Akihiko Nakase

Parallelization of decision tree algorithm and its performance evaluation

Data mining is a typical application of high performance computing in the business field. An efficient data mining system which can deal with huge amount of data is desired. This paper describes the parallel processing of decision tree which is a typical algorithm for classification of large database. A free software C4.5 is parallelized for SMP machine using thread library. Parallelism in generating a decision tree can be classified into intra-node parallelism and inter-node parallelism. Intra-node parallelism can be further classified into record parallelism, attribute parallelism, and their combination. We have implemented these four kinds of parallelizing methods, and evaluated their effects with four kinds of test data. The result shows that there is a relation between the characteristics of data and the parallelizing methods, and combination of multiple parallelizing methods is the most effective one.

Distributed memory implementation of KLIC

This paper describes external reference management and distributed unification in a distributed implementation of a concurrent logic programming language KL1. This implementation is based on the KLIC system. KLIC has a feature calledgeneric objects that enable easy modification and extension of the system without changes in the core implementation. This distributed implementation is built upon the same core and external references are represented using generic objects. Unification operations are defined as methods of generic objects. Since creation of interprocessor reference loops cannot be avoided, we studied a new unification scheme that can cope with interprocessor reference loops. We built several experimental distributed systems that all demonstrate reasonable efficiency.

Mining WWW access sequence by matrix clustering

Sequence pattern mining is one of the most important methods for mining WWW access log. The Apriori algorithm is well known as a typical algorithm for sequence pattern mining. However, it suffers from inherent difficulties in finding long sequential patterns and in extracting interesting patterns among a huge amount of results. This article proposes a new method for finding generalized sequence pattern by matrix clustering. This method decomposes a sequence into a set of sequence elements, each of which corresponds to an ordered pair of items. Then matrix clustering is applied to extract a cluster of similar sequences. The resulting sequence elements are composed into a generalized sequence. Our method is evaluated with practical WWW access log, which shows that it is practically useful in finding long sequences and in presenting the generalized sequence in a graph.

Implementation and performance evaluation of dynamic scheduling for parallel decision tree generation

This paper proposes a parallel data-mining algorithm and its implementation on a PC cluster. The decision tree is a widely used data-mining algorithm for classifying records in a database. Simple parallelization of decision tree generation is not efficient because of the load imbalance caused by the form of the generated tree. The SPRINT algorithm solves this problem by grouping a set of nodes in the same level of the tree and balancing the load; however, frequent disk access is required when the data size exceeds the memory size. We propose an improved parallel algorithm of SPRINT by incorporating a dynamic scheduling. Dynamic scheduling is effective in reducing the amount of disk access for storing intermediate results; however, it may cause imbalance in data distribution on PEs (Processing Elements). We solved this problem by incorporating data redistribution. The evaluation result shows that our method realizes an improvement in speed of 3.5 times, for the best case, and equal performance even in the worst case, compared with SPRINT. We also discuss how further performance enhancement may be possible by improving the communication performance.

Quiescence Detection in a Distributed KLIC Implementation

Quiescence detection is a fundamental facility for parallel and distributed processing. This paper describes schemes for quiescence detection in a distributed KLIC implementation. KLIC is a portable implementation of concurrent logic programming language KL1. Termination is detected using the weighted throw counting (WTC) scheme. Based on the scheme a scheme for global suspension was invented. The postmortem system built-in predicate which provides meta programming facilities was designed, and its distributed implementation is also presented.