Machine Learning with Known Input Data Uncertainty Measure

Abstract

Part 7: AlgorithmsInternational audienceUncertainty of the input data is a common issue in machine learning. In this paper we show how one can incorporate knowledge on uncertainty measure regarding particular points in the training set. This may boost up models accuracy as well as reduce overfitting. We show an approach based on the classical training with jitter for Artificial Neural Networks (ANNs). We prove that our method, which can be applied to a wide class of models, is approximately equivalent to generalised Tikhonov regularisation learning. We also compare our results with some alternative methods. In the end we discuss further prospects and applications

Machine Learning-Based Microclimate Model for Indoor Air Temperature and Relative Humidity Prediction in a Swine Building

“…To achieve the desired accuracy, it is essential to generate a reference for selecting the parameters that need to be recorded. The current study considers the use of different datasets as a useful method to ascertain the appropriate data that may have fewer variables and significant implications for predictions indeed [ 20 , 21 , 36 , 37 ]. So that the current study adopts the Spearman rank correlation coefficient approach in order to extract the best features, which is a commonly followed method to explore the relationships between attributes.…”

“…Determination of input data is the bottom line of any modelling criteria yet crucial consideration in diagnosing the exquisite functional form of ML models. Choosing the right input variables involves improving the accuracy of the algorithm; also, it dominates the calculation speed, training time, training complexity, comprehensibility, and computational effort of the simulation [ 20 , 21 , 22 ]. The present study analyzes the performance of the models with feature-selected datasets and available datasets; it also suggests the optimal input selection to feed the models from the available datasets.…”

Integrating Digital Technologies to Aid Grassland Productivity and Sustainability

“…Uncertainty in data, and therefore in all types of data science models, introduces the risk of poor decision outcomes because of biases, drift and lack of precision in individual sensor systems (Wolfert et al, 2017). Further, as the volume and variety of data increases, so do the uncertainties inherent within (Czarnecki and Podolak, 2013;Hariri et al, 2019)-big data is often subject to noise, incompleteness, bias and inconsistency (Hariri et al, 2019;Sharifi et al, 2020), and may often be disparate, dynamic, untrustworthy, and inter-related (Wang and Jones, 2017).…”

“…The risk of poor decision outcomes is particularly true in analytics that combine non-traditional information sources such as rapidly arriving data from sensors, process models, qualitative information and user behavior (Wynne, 1992). Using multiple disparate data sources means compounding data uncertainty originating from the data collection, data curation and combination from multiple sources (Czarnecki and Podolak, 2013;Hariri et al, 2019).Communicating uncertainty in data can introduce further complexities, and uncertainties are sometimes ignored, or even explicitly denied (van der Bles et al, 2019). Uncertainty in the data collection, analysis and knowledge extension processes can lead to a lack of confidence in the resulting model outputs and decision made thereof.…”

“…Algorithms need to be created, calibrated and validated to predict biophysical information from digital data. Data, both digital and biophysical contains uncertainty through both the accuracy and precision of measurement (Czarnecki and Podolak, 2013) and inherent expression of the observation (Steel and Torrie, 1980).…”