Hi Kong,
The approach for using "pcares" function for classification seems reasonable, but there are some things to consider when using "pcares" for the use case of video classification, especially if you're getting low accuracy after classification.
[residuals,reconstructed] = pcares(X,ndim)
The 'reconstructed' data represents an approximation of your original data using the top 'ndim' principal components. If too few components are used (e.g., 5), some significant variance and important features might be lost. For this purpose you might need to increase the number of components to retain enough information.
Spatial and Temporal Structure: Also flattening the reconstructed data into 1D vectors may discard spatial and temporal relationships, which are important in video data. For 2D or 3D data (like video), it's often better to preserve these relationships, as they are critical for accurate predictions.
You can experiment by increasing the number of components to capture more variance. You could also consider using "pca" (without the residuals) to retain the most important features. Additionally, for better preservation of spatial features, you can consider using 'convolutional neural networks': https://www.mathworks.com/help/deeplearning/ug/introduction-to-convolutional-neural-networks.html
Hope this helps!
