- Decide which sets of metafeatures to use
- original, model-based (knn, perceptron, etc), graph-based, time-based, etc
- Decide which datasets to use
- UCI, synthetic, D3M, etc
- Compute all metafeatures for all datasets
- Decide pipeline style (single classifier? fixed structure? dynamic pipelines?)
- Decide on meta-task
- algorithm selection? pairwise comparison? hyper-parameter optimization?
- Run meta-task with various subsets and feature selected metafeatures
- Decide on deep-mfe architecture
- double attention, generative, invertible network generator, double PCA/LDA
- Attach to previous meta-task
- Run meta-task with deep-learned metafeatures
- Compare results
- hand-crafted mfs, dataset2vec,