This study proposes a knowledge graph-based big data analysis model for course quality evaluation, aiming to address issues in online education course evaluations such as semantic bias, grammatical deficiencies, vocabulary limitations, false evaluations, information distortion, and imbalanced evaluation categories. The model incorporates three innovative strategies: topic modeling sentiment scoring, sentiment label correction, and comprehensive course quality evaluation. It extracts topics from massive course data, integrates sentiment labels through knowledge graph embedding, employs optimal classifier sequences for predicting course indicators, and utilizes game theory calculations to obtain global importance values. Experimental results demonstrate that our model outperforms existing methods: compared to XGBoost, accuracy and macro-F1 scores increased by 3.21% and 4.86%, respectively, on the China University MOOC dataset; compared to Naive Bayes, they improved by 4.31% and 3.35% on the Coursera dataset. The model also performed well on imbalanced NetEase Cloud Classroom and Udemy datasets, confirming its generalizability and robustness. This study provides strong technical support for improving online teaching quality and educational decision-making.