The NURSERY dataset

This data set contains 12,958 entries [Blake, Merz 1998], consisting of 8 discrete attributes and one class variable taking 4 values¹. The data were randomly separated into a training set of size $N_{train}=11,000$ and a test set of size $N_{test}=1958$. In the case of MTs and MFs the former data set was further partitioned into 9000 examples used for training the candidate models and 2000 examples used to select the optimal $m$. The TANB and naive Bayes models were trained on all the 11,000 examples. No smoothing was used since the training set was large. The classification results are shown in figure 14(b).

Figure: Comparison of classification performance of the MT and other models on the SPLICE data set when $N^{train}=2000,\;N^{test}=1175$. Tree represents a mixture of trees with $m=1$, MT is a mixture of trees with $m=3$. KBNN is the Knowledge based neural net, NN is a neural net.

Figure: Comparison of classification performance of the mixture of trees and other models trained on small subsets of the SPLICE data set. The models tested by DELVE are, from left to right: 1-nearest neighbor, CART, HME (hierarchical mixture of experts)-ensemble learning, HME-early stopping, HME-grown, K-nearest neighbors, LLS (linear least squares), LLS-ensemble learning, ME (mixture of experts)-ensemble learning, ME-early stopping. TANB is the Tree Augmented Naive Bayes classifier, NB is the Naive Bayes classifier, and Tree is the single tree classifier.