Automated machine learning (AutoML) has become a new trend which is the process of automating the complete pipeline from the raw dataset to the development of machine learning model. It not only can relief data scientists' works but also allows non-experts to finish the jobs without solid knowledge and understanding of statistical inference and machine learning. One limitation of AutoML framework is the data quality differs significantly batch by batch. Consequently, fitted model quality for some batches of data can be very poor due to distribution shift for some numerical predictors. In this dissertation, we develop an intelligent binning to resolve this problem. In addition, various regularized regression classifiers (RRCs) including Ridge, Lasso and Elastic Net regression have been tested to enhance model performance further after binning. We focus on the binary classification problem and have developed an AutoML framework using Python to handle the entire data preparation process including data partition and intelligent binning. This system has been tested extensively by simulations and real datasets analyses and the results have shown that (1) All the models perform better with intelligent binding for both balanced and imbalance binary classification problem. (2) Regression-based methods are more sensitive than tree-based methods using intelligent binning. RRCs can work better than other tree methods by using intelligent binning technique. (3) Weighted RRC can obtain the best results compared to other methods. (4) Our framework is an effective and reliable tool to conduct AutoML.

Understanding data on novel coronavirus (COVID-19) pandemic, and modeling such data over time are crucial for decision making at managing, fighting, and controlling the spread of this emerging disease. This thesis work looks at some aspects of exploratory analysis and modeling of COVID-19 data obtained from the Florida Department of Health (FDOH). In particular, the present work is devoted to data collection, preparation, description, and modeling of COVID-19 cases and deaths reported by FDOH between March 12, 2020, and April 30, 2021. For modeling data on both cases and deaths, this thesis utilized an autoregressive integrated moving average (ARIMA) times series model. The "IDENTIFY" statement of SAS PROC ARIMA suggests a few competing models with suggested values of the parameter p (the order of the Autoregressive model), d (the order of the differencing), and q (the order of the Moving Average model). All suggested models are then compared using AIC (Akaike Information Criterion), SBC (Schwarz Bayes Criterion), and MAE (Mean Absolute Error) values, and the best-fitting models are then chosen with smaller values of the above model comparison criteria. To evaluate the performance of the model selected under this modeling approach, the procedure is repeated using the first six month's data and forecasting the next 7 days data, nine month's data and forecasting the next 7 days data, and then all reported FDOH data from March 12, 2020, to April 30, 2021, and forecasting the future data. The findings of exploratory data analysis that suggests higher COVID-19 cases for females compared to males and higher male deaths compared to females are taken into consideration by evaluating the performance of final models by gender for both cases and deaths' data reported by FDOH. The gender-specific models appear to be better under model comparison criteria Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) compared to models based on gender aggregated data. It is observed that the fitted models reasonably predicted the future numbers of confirmed cases and deaths. Given similarities in reported COVID-19 data, the proposed modeling approach can be applied to data in the USA and many other States, and countries around the world.

Recent advances in technology have increased the need for using simulation models to analyze tasks and obtain human performance data. A variety of task analysis approaches and tools have been proposed and developed over the years. Over 100 task analysis methods have been reported in the literature. However, most of the developed methods and tools allow for representation of the static aspects of the tasks performed by expert system-driven human operators, neglecting aspects of the work environment, i.e. physical layout, and dynamic aspects of the task. The use of simulation can help face the new challenges in the field of task analysis as it allows for simulation of the dynamic aspects of the tasks, the humans performing them, and their locations in the environment. Modeling and/or simulation task analysis tools and techniques have been proven to be effective in task analysis, workload, and human reliability assessment. However, most of the existing task analysis simulation models and tools lack features that allow for consideration of errors, workload, level of operator's expertise and skills, among others. In addition, the current task analysis simulation tools require basic training on the tool to allow for modeling the flow of task analysis process and/or error and workload assessment. The modeling process is usually achieved using drag and drop functionality and, in some cases, programming skills. This research focuses on automating the modeling process and simulating individuals (or groups of individuals) performing tasks in a dynamic work environment in any domain. The main objective of this research is to develop a universal tool that allows for modeling and simulation of task analysis models in a short amount of time with limited need for training or knowledge of modeling and simulation theory. A Universal Task Analysis Simulation Modeling (UTASiMo) tool can be used for automatically generating simulation models that analyze the tasks performed by human operators. UTASiMo is a multi-method modeling and simulation tool developed as a combination of agent-based, discrete event, and system dynamics simulation models. A generic multi-method modeling and simulation framework, named 3M&S Framework, as well as the Unified Modeling Language have been used for the design of the conceptual model and the implementation of the simulation tool. UTASiMo-generated models are dynamically created during run-time based on user inputs. The simulation results include estimations of operator workload, task completion time, and probability of human errors based on human operator variability and task structure.

Attributed graph is a powerful tool to model real-life systems which exist in many domains such as social science, biology, e-commerce, etc. The behaviors of those systems are mostly defined by or dependent on their corresponding network structures. Graph analysis has become an important line of research due to the rapid integration of such systems into every aspect of human life and the profound impact they have on human behaviors. Graph structured data contains a rich amount of information from the network connectivity and the supplementary input features of nodes. Machine learning algorithms or traditional network science tools have limitation in their capability to make use of both network topology and node features. Graph Neural Networks (GNNs) provide an efficient framework combining both sources of information to produce accurate prediction for a wide range of tasks including node classification, link prediction, etc. The exponential growth of graph datasets drives the development of complex GNN models causing concerns about processing time and interpretability of the result. Another issue arises from the cost and limitation of collecting a large amount of annotated data for training deep learning GNN models. Apart from sampling issue, the existence of anomaly entities in the data might degrade the quality of the fitted models. In this dissertation, we propose novel techniques and strategies to overcome the above challenges. First, we present a flexible regularization scheme applied to the Simple Graph Convolution (SGC). The proposed framework inherits fast and efficient properties of SGC while rendering a sparse set of fitted parameter vectors, facilitating the identification of important input features. Next, we examine efficient procedures for collecting training samples and develop indicative measures as well as quantitative guidelines to assist practitioners in choosing the optimal sampling strategy to obtain data. We then improve upon an existing GNN model for the anomaly detection task. Our proposed framework achieves better accuracy and reliability. Lastly, we experiment with adapting the flexible regularization mechanism to link prediction task.

