Finding Discrimination Rules Using the Cascade Model

抄録

The cascade model is proposed as a new method for finding discrimination rules. The method is based on the itemset lattice used in association rule mining. Items arising from explanation attributes are employed to construct the lattice, while the distribution of the class values of the supporting instances is attached to each node in the lattice. The gini-index is used to indicate the potential of the node. The introduction of node potential suggests an image of cascades, with nodes as lakes and links as waterfalls, in which an instance corresponds to a drop of water falling down the cascades. The power of a link is analogous to the hydroelectric power of a waterfall. The problem of finding discrimination rules is then formulated as a search for powerful waterfalls in the cascades. the model has been implemented as the DISCAS system, using a new pruning criterion to avoid combinatorial explosion of the number of nodes in the lattice. Algorithms for node generation and rule extraction are described. Application to House voting records shows that the resulting rules are simple and comprehensible.