情報処理学会 第80回全国大会講演要旨

Both fitting and cross validating models from small data sets is challenging due to the difficulty of avoiding bias. Our goal is to predict the failure rate as well identifying the failures key factors of a certain type of electronic component from part of their specifications and their actually recorded failure rate, while having available to us around 100 failure records spread over around 30 components plus 100 non failing components. Toward that goal, we made a generalized linear model from our data set, which was able to predict their actual failure rate from their specs within a 95% confidence interval. To make this possible we used an expert's knowledge in this component type to alleviate the lack of data.