Un-pausing Mondrian Blocks Sample Clauses

Un-pausing Mondrian Blocks. In the original algorithm, splits are not inserted for nodes where all samples have the same label (see line 5 of Alg. 1). The authors call this “pausing” a Mondrian block, and they do this to make Mondrian forests “comparable” to standard Random Forests. However, from our experiments 1Note that Alg. 1 describes the offline method to build a Mondrian forest from a given data set. For brevity, we omit the description of the actual online method that updates the tree with a single new data sample [18].

Un-pausing Mondrian Blocks. In the original algorithm, splits are not inserted for nodes where all samples have the same label (see line 5 of Alg. 1). The authors call this “pausing” a Mondrian block, and they do this to make Mondrian forests “comparable” to standard Random Forests. However, from our experiments 1Note that Alg. 1 describes the offline method to build a Mondrian forest from a given data set. For brevity, we omit the description of the actual online method that updates the tree with a single new data sample [18]. 19 else 20 ⌧j Z 21 AddToLeaves(vj,T ) we found that this can cause a bad classification performance of the algorithm, especially if many samples from the same class are observed in a row. In that case, a large amount of these samples will be lost, because the trees only store the bounding boxes of the data. Therefore, if then samples from a new class arrive, the inserted split can not take these “lost” samples into account, which often leads to inadequate splits. To mitigate this problem we define a split threshold parameters ✓s, which determines the maximum number of samples with the same label that can be contained in a leaf node. Thus, we modify the conditon in line 5 into