Refining diagnostic knowledge extracted from interferon therapy by graph-based induction

A machine learning technique called graph-based induction (GBI) extracts patterns from graph-structured data by stepwise pair expansion. GBI has been extended to 1) beam-wise GBI (B-GBI) by incorporating a beam search to improve its search capability, and 2) decision tree graph-based induction (DT-GBI) to construct a decision tree for graph-structured data. We applied B-GBI and DT-GBI to analyze the effectiveness of interferon therapy in the hepatitis dataset provided by Chiba University Hospital. Descriptive patterns were extracted by B-GBI and discriminative ones by DT-GBI using only the time sequence data of blood inspection and urinalysis. The discriminative patterns extracted by DT-GBI tend to be included in only relatively small number of patients and thus too specific. Thus, we tried to extract patterns, which are both discriminative and descriptive by B-GBI. Furthermore, since there are exceptional situations (patients) with the extracted patterns, these patterns are further utilized to extract refined knowledge from the dataset. The preliminary results are reported with some of extracted patterns.