Algorithm Portfolio
No single learning algorithm wins on every dataset, so rather than hunt for one champion, an AutoML system can carry a small squad of specialists and let the data decide which one plays. That squad is an algorithm portfolio.
What a portfolio is
A portfolio is a curated set of algorithms — or of complete algorithm-plus-hyperparameter configurations — chosen because their strengths do not overlap. Diversity is the design goal: five near-identical gradient-boosting configurations are worth less than a set holding one boosted-tree model, one linear model, a nearest-neighbour baseline and a neural network, because that set covers far more kinds of data. Portfolios are built offline, by evaluating candidates across a large benchmark of datasets and greedily adding whichever configuration most improves results on the datasets the current set handles worst.
How AutoML systems use one
- As a warm start: Auto-sklearn 2.0 runs a precomputed portfolio before any Bayesian optimisation, so search begins from configurations already known to work broadly.
- As the whole search: AutoGluon's presets train a fixed portfolio and stack the results, skipping expensive tuning altogether.
- As a selector: given meta-features of a new dataset, a meta-model picks which portfolio member to run — the per-instance strategy SATzilla made famous in solver competitions.
Why it matters
A portfolio turns expensive search into cheap lookup. Under a tight time budget, running ten configurations known to be complementary beats exploring a vast space blindly, and the run becomes reproducible: the same portfolio on the same data yields the same shortlist every time.
An algorithm portfolio is a small, deliberately diverse set of algorithms or model configurations that an AutoML system keeps on hand because their strengths complement one another. Instead of searching a huge configuration space from scratch, the system runs or selects from the portfolio, which is assembled offline by greedily adding members that cover the datasets the existing set handles worst. Portfolios deliver strong results under tight time budgets and make AutoML runs cheaper and reproducible.
Think of a repair van. The driver doesn't carry five versions of the same wrench — they carry a spanner, a screwdriver, a drill and a hammer, so that whatever the job turns out to be, something in the van will fit.
An algorithm portfolio is that van for machine learning. Different learning methods suit different kinds of data, and nobody knows in advance which one suits yours. So automated systems keep a small, varied collection of methods that have already proved themselves, and try those first instead of hunting through millions of possibilities. It saves a great deal of time and still lands on something that works well.