Induction of supermodels
Hendrik Blockeel · Lirias · 2002
It is rarely the case in machine learning and data mining projects that only a single inductive process is run. Often, experiments are performed with a number of different algorithms, different settings for a single algorithm, dierent feature sets, and so on. While most research focuses on algorithms that induce a single model, given this practice it would be useful to look for algorithms that induce a kind of modelgenerator or "super"-model, from which multiple concrete models can efficiently be generated (this is a kind of partial evaluation). Thus one can avoid executing computationally heavy induction processes many times in very similar contexts. The process is similar to partial evaluation. This paper discusses some possibilities and challenges for research on supermodels.