Improvement of the model by preprocessing big data of tapping temperature prediction industry

Abstract

The quality of steel is highly related to the tapping temperature. At present, many models can predict the tapping temperature to a certain extent. However, after in-depth exploration of the model, it was found that most of the models have a better performance for the tapping temperature prediction of common steel, while the special steel was ignored. It is because that most models are rough in the data processing. The neglect of incomplete data, uneven distribution problems have caused models to over-emphasize some types of steel, while it does not agree with the production demand of steel mills. This paper proposes Bayes neural network model with the improved filling method based on K-Means algorithm and data homogenization method based on K-Means algorithm. The statistical results show that the model has a better performance at error within ±5°C ratio, mean absolute error, root means square error, etc. than the traditional ones.

SIGANEO: Similarity network with GAN enhancement for immunogenic neoepitope prediction

“…It is also worth noting that the average performance of zero-filling-coupled network is higher than mean-value-filling-coupled network. One possible explanation is that zero-filling method can mask missing values by 0, whereas mean-filling method introduces more inaccurate information to the network [51] . To gain a more intuitive understanding of the impacts of missing-value imputation methods on neoantigen prediction performance, we randomly removed 42.5 % of values from the test dataset according to the rate of missing value in the training dataset.…”