Transferring Knowledge by Prior Feature Sampling

The paper presents a novel method for transfer learning through prior variable sampling. A set of problems defined in the same feature space with similar dependencies of target on features is considered. We suggest a method for learning a decision tree ensemble on each of the problems by prior estimation of variable importance on other problems in the set and using it for regularizing model learning for a small amount of training samples. The method is tested on several simulated and real datasets. In particular, we apply our method for a set of time series classification (TSC) problems. Our analysis demonstrates an intriguing result: a model trained on several TSC problems can learn a new problem with high accuracy from a low number of samples.