The paper presents the bottom-up subspace cluster- ing approach and discusses some drawbacks of clustering methods in broad analysis of complex, high-dimensional data. The aim of this paper is to propose some improve- ments of existing bottom-up subspace clustering methods. A novel grid-based bottom-up subspace clustering algo- rithm is presented which is able to handle both numerical and nominal attributes and requires only one single param- eter. Clusters are represented as hyper-rectangles in sub- spaces of attributes and can be easily interpreted by a human as decision rules. The results of experiments con- ducted on artificial and real data sets are included.