CART (Classification and Regression Trees) predicts outcomes based on binary splits of input variables. It can be used for scenario discovery to identify vulnerable scenarios for policy measures.
When to use
CART [1] is a rule-based method that can be used for scenario discovery to discover relationships between a set of uncertain input variables (predictors) and a target objective.
CART generates binary trees based on the application of an iterative binary partitioning procedure.
The final tree includes a series of nodes and branches, each branch being representative of a scenario.
There are two main requirements [2]:
- Generation of a database representing the results for a target objective from different combinations of predictors' values and based on the use of simulation model (see scenario discovery)
- Definition of a splitting rule for each predictor to divide the tree into branches and construct a fully grown decision tree
How
Application of CART is usually performed with existing software packages (see Tools), specifying the target outcome (e.g. success vs failure of a policy) and the predictor variables to be split (e.g. variables describing states of multiple plausible futures).
The algorithm involves three major steps can be identified:
- A 'splitting' stage in which a decision tree is grown [2], involving:
- the definition of a threshold value for each predictor explaining the target outcome [16]. One example of a splitting method (i.e., heterogeneity criteria) is the Gini index [2]
- the division of the main tree into two child trees split at a node featuring a predictor in a given input combination, with each branch representing the outcome of interest when the predictor is under or above the defined threshold [16]
- the continuation of the procedure with the remaining predictors on the newly formed branches, forming new nodes and branches, until no more splitting is possible [16]
- A 'pruning' stage in which the maximal-sized tree is pruned back to the root and where a split contributing the least to the overall performance of the tree is removed [16]. Remaining nodes are then recombined. The process continues to find the tree with the best predictive power in terms of coverage (i.e., ratio of relevant cases) and density (i.e., fraction of relevant points) [2]. Examples of pruning methods include cost-complexity pruning, weakest link pruning [8].
- User selection among the generated trees (fully grown one and pruned alternatives) of a tree with high coverage and high density [2].