Cautious weighted random forests

Random forest is an efficient and accurate classification model, which makes decisions by aggregating a set of trees, either voting or averaging class posterior probability estimates. However, tree outputs may be unreliable in presence scarce data. The imprecise Dirichlet model (IDM) provides workaround, replacing point estimates with interval-valued ones. This paper investigates new aggregation method based on the theory belief functions to combine such intervals, resulting cautious random classifier. In particular, we propose strategy for computing weights minimization convex cost function, takes both determinacy accuracy into account it possible adjust level cautiousness model. proposed evaluated 25 UCI datasets demonstrated more adaptive noise training data achieve better compromise between informativeness cautiousness. • A classifier forests proposed. Tree are automatically learned from using function. tuned single parameter. Extensive experiments demonstrate interests method.