Research Of Conformal Predictor Based On Distance Metric Learning And Clustering

Conformal predictor is an region prediction machine learning algorithm, which provide a predefined confidence level for each prediction. Being different from the traditional machine learning algorithms whose goals are to get a high point predictive accuracy by offline learning, conformal predictor is better candidate for high risk applications.Conformal predictor is online by nature, and it needs to storage and access the original examples frequently when calculates the randomness test value. It is a load for some large data applications to bear. What’s more, conformal predictor has also the imbalance problems, the prediction risk of small class data can’t be controlled accurately. Currently, researches about the online computational efficiency and imbalance problem of conformal predictor are less. This paper is mainly dedicated to these two problems.For online computational efficiency problem, we propose the "Distance metric learning-based conformal predictor" algorithm. The results show that:the algorithm can guarantee the prediction efficiency and improve the online computational efficiency.For imbalance problem, we propose "Cluster sampling-based conformal predictor" and "Cluster partition-based conformal predictor" on dataset processing level and method improving level respectively. The results show that:these two methods can reduce the influence of imbalance data effectively, the prediction risk of small class data can be controlled accurately.