Browsing by Author "Roy, Arighna"

Now showing 1 - 2 of 2

Pattern Recognition and Quantifying Associations Within Entities of Data Driven Systems for Improving Model Interpretability
(North Dakota State University, 2022) Roy, Arighna
Discovering associations among entities of a system plays an important role in data science. The majority of the data science related problems have become heavily dependent on Machine Learning (ML) since the rise of computation power. However, the majority of the machine learning approaches rely on improving the performance of the algorithm by optimizing an objective function, at the cost of compromising the interpretability of the models. A new branch of machine learning focuses on model interpretability by explaining the models in various ways. The foundation of model interpretability is built on extracting patterns from the behavior of the models and the related entities. Gradually, Machine learning has spread its wing to almost every industry. This dissertation focuses on the data science application to three such domains. Firstly, assisting environmental sustainability by identifying patterns within its components. Machine learning techniques play an important role here in many ways. Discovering associations between environmental components and agriculture is one such topic. Secondly, improving the robustness of Artificial Intelligence applications on embedded systems. AI has reached our day-to-day life through embedded systems. The technical advancement of embedded systems made it possible to accommodate ML. However, embedded systems are susceptible to various types of errors, hence there is a huge scope of recovery systems for ML models deployed on embedded systems. Third, bringing the user communities of the entertainment systems across the globe together. Online streaming of entertainment has already leveraged ML to provide educated recommendations to its users. However, entertainment content can sometimes be isolated due to demographic barriers. ML can identify the hidden aspects of these contents which would not be possible otherwise. In subsequent paragraphs, various challenges concerning these topics will be introduced and corresponding solutions will be followed that can address those challenges.
Quantifying Relationships Between Two Time Series Data Sets
(North Dakota State University, 2016) Roy, Arighna
One of the popular methods for quantifying the relationship between two time series data sets is canonical correlations; however, it is linear and cannot accommodate more complex scenarios, such as time series data for which distance relationships are best characterized through dynamic time warping. I introduce a nearest-neighbor-overlap method that resolves both problems and allows a reliable determination of signi cant relationships. The nearest neighbor algorithm also does not depend on the normal distribution of the variables, unlike canonical correlations. Also, it is not sensitive to singularity (when one variable is derivable from another) of the data. I demonstrate that our method substantially outperforms canonical correlation analysis for time series data sets from the UCR repository as well as the environmental data of Red River Valley region.