Douglas Honors College (DHC) Classifying Two-Dimensional Data by Detecting Compensations Carson Gavin Project Mentor(s): Boris Kovalerchuk, PhD
Machine learning algorithms are efficient methods to build data classification models. However, many methods produce black-box models that domain experts cannot understand, making them difficult to trust and deploy. Alternative visual and analytical methods are needed to provide explanation for such models. The purpose of this project is to develop such a method based on the concept of compensation. This method builds a linear classification line by detecting dominant directions in a data set. This is done by finding groups of points that share compensation properties, which can be understood by a domain expert. Dominant angles can be defined as those that intersect the most points. This can be calculated by finding how many points are intersected when drawing several lines of that angle. By averaging the angle between the dominant angles of two classes and taking the center point between them, a linear classifier is constructed. This method has been run on the Iris data set, obtaining 100% accuracy when classifying the Setosa class and 95% accuracy when comparing the Versicolor and Virginica classes. These accuracies are on par with less explainable models. Several improvements can be made to the current algorithm, such as testing several dominant angles or creating multiple starting points to find the best possible linear classifier. The approach can also be expanded to any number of data dimensions. The utility of this algorithm is to create an algorithm that can be understood by domain experts, and to avoid reliance on black-box machine learning models. Presentation Type: Poster Presentation (May 21, 9:30am–3:00pm) Keywords: Compensation, Linear Classification, Two-dimensional data, Model explanation SOURCE Form ID: 94 An Interactive Stochastic Population Projection Model on Florida Manatees Koah Ghrist Project Mentor(s): Yuba Siwakoti, PhD Making informed decisions on species conservation requires careful consideration to avoid misplacing major resources to minor threats or failing to address threats all together. Past studies have developed stochastic population projection models to best inform conservation decisions for the Florida manatee; however, a publicly available, user friendly, and interactive model does not yet exist. This study develops such a model, inspired by PhET simulations. The model is programmed in Python alongside the PyGame and MatPlotLib libraries for GUI and graphing purposes respectively. Manatee and climate behavior is programmed using current data and theories provided by relevant peer reviewed research. Warm water refuges are a focal point for this model due to their immense impact on manatee survivability. Users drag and drop warm water refuges onto a map of Florida's waters, establishing environmental factors and starting population through sliders. Mortality events may be created at any point during a model instance to apply stress to any sub- population. Model instances may be constrained to a time limit or continue indefinitely until extinction or user intervention. This model produces results approximate to those before it regarding the importance of proactive conservation measures. Future models may aim for a browser-based environment, to further public accessibility; alternatively, future versions may contain manatee interactions with feeding grounds, to demonstrate another important factor to manatee survivability. Presentation Type: Oral Presentation (May 20, 9:30am–5:00pm) Keywords: Manatee, Warm Water Refuge, Simulation, Conservation, Education SOURCE Form ID: 156
164
Made with FlippingBook interactive PDF creator