AutoGOAL is an open-source AutoML framework that finds machine learning pipelines for a given problem by searching a huge space of algorithm combinations. That search is powerful but starts from zero on every new problem. My contribution attacks that: a meta-learning subsystem that lets the framework learn from its own past experiments.
The idea is to treat machine learning itself as a learning problem. A set of meta-features (size, number of classes, statistical properties) describes every dataset, and every past experiment records how well each pipeline performed. From that experience, a meta-learner predicts which pipelines look promising for a dataset it has never seen, so the search starts from likely winners instead of random guesses, cutting training time and improving the solutions found.
The work spans the full pipeline: automated extraction of meta-features from datasets, processing of past pipeline results into training data, two meta-learner implementations (nearest-neighbors and a neural network with different ranking strategies), and the experimental infrastructure to test the approach across many datasets.
Machine learning research at the University of Havana, building on the AutoGOAL framework created by the university’s AI research group (over 100 commits on the research branch).