China
Africa
Explore chapters and articles related to this topic
The Use of Machine Learning in Heat Transfer Analysis for Structural Fire Engineering Applications
Published in M.Z. Naser, Leveraging Artificial Intelligence in Engineering, Management, and Safety of Infrastructure, 2023
Yavor Panev, Tom Parker, Panagiotis Kotsovinos
As seen from Fig. 16.4, SVMs are developed to assess the Euclidian (“straight line”) distance between the prediction and the fitted decision boundary (hyperplane). However, this distance is model-specific which can make it difficult for the user to gauge its magnitude without having profound knowledge of the contents of the used dataset. To address this model dependency, the Platt scaling transformation technique (Platt John, 2000) was used to map the distances between the data points and the decision boundaries between values ranging from 0.5 (achieved at the boundary) and 1 (achieved further away from the boundary). In basic principle, Platt scaling works by fitting a logistic regression to the scores produced by the SVM. It should be noted that the technique was adopted only to provide a “consistent measure” of relative strength of prediction based on distance from a decision boundary within the same dataset. This implementation was not intended to compare prediction results between different ML models or to infer an empirical probability distribution.
Mapping near-real-time power outages from social media
Published in International Journal of Digital Earth, 2019
Huina Mao, Gautam Thakur, Kevin Sparks, Jibonananda Sanyal, Budhendra Bhaduri
In practice, not only do we want to know if a tweet is actually about a power outage, but we also want to know the confidence score on the prediction. To obtain the confidence score, we performed probability calibration for the SVM outputs based on Platt's sigmoid model (Platt 1999), as shown in Equation (2).where x is the input and is the output of the classifier. A and B are parameters learned by the algorithm. Briefly speaking, Platt scaling fits a logistic regression model to classifier outputs and produces probability estimates (from 0 to 1). Classification results have shown that, among the set of 621,544 tweets containing power outage-related keywords, 90,060 (14%) tweets had a probability , which we consider positive tweets. Among the set of positive tweets, the percentage of tweets with is 85% (76,588); , 73% (65,667); , 55,234 (61%); , 40,316 (45%). Figure 3 shows the number of power outage tweets with a probability of over the study period. Spikes of outage tweets can clearly be seen across two hurricanes–Hermine (28 August to 8 September 2016) and Matthew (28 September to 10 October 2016)–and on September 13, when there were major power outages in north New Jersey (including the Newark Airport), as well as on 22 September 2016, when a fire at a Puerto Rico power plant caused utility failure.