to improve aspect-level sentiment analysis. The model achieved 88.4% accuracy and 0.86 F1-score, incorporating hierarchical clustering and dependency parsing to refine aspect categorization and improve accuracy. Jain et al. [27] implemented a CNN-LSTM model for consumer sentiment analysis and the model demonstrated 91.7% accuracy with 0.89 precision and 0.88 recall, higher than standalone deep learning and machine learning models. Furthermore, Priyadarshini et al. [28] develop a hybrid LSTM-CNN model with Grid Search Optimization. The model achieved the highest performance with 96.1% accuracy, 0.95 precision, 0.94 recall, and 0.94 F1-score. The model is compared with baseline algorithms, including CNN (88.2% accuracy), KNN (82.1% accuracy), LSTM (90.3% accuracy), CNN-LSTM (92.8% accuracy) and LSTM-CNN (93.4% accuracy), and LSTM-CNN-Grid Search achieved higher performance as the research incorporated grid search as a hyperparameter optimization technique to minimize pre-defined losses.

As illustrated in Figure 2 , the machine learning workflow for sentiment analysis often integrates various stages, including data input, feature extraction, model selection, and performance evaluation.

As sentiment analysis techniques continue to evolve, there is growing interest in applying these models in real-time systems to provide immediate feedback and actionable insights. Several studies focus on real-time sentiment analysis and AI-driven emotion visualization. Jabbar et al. [29] implement a SVM-based real-time sentiment analysis system within a mobile application using Python and Flask. The system achieved 87.3% accuracy with an average response time of 0.8 seconds per query. When users submit a review, the application processes the text through the model, which predicts the sentiment polarity and provides immediate feedback. Li et al. [30] propose an explanation framework for AI-driven sentiment analysis, leveraging Bi-LSTM with attention mechanisms to visualize emotion-triggering words. The proposed framework achieved 91.2% accuracy and 0.88 F1-score while providing 78% user satisfaction in explainability surveys. The framework emphasizes considering the cause and trigger of emotions as the explanation for the deep learning-based emotion analysis. The proposed method has more functionality, usability, reliability, and performance compared to SHapley Additive exPlanations (SHAP), which is a model interpretation package developed by Python and Local Interpretable Model-agnostic Explanations (LIME) which is an explainable framework for black box deep learning models using a local explainable model to explain each separate prediction, with 15% higher user comprehension scores and 20% faster explanation generation time.

This literature review highlights the diverse methodologies utilized in sentiment analysis, including traditional machine learning, deep learning, hybrid approaches, and real-time applications, reflecting the dynamic evolution of sentiment analysis in e-commerce. Together, these developments from traditional machine learning to real-time deep learning systems highlight the dynamic landscape of sentiment analysis in e-commerce, underscoring its critical role in enhancing user engagement, personalization, and business intelligence.