Try to Add Meta Learning to TPOT

[yhw741632153]

Hello, I'm a graduate student working on automated machine learning and I'm currently planning to try adding a meta-learning module to TPOT to warm up the pipeline, I'd like to know if this is feasible? My general idea is as follows: in the meta-learning phase, compute the meta-features of the collected metadata, then let them run in TPOT, and afterwards store the best number of pipeline results, say, 20. Subsequently, in the test data, the corresponding meta-features results of the test data are calculated first, and afterwards, based on the clustering results, the twenty pipelines of the closest neighboring metadata are obtained and added to the pipeline of this test data, and the remaining 80 are randomly generated as an attempt to preheat the TPOT pipeline. This is still only my general idea, the details of which have not yet been implemented. May I ask the authors whether this idea of mine is meaningful or practicable?