Real-World-Data always follows natural patterns with various undesirable situations. Chinese language is more prone to susceptibility and influence for its rich semantic information. When confronted with severely imbalanced distributions, the existing methods always tend to favor the majority class while neglecting the minority class. However, the performance of the minority class cannot be ignored, and it is particularly crucial not to overlook the performance of the minority class when dealing with situations involving offensive text filtering. In this work, we mainly focus on three aspects, (i): We developed a Chinese Offensive Message Text Dataset, which includes seven categories: Normal, Spam, Advertisement, Fraud, Violent, Pornography, Politics, exhibiting a significant imbalance between Normal and Abnormal. (ii): We proposed a feature space manipulation which combines samples optimization and spatial position optimization for the majority / minority and a hybrid feature prediction module which integrates multi-level feature fusion for comprehensive utilization of sample information. (iii): Experimental results show that our method achieves significant improvement on the offensive classes in our dataset.