Hans‐J. Lenz

Data warehouses are used to store large amounts of data which is often used for On-Line Analytical Processing (OLAP). Short response times are essential for on-line decision support. Common approaches to reach this goal in read-mostly environments are the precomputation of materialized views and the use of index structures. This paper focuses on the use of index structures for supporting fast access to data. The performance of index structures depends on many different parameters. Here, we focus on a set of nine parameters. Two approaches are presented to support the decision making process which index structure should be applied. The first approach is based on classification trees. The second approach uses an aggregation and scatter diagram method. Both approaches are applied to four distinct index structures: a tree-based index structure without aggregated data, a tree-based index structure with aggregated data and two bitmap index structures. This paper presents results of the comparison with both approaches.

Research and industry has tackled the object identification problem of data integration in many different ways. This paper presents a framework, that allows the evaluation of competing approaches. To this end, complexity measures and data characteristics are introduced, which reflect the hardness of a given object identification problem. All characteristics can be estimated by use of simple SQL queries and simple calculations. Following the principle of benchmark definitions we specify a test framework. It consists of a test database and its characteristics, quality criteria, and a test specification. Adequate measures needed for the correctness criterion of the benchmark are given. A running example of the Berlin Online Apartment-Advertisements database (BOA) illustrates the approach. The BOA-database is freely available at www.wiwiss.fu-berlin.de/lenz/boa/.

We evaluate the rough set and the association rule method with respect to their performance and the quality of the produced rules. It is shown that despite their different approaches, both methods are based on the same principle and, consequently, must generate identical rules. However, they differ strongly with respect to performance Subsequently an optimized association rule procedure is presented which unifies the advantages of both methods.