Sequential multiple change point detection concerns the identification of multiple points in time where the systematic behavior of a statistical process changes. A special case of this problem, called online anomaly detection, occurs when the goal is to detect the first change and then signal an alert to an analyst for further investigation. This dissertation concerns the use of methods based on kernel functions and support vectors to detect changes. A variety of support vector-based methods are considered, but the primary focus concerns Least Squares Support Vector Data Description (LS-SVDD). LS-SVDD constructs a hypersphere in a kernel space to bound a set of multivariate vectors using a closed-form solution. The mathematical tractability of the LS-SVDD facilitates closed-form updates for the LS-SVDD Lagrange multipliers. The update formulae concern either adding or removing a block of observations from an existing LS-SVDD description, respectively, and thus LS-SVDD can be constructed or updated sequentially which makes it attractive for online problems with sequential data streams. LS-SVDD is applied to a variety of scenarios including online anomaly detection and sequential multiple change point detection.

Concept drift detection, the identfication of changes in data distributions in streams, is critical to understanding the mechanics of data generating processes and ensuring that data models remain representative through time [2]. Many change detection methods utilize statistical techniques that take numerical data as input. However, many applications produce data streams containing categorical attributes. In this context, numerical statistical methods are unavailable, and different approaches are required. Common solutions use error monitoring, assuming that fluctuations in the error measures of a learning system correspond to concept drift [4]. There has been very little research, though, on context-based concept drift detection in categorical streams. This approach observes changes in the actual data distribution and is less popular due to the challenges associated with categorical data analysis. However, context-based change detection is arguably more informative as it is data-driven, and more widely applicable in that it can function in an unsupervised setting [4]. This study offers a contribution to this gap in the research by proposing a novel context-based change detection and adaptation algorithm for categorical data, namely Fine-Grained Change Detection in Categorical Data Streams (FG-CDCStream). This unsupervised method exploits elements of ensemble learning, a technique whereby decisions are made according to the majority vote of a set of models representing different random subspaces of the data [5]. These ideas are applied to a set of concept drift detector objects and merged with concepts from a recent, state-of-the-art, context-based change detection algorithm, the so-called Change Detection in Categorical Data Streams (CDCStream) [4]. FG-CDCStream is proposed as an extension of the batch-based CDCStream, providing instance-by-instance analysis and improving its change detection capabilities especially in data streams containing abrupt changes or a combination of abrupt and gradual changes. FG-CDCStream also enhances the adaptation strategy of CDCStream producing more representative post-change models.

<p>It has been claimed that there are considerable differences between pri-vate and non-private patients with regard to the outcome of back surgery.This can be found in the yearly report from the register concerning backsurgery in Sweden. However, the results seem doubtful and the referencescould not be found. Therefore, we analyze the data about nearly 1200patients from the clinic of back surgery in Str¨angn¨as (CSS). It includesthree time periods with somewhat different questionnaires from 1986 to2007 with both private and non-private patients. In the third period,the patients have been evaluated using the SF-36 questionnaire. The re-sults show that most of the differences between private and non-privatepatients are minor and not statistically significant.</p>

It has been claimed that there are considerable differences between pri-vate and non-private patients with regard to the outcome of back surgery.This can be found in the yearly report from the register concerning backsurgery in Sweden. However, the results seem doubtful and the referencescould not be found. Therefore, we analyze the data about nearly 1200patients from the clinic of back surgery in Str¨angn¨as (CSS). It includesthree time periods with somewhat different questionnaires from 1986 to2007 with both private and non-private patients. In the third period,the patients have been evaluated using the SF-36 questionnaire. The re-sults show that most of the differences between private and non-privatepatients are minor and not statistically significant.

Biclustering is a technique for clustering rows and columns of a data matrix simultaneously. Over the past few years, we have seen its applications in biology-related fields, as well as in many data mining projects. As opposed to classical clustering methods, biclustering groups objects that are similar only on a subset of variables. Many biclustering algorithms on continuous data have emerged over the last decade. In this dissertation, we will focus on two Bayesian biclustering algorithms we developed for discrete data, more specifically categorical data and ordinal data. / Statistics

Many people in widely varied fields are exposed to categorical data describing myriad observations. The breadth of applications in which categorical data are used means that many of the people tasked to apply these data have not been trained in data analysis. Visualization of data is often used to alleviate this problem since visualization can convey relevant information in a non-mathematical manner. However, visualizations are frequently static and the tools to create them are largely geared towards quantitative data. It is the purpose of this thesis to demonstrate a method which expands on the parallel coordinates method of visualization and uses a 'Google Maps' style of interaction and view dependent data presentation for visualizing and exploring categorical data that is accessible by non-experts and promotes the use of domain specific knowledge. The parallel coordinates method has enjoyed increasing popularity in recent times, but has several shortcomings. This thesis seeks to address some of these problems in a manner which involves not just addressing the final static image which is generated, but the paradigm of interaction as well.