Meta-information for data interference reduction in classification and forecasting

To achieve interference-less machine learning, therefore avoiding the negative impact of excessive noise/outlier on training and testing accuracy, we establish four fundamental hypotheses for application in classification and forecasting tasks. A spatial transformation, such as Density-Based Spatial Clustering of Applications with Noise (DBSCAN), and a temporal transformation, such as Empirical Mode Decomposition (EMD), have shown the potential to form meaningful representations of the data in classification and forecasting tasks respectively. Using these hypotheses, the dataset is preprocessed to generate meta-information. This meta-information is utilized to guide the model building stage and noise reduction is evident. Several learning algorithms—for instance, the Constructive Backpropagation (CBP) for classification and the long short-term memory (LSTM) neural network for forecasting—have been augmented and tested on real-world benchmark datasets and our results reported in several research proceedings reveal significant performance enhancement when conditions for these hypotheses are satisfied. This paper presents an overview of the techniques and potential areas of application.