The paper presents a methodology with a hierarchical approach to ranking construction. In the process, different discrete forms of the input data are explored, and advantage is taken of the wisdom of the crowd offered by the random forest algorithm. The experiments were conducted in the stylometry domain on a task of authorship attribution. The results obtained validated the proposed framework, as it led to many cases with a noticeable improvement in accuracy for reduced sets with feature selection relying on the proposed method of weighting attributes.