The article introduces a new method called local Fisher discriminant analysis (LFDA) for reducing the complexity of data with class labels. Traditional methods struggle with data that forms multiple clusters within a single class, but LFDA considers the local structure of the data to handle this issue. The researchers show that LFDA can also be adapted for non-linear data reduction using the kernel trick.