Using Genetic Programming to Generate and Evaluate Features
Summary
The article introduces gplearn’s Symbolic Transformer as a supervised way to create features by combining input variables into mathematical expressions. It describes the underlying genetic algorithm: candidate formulas are generated, scored for fitness, and iteratively selected, recombined, or mutated. This offers an alternative to manually proposing ratios and other combinations when domain knowledge is limited, while aiming to discard combinations with weak relationships to the target.
The author proposes assessing generated features through their correlation with the label, collinearity with one another, model importance, and effect on model performance, using a Home Credit default-risk project as the intended setting. The text does not report the results of that assessment, so it establishes a workflow and rationale rather than evidence that the method improves predictions. It also notes that the article concerns an older version of the platform and may not apply to its current tools; supervised feature generation requires careful validation to avoid overfitting.
Key ideas
- Symbolic Transformer uses genetic programming to discover mathematical combinations of input features.
- Candidate expressions are iteratively selected and varied according to a supervised fitness measure.
- Generated features should be assessed for target association, collinearity, model importance, and validation performance.
- The article proposes an evaluation but does not provide its results, and its platform guidance is outdated.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.