Support Vector Machine (SVM) is a reliable supervised learning model extensively utilized for classification and regression tasks, owing to its remarkable ability to achieve strong generalization performance. This study focuses on two key factors in the SVM model: the error penalty parameter C and the kernel function. The C parameter is used to balance the model’s complexity and empirical risk, and its selection is crucial for SVM performance. A smaller C value may lead to underfitting, while a larger C can result in overfitting. Additionally, the choice of the kernel function also significantly impacts SVM performance. We will investigate the effects of different kernel functions and parameter settings in the classification task of the Iris dataset and visualize their impacts through a visual approach. The study’s results indicate that, in most cases, the Gaussian kernel outperforms other kernel functions, exhibiting superior classification performance and generalization capability. Therefore, we opt for the Gaussian Radial Basis Function (RBF) kernel and conduct experiments to evaluate the influence of different parameter configurations on classification performance.

Repeated-Measures longitudinal data is common in drug research, where every patient is repeatedly measured across time. Responses could either be continuous variables such as blood pressure or binary variables such as drug test positive/negative. Issues to be addressed are within-subject observation dependence, as well as the between-subject differences (mixed effects). Another important problem to address is the missingness and dropouts.

Full likelihood-based models such as the generalized linear mixed model (GLMM) together with EM algorithm could be utilized, given simplified parametric correlation structure between random components. If the interest is only the mean parameters, little in subject effects, the non-likelihood-based generalized estimating equations (GEE) is a good alternative. GEE circumvents the structural and computational complexities of likelihood-based models, and it is robust to misspecification of the working correlation structures of marginal observations. However, as a non-likelihood frequentist marginal model, GEE itself lacks strength dealing with missing data mechanism beyond missing completely at random (MCAR). Therefore, integration of multiple imputation and GEE (MI-GEE) is a great solution to longitudinal data with dropouts.

Real data application is performed on MI-GEE and GLMM. Our results for the bupropion study dataset show effective but not significant strength of Bupropion for treating methamphetamine dependence, which is consistent with previous studies and biological sense.

High blood pressure is a prevalent health concern worldwide, and identifying the factors that contribute to its development is crucial for prevention and management strategies. This study aimed to investigate the influence of sex, hereditary factors, habitats, and BMI on the risk of high blood pressure using machine learning techniques. Several models, including Logistic Regression, Decision Trees, Random Forests, XGBoost, Support Vector Machines, and Neural Networks are employed on the public-use sample from the Add Health dataset.

Hospital readmission is a costly, undesirable, and often preventable patient outcome of inpatient care. Early readmission prediction can effectively prevent life-threatening events and reduce healthcare costs. However, imbalanced class distribution and high missing value rates are usually associated with readmission data and need to be handled carefully before building classification models. In this paper, we investigate the prediction of hospital readmission on a dataset with high percentage of missing values and class imbalance problem. Different methods are applied to impute missing values in the categorical variables and numerical variables. In addition, SMOTE (Synthetic Minority Over-sampling Technique) and cost-sensitive learning are combined with different classification methods (LASSO logistic regression, random forest, and gradient boosting) to explore which one will yield the best classification performance on the readmission data. Total misclassification cost and area under ROC curve are used as evaluation metrics for model comparison. Our results show that the SMOTE method causes overfitting on our readmission data and cost-sensitive learning outperforms SMOTE in terms of total misclassification cost.

Yelp has been a dominating company in the business rating industry. In the current days, reviews and ratings critically affect people’s decisions. A model that effectively predicts the ratings could be supportive to businesses. In this research, we develop several models based on various NLP and regression methods based on the text review and other traits of the reviews to predict the final star rating of each. Models including LSTM, GRU, BERT, and multinomial logistic regression are utilized. The fine-tuning BERT model has the best result of 68.8% on the testing dataset. Although the model predicts well in general, we could further investigate and evaluate text reviews on multiple aspects.

The unit price of an electronic component with certain specifications and purchase details could be crucial for the decision-making of customers. With the massive historical purchasing data at Supplyframe, Inc., classical statistical and Machine Learning (ML) models are used to capture the underlined relationship and predict accurate price. The Naive model using mean estimation is adopted as baseline models, followed by the exploration of a wide range of machine learning models including Ordinary Least Squares, Supporting Vector Machine, $k$-Nearest Neighbors, Random Forests (RF), Extreme Gradient Boost (XGB). To make better use of the unstructured features, Deep Neural Networks (DNN) are built based on Convolutional Neural Networks and feature embedding, which maps unstructured features to higher dimension vectors. We observe that the RF and XGB models outperform other classical statistical and ML models when only the structured features are used while the DNN model is proved to be the most powerful by combining both structured and unstructured features. Consistent superior performances are found for the DNN model in terms of the root mean squared error, the prediction interval of the ratios of observed and predicted values, the prediction coverage rate and the capture of the monotonic decreasing relationship between unit prices and purchase quantities.

Pledgeling is a platform that powers corporate giving and social impact programs for busi-

nesses of all sizes by integrating donation features in the backend. One of Pledgeling's part-

ners is Evite, the world's leading digital platform for bringing people together with event

invitations. In this study, the data from Pledgeling is used to train a logistic regression

model to determine which aspects of RSVP events lead to hosts turning on the donation

features. From the rst study, it was statistically signicant that categories of events that

fall into Organizations, Weddings, and Animals were more likely to add donation features.

Also events from the West and Northeast regions during Q1 and Q4 during the calender

year were more likely to add donation features. Additionally, data is used to train a linear

regression model to study which features lead to users donating more for each event. From

the second study, there were statistically signicant results that categories of events that fall

into Organizations, Weddings, and Get Togethers were donating more money. Also, results

indicated that events in West and Northeast regions were donating more from the beginning

of summer to the end of the year. Lastly, events donated more to Science-related as well as

Public Health-related causes compared to causes for Animals organizations.

This thesis presents a study on location optimization for franchisee restaurants using Graph Neural Network (GNN) models, namely GCN, GAT, and GraphSAGE. The research employs these models to analyze geographic coordinates and other relevant data to predict optimal franchise locations. By incorporating real-world data such as Yelp reviews, census information, and city demographics, the study attempts to model the significant factors that influence the success of franchise locations. The primary contribution of this work is the development of a tool that aids market research teams in making informed decisions about where to establish new restaurant outlets and optimizing location selection through advanceddata analytics and machine learning techniques. Among the three models, the GraphSAGE model performed best. It achieved a loss of 0.48, an accuracy of 77%, and an ROC score of 0.78, outperforming the other models across various assessments.

In this application, a predictive model is built for finding the price of an item from their characteristics, especially the description. The implementation of word vectors, optimized gradient descent, and long short-term memory are employed to tackle the problem. And the relatively low loss for both training and test set proves this is a reasonable approach.

The coronavirus disease 2019 (COVID-19) pandemic has already become a global threat. To fight against COVID-19, effective and fast screening methods are needed. This study focuses on leveraging deep learning techniques to automatically detect COVID‐19 pneumonia in chest X-ray images. Two models are trained based on transfer learning and residual neural network. The first one is a binary classifier that separates COVID-19 pneumonia and non-COVID-19 cases. It classifies all test cases correctly. The second one is a four-class classifier that distinguishes COVID-19 pneumonia, viral pneumonia, bacterial pneumonia and normal cases. It reaches an average accuracy, precision, sensitivity, specificity, and F1-score of 93\%, 93\%, 93\%, 97\%, and 93\%, respectively. To understand on how the four-class classifier detects COVID-19 pneumonia, we apply Gradient-weighted Class Activation Mapping (Grad-CAM) method and find out that the classifier is able to focus on the patchy areas in chest X-ray images and make accurate predictions.