An adaptive segmentation method based on attribute weights appropriate to the decision context

When segmenting mixed-type data in corporate information systems, geometric similarity of objects and significant differences for management indicators may not be the same. Along with the selection of the algorithm and the number of clusters, it is also necessary to determine the contribution of the attributes to the distance. In this work, we developed an adaptive segmentation protocol that combines attribute weighting, decision signaling, and configuration selection. For each number of clusters, attribute weights are updated based on relevance to current segments, informativeness to the decision variable, and stability of the distribution over time. Numeric and categorical characters are weighted at the initial attribute level. K-Means, medoid, and agglomerative configurations are then evaluated through geometric separation, decision relevance, reselection stability, and segment balance. In order to limit the computational cost, the stability is calculated by first relearning the weights on the fixed geometry and then on the selected candidates. Experiments covered three open UCI datasets and five synthetic scenarios. The adaptive method recovered the number of true clusters in 23 out of 25 cases in synthetic experiments and reduced the average contribution of noise attributes. According to Bank Marketing, the decision correlation increased over K-Means, but no such advantage was observed in the other real sets. The overall result of K-Means for synthetic structure recovery was higher. The temporal component gave a positive mean change in a separate control experiment; this effect was not uniform across realizations. The contribution of the work is to generate a geometry relevant to the decision context and to openly evaluate its benefits and limits. Results relate to selected data and evaluation protocol; A direct comparison with W-k-means and sparse K-means was not performed.