A general-purpose knowledge finding method for efficient knowledge finding by selectively sampling only data in large information amounts from a database. Learning means 104 causes a lower-order learning algorithm, inputted via an input unit 107, to perform learning on plural partial samples generated by sampling from data stored in a high-speed main memory 120, to obtain plural hypotheses. Data selection means 105 uses the hypotheses to estimate information amounts of respective candidate data read from a large-capacity data storage device 130, and additionally stores only data in large information amounts into the high-speed main memory 120. A control unit 106 repeats the processing a predetermined number of times, and stores obtained final hypotheses. A prediction unit 102 predicts a label value of unknown-labeled data inputted into the input unit 107 by the final hypotheses, and an output unit 101 outputs the predicted value.