A Comparison of Hardware and Software in Sequence Rule Evolution
Magnus Lie Hetland, Pål Sætrom · 2003
Sequence rule mining is an important problem in the field of data mining. Many algorithms have been devised that are based on counting candidate rules and excluding those with low support. Recently, the techniques of heuristic search, and evolutionary algorithms in particular, have been applied to various data mining problems, including sequence mining. In our previous work we have used specialized hardware to make mining certain rule formats feasible. In this paper we compare the performance of this hardware with realistic software alternatives. We show that these software alternatives give acceptable, although significantly slower, running times for restricted rule formats. We also demonstrate that the increased expressiveness available with the hardware rule formats does not necessarily have a great impact on predictive power, and may be more useful as a way of tailoring rule formats to specific tasks.