A computing device determines a sparse feature representation for a machine learning model. Landmark observation vectors are randomly selected. Neighbor observation vectors are randomly selected that are less than a predefined distance from a selected landmark observation vector. The observation vectors are projected into a neighborhood subspace defined by principal components computed for the neighbor observation vectors. A distance vector includes a distance value computed between each landmark observation vector and each observation vector of the projected observation vectors. Nearest landmark observation vectors are selected from the landmark observation vectors for each observation vector. A second distance vector that includes a second distance value computed between each observation vector and each landmark observation vector is added to a feature distance matrix, where the second distance value is zero for each landmark observation vector not included in the nearest landmark observation vectors. A model is trained using the feature distance matrix.