Autoencoder-based One-class Classification Technique for Event Prediction

This paper proposes an autoencoder-based one-class classification technique to predict a specific event such as the occurrence of a fire in a specific building. Basically, a binary classification system that uses machine learning to identify fire-risk buildings requires 'positive' fire data and 'negative' non-fire data. However, the fire-risk building data that can be actually obtained have a single class data that includes only the data of the occurrence of the fire and does not include the data of the 'non-occurrence'. In this situation, PU (Positive-Unlabeled) learning which uses 'unlabeled' data can be an effective way of generating the fire prediction model. The autoencoder generates new features from the unlabeled data, with which a predictive model for predicting the fire-risk buildings is built through PU learning.