case 2 of Hunt’s method, a test based on a single attribute is chosen for expandingthe current node. The choice of an attribute is normally based on the entropygains [Qui93] of the attributes. The entropy of an attribute, calculated fromclass distribution information, depicts the classification power of the attributeby itself. The best attribute is selected as a test for the node expansion.Highly parallel algorithms for constructing classification decision trees aredesirable for dealing with large data sets in reasonable amount of time. Classi-fication decision tree construction algorithms have natural concurrency, as oncea node is generated, all of its children in the classification tree can be generatedconcurrently. Furthermore, the computation for generating successors of a classi-fication tree node can also be decomposed by performing data decomposition onthe training data. Nevertheless, parallelization of the algorithms for constructionthe classification tree is challenging for the following reasons. First, the shape ofthe tree is highly irregular and is determined only at runtime. Furthermore, theamount of work associated with each node also varies, and is data dependent.Hence any static allocation scheme is likely to suffer from major load imbalance.Second, even though the successors of a node can be processed concurrently,they all use the training data associated with the parent node. If this data isdynamically partitioned and allocated to different processors that perform com-
putation for different nodes, then there is a high cost for data movements. If the
data is not partitioned appropriately, then performance can be bad due to the
loss of locality.
Several parallel formulations of classification decision tree have been pro-
posed recently [Pea94,GAR96,SAM96,CDG+97,Kuf97,JKK98,SHKS99]. In this
section, we present two basic parallel formulations for the classification decision
tree construction and a hybrid scheme that combines good features of both of
these approaches described in [SHKS99]. Most of other parallel algorithms are
similar in nature to these two basic algorithms, and their characteristics can be
explained using these two basic algorithms. For these parallel formulations, we
focus our presentation for discrete attributes only. The handling of continuous
attributes is discussed separately. In all parallel formulations, we assume that N
đang được dịch, vui lòng đợi..