Improved Algorithms Based On Extreme Learning Machine For Handing With Missing Data And Application

Since the beginning of the21st Century, the human beings speed up theexploration of marine resouses. There are so abundant amount of marine resourses indeep oceans that lots of countries strive for their maritime interests. The explorationdata of seabed mineral resources conclude a large amount of missing values, and thereare less methods to handle with missing data in mineral resource evaluations. Thispaper is aimed at researching on the machine learning methods handling with missingdata which are used to evaluate seabed mineral resources.Extreme learning machine (ELM) is one of the recent successful machine learningapproaches for its extremely fast traing speed and good generalization performance.The sparse Bayesian ELM (SBELM) approach and the TROP-ELM approach, whichare two recent variants of ELM, can result in more accurate and compact model.However, the SBELM and TROP-ELM can not deal with the missing data in theirstandard forms. To solve the problem, we design two novel methods, by adjusting thecalculation of outputs of the hidden layers in SBELM and TROP-ELM, as: additivemodels for missing data (AM-MD) and self-adjusting neuron state models for missingdata (SNSM-MD). Experimental results on several data sets from the UCI repositoryindicate that the proposed modied SBELM methods and TROP-ELM methods havesignicant advantages: high accuracy or low error and good generalization performancecompared with several other existing methods. Moreover, the proposed methods enrichELM with new tools to solve missing data problem for multi-class classication andregression even with up to50%of the features missing in the input data. At last, themodified TROP-ELM will be used in the evaluations of seabed mineral resources withmissing values.