S-Shaped Versus V-Shaped Transfer Functions for Salp Swarm Algorithm in Sentiment Analysis Feature Selection Problem
Dinar Ajeng Kristiyanti, Akhmad Hairul Umam · 2025
This paper proposed an advanced method to enhance sentiment analysis (SA) for large-scale datasets. Given the complexity of SA, especially when dealing with high-dimensional data rich in features, effective feature selection (FS) becomes crucial to manage and reduce the data's dimensionality. Conventional machine learning techniques often fall short in this regard, highlighting the need for more sophisticated FS algorithms. In this study, we propose an improved bio-inspired optimization technique known as the Salp Swarm Algorithm (SSA) as a wrapper method for feature selection to enhance sentiment analysis of Indonesian-language tweets. Feature selection is inherently an NP-hard problem, making meta-heuristic algorithms like SSA a more efficient choice than exact methods. However, SSA typically faces challenges in terms of slow convergence, particularly in binary problems and large datasets. To address this, we introduce eight SSA variants, each employing S-shaped and V-shaped transfer functions to map the continuous search space into a discrete one, thus facilitating efficient feature selection. The proposed SSA-based approach is benchmarked against existing metaheuristic algorithms, including both Ant Lion Optimization (ALO) and Particle Swarm Optimization (PSO). Comparative experiments demonstrate that the enhanced SSA method outperforms both PSO and ALO, showing superior performance in terms of classification accuracy and processing time. Specifically, the experimental results reveal that the SSA combined with the V2-shaped transfer function achieved in just 5.16 seconds using the Support Vector Machine (SVM) classifier, with an accuracy of 89.76%. Similarly, the SSA integrated with the S1-shaped transfer function closely follows, completing the processing in 4.38 seconds with an accuracy of 89.43%. These findings indicate that the proposed SSA-based feature selection approach efficiently explores the feature space, selects the most informative features, and significantly improves the classification accuracy for sentiment analysis tasks.