A Meta-Learning Method to Learn from Small Datasets
洪秀芳, Meafen Hung · 2005
The nature of survival suggests that learning from fewer examples is often important, but machine learning has not yet learned well from small datasets. In contrast, human beings often learn well from very small examples, even if the number of potentially features is large. To do so, they successfully use previously learned concepts to improve performance of the current task. This thesis is an approach to develop a classifier for a small dataset using other datasets and their learning result. Meta-learning aims at how learning systems can increase performance through experience, but researches in characterization of datasets are still lack. We propose a measurement to select a support dataset, a process to acquire prior knowledge and apply it to new learning problem to improve performance. A proper characterization of datasets to match naive Bayes algorithm is key to the research.