ROCsearch in a Wider Context — A ROC-Guided Search Strategy for Subgroup Discovery and Beyond

ROCsearch is a ROC-based beam search variant, initially developed for Subgroup Discovery (SD). In ordinary beam search, on each search level, a fixed number of best-scoring candidates are selected to generate candidates for the next search level. This fixed number, the beam width, is typically hard to set, and its setting strongly influences the outcome of the mining process. In ROCsearch, however, on each search level, the beam width is set dynamically by analyzing the intermediate results in ROC space. Thus, setting the beam width parameter is taken out of the domain expert’s hands, lowering the threshold for using Subgroup Discovery. Also, ROCsearch automatically adapts its search behavior to the properties and resulting search landscape of the dataset at hand. In Subgroup Discovery, ROCsearch has been shown to be an order of magnitude more efficient than traditional beam search, while its results are equivalent and on large datasets even better than traditional beam search results. However, ROCsearch has not been investigated beyond SD, while it should be readily applicable in machine learning and data mining outside of the SD subfield. As work in progress, we propose this wider outlook for ROCsearch.