Probability-weighted Extreme Learning Machine for Classification with Uncertain Data

Learning uncertain data is a hot research topic in recent years. Traditional methods for uncertain data depend on data pre-processing. For example, substituting uncertain value with expected value or ignoring uncertain samples. However, preprocessing can introduce false value and ignore uncertainty. In this paper, we focus on classification problem. Specially, we solve the problem of classifier training with uncertain data. Based on a general uncertain data model, we propose a novel probability-weighted extreme learning machine classification algorithm for uncertain data. In the proposed algorithm, due to huge increase of training set arise from uncertainty, training process can be computationally intensive. For improving training efficiency, we optimize the solution by making it independent of the number of uncertain instances. Extensive experiments over different levels of uncertainty demonstrate that uncertain information does contribute to training and is even necessary within a certain degree of uncertainty. The results show that probability extreme learning machine beats traditional methods in most cases.