METHOD FOR CUSTOMER FEEDBACK SENTIMENT ANALYSIS BASED ON HYBRID CLASSIFICATION MODELS FOR FEEDBACK AUTOMATION IN IT PROJECT MANAGEMENT

Abstract

The article presents a comprehensive approach to automating customer feedback processing within IT project management systems. Effective monitoring of user satisfaction is a critical factor for ensuring software product quality (Quality Assurance) and managing the product backlog. However, manual analysis of large volumes of textual data is labor-intensive, slow, and prone to subjectivity. This study proposes a method for customer review sentiment analysis based on hybrid classification models, which combines the computational efficiency of classical machine learning algorithms with the generalization capabilities of deep learning.

The foundation of the proposed method is a sequential hybrid architecture that integrates a Support Vector Machine (SVM) classifier with a Radial Basis Function (RBF) kernel and a deep neural network (Dense Neural Network). The method utilizes a stacking technique where probabilistic estimates generated by the SVM are used as additional input features for the neural network. This approach ensures effective error correction and increases the overall system confidence. Additionally, the feature extraction stage employs TF-IDF vectorization with the inclusion of bigrams, allowing the model to consider the local context of the text and improve the recognition of complex linguistic constructs. The system is implemented as a modular architecture, ensuring flexibility for integration into existing project management tools (e.g., Jira, Trello dashboards) and scalability for distributed processing.

Experimental studies conducted on a balanced dataset of customer reviews demonstrated the high effectiveness of the proposed hybrid approach compared to individual models. The optimized SVM achieved an accuracy of 86.2%, the standalone neural network reached 88.9%, while the final hybrid model achieved the highest accuracy of 89.0% with an average F1-Score of 0.890. The complementarity of the components ensured a consistent advantage of the hybrid model, particularly in classifying ambiguous and mixed-sentiment texts.

The obtained results confirm the feasibility of using the developed method for feedback automation, enabling IT project managers to rapidly obtain objective product quality metrics, identify critical issues in real-time, and make informed management decisions to improve the software development lifecycle.