Decoding brain activity from non-invasive motor imagery electroencephalography (MI-EEG) signals is vital for brain-computer interface (BCI). Many studies overlook the feature-classifier interaction, impacting decoding performance. To tackle this issue, we propose a decoding framework for MI-EEG signals that uses multi-view weighted features. Our approach begins by employing a multi-view feature fusion mechanism to capture both global and local features from raw MI-EEG signals. Following feature extraction and fusion, we utilize the Expectation-Maximization (EM) algorithm to partition the samples to distinct soft subspaces. The subset information is subsequently used to classify uncategorized MI-EEG samples. Within this framework, features that are more relevant to the subsequent classification model are assigned greater weights. Results from public benchmark datasets demonstrate that, compared to commonly used models, our method achieves superior classification results.