Alexandra Castelazo

Analysis on trends, seasonality, and forecasting of police shootings with a dataset with information on daily shootings, as well as the ethnicity of the people that had been shot from 2015-2020. The data included six ethnical categories: White, Black, Hispanic, Asian, Native, and Other.

Above are predictions for the different ethnicity groups included in the data. An important note about the data is that the totals for each ethnicity group vary, so some of the predictions will be much smaller than others.

Prediction and detection of credit card fraud from a dataset of 1000 observations of accounts that have at least one recently declined charge and other demographic and account information. Methods for model selection are Logistic regression, Linear Discriminant Analysis, GAM's, Decisions Trees, and Random Forests and an implementation of a probability threshold of 0.25 to help with selecting the model that overestimates the number of compromised credit cards to help minimize overall costs.

The final chosen predictive model, out of a total of 15 models, was a Random Forest for the following model: fraud ∼ number of declines + age + amount. This model had an AUC (for both training and test sets) of 0.8919.

Prediction of housing price estimates for a collection of ten residential buildings to be sold in January 2020, where the predictions are not to be higher than 10% of the true market prices so the houses sell. The data has 1587 known property transactions that took place in Taipei City and contains information about each property including the transaction price, the district in which the building is located, the land and building areas, the main building material, the numbers of rooms, hallways, and bathrooms, the year and month of the transaction, and the year the building was made.

The above plot shows the residuals vs fitted values. The red line is a scatterplot smoother and represents the average value of the residuals (the difference between the observed and predicted prices) for different fitted values. Since the red line trends above the middle dotted line (which represents zero error, a perfect prediction), this means the true values of the house prices are on average larger than our predicted prices. This plot suggests that our model tends be most accurate for smaller housing prices but tends to underestimate for bigger housing prices

Prediction of the mortality rates of 418 passengers who were aboard the Titanic with the use of machine learning. The data used included information on 891 passengers of the titanic, which contains information about each individual passenger, including their ticket class (which is another measure of socioeconomic status), sex, age, the number of siblings, spouses, children, and parents they traveled with, assigned ticket number, passenger fare, cabin number, the port that they embarked from, and whether they survived the shipwreck or not. Models chosen were prediction accuracy oriented rather than an inferential oriented. Variable data imputation was also necessary for some variables with missing information.

The above are ROC curves for each of the 10 folds with probability predictions for the support vector machine model. Out of all of the other models created, the support vector machine model had the best prediction accuracy with a lower deviation.

Analysis on variables relating to experiencing stroke and determining whether there are actionable steps one could take to reduce their risk of stroke. To evaluate the model's ability to predict stroke to that of other potential stroke-predictor algorithms. The data used had approximately 25k people, containing information potentially-relevant lifestyle variables. The data was split into a training set, which additionally contained the stroke outcome meant to build our model on, and a testing set meant to test the accuracy of our model. These samples contained 15,303 and 10,204 observations respectively.

Based on the table above, the most accurate model was the Ensemble model, which was based off Random Forest, Lasso GLM, K Nearest Neighbors (KNN), and Gradient Boosting Machine (GBM) predictors. Although the model has a high accuracy, the model is not interpretable.

The analysis on the number PPE usage requirements. The data was collected randomly from the Cal Poly Pomona University Library. The goal of this analysis is to ensure that at least 65% of students in the University Library are wearing a mask of surgical grade or better at any given time. If the requirement is not met, the data will allow the university to determine how to best proceed in changing policies to ensure student safety from COVID-19.

Cluster analysis on the Global School-Based Health Survey which was done in schools across Argentina in 2018. A total of 56,981 students participated and more than 19 variables were recorded for each of these students. Before utilizing the data, it was cleaned to only include 19 variables. