The AGARICUS-LEPIOTA dataset

Figure 14: Classification results for the mixtures of trees and other models: (a) On the AGARICUS-LEPIOTA data set; the MT has $m=12$, and the MF has $m=30$. (b) On the NURSERY data set; the MT has $m=30$, the MF has $m=70$. TANB and NB are the tree augmented naive Bayes and the naive Bayes classifiers respectively. The plots show the average and standard deviation test set error rate over 5 trials.

The AGARICUS-LEPIOTA data [Blake, Merz 1998] comprises 8124 examples, each specifying the 22 discrete attributes of a species of mushroom in the Agaricus and Lepiota families and classifying it as edible or poisonous. The arities of the variables range from 2 to 12. We created a test set of $N_{test}=1124$ examples and a training set of $N_{train}=7000$ examples. Of the latter, 800 examples were kept aside to select $m$ and the rest were used for training. No smoothing was used. The classification results on the test set are presented in figure 14(a). As the figure suggests, this is a relatively easy classification problem, where seeing enough examples guarantees perfect performance (achieved by the TANB). The MT (with $m=12$) achieves nearly optimal performance, making one mistake in one of the 5 trials. The MF and naive Bayes models follow about 0.5% behind.