CP-IMOS: A Cross-Platform Imbalance-Aware Methodology for Sentiment Classification of MOOC Reviews in IT Education

MOOC review datasets offer a practical way to study course strengths, learner frustrations, and platform-level issues. However, sentiment models built from these data can give a distorted picture when positive ratings dominate the corpus. This paper presents CP-IMOS, a Cross-Platform Imbalance-Aware MOOC Sentiment Classification methodology, using public review data from Coursera, Udemy, and Stepik. The Coursera file contained 1,454,711 review rows; after cleaning and exact duplicate removal, 498,401 review records remained. Keyword filtering of course identifiers and titles then yielded 258,828 IT-related Coursera reviews from 215 course identifiers. For Udemy, a 200,000-comment Kaggle sample was processed with Course_info.csv to support second-platform validation. After cleaning, 46,377 comments from Development and IT & Software courses were retained from 15,836 course identifiers. The Stepik Russian-language corpus was included to test multilingual ingestion, duplicate removal, and Cyrillic-character validation during retrieval, but it was excluded from supervised sentiment training because it did not provide rating or sentiment labels. CP-IMOS maps rating-derived labels into five classes that correspond to ratings 1 through 5 and adds the Platform Sentiment Imbalance Index (PSII) to measure majority-class dominance before model interpretation. The classifier stage uses TF-IDF features with transparent machine-learning components. In the Coursera analysis sample, SGD Logistic Regression obtained the highest macro-F1 (0.383), while Multinomial Naive Bayes produced higher accuracy (0.759) but a weaker macro-F1 (0.250). In the Udemy validation set, SGD Logistic Regression also produced the highest macro-F1 (0.430). The findings show that cross-platform MOOC review data are strongly skewed toward positive ratings, although Udemy contains more non-positive ratings than Coursera. The main contribution is a reproducible retrieval-to-classification workflow that uses PSII, macro-F1, per-class metrics, and platform-specific interpretation before sentiment results are used in IT education analytics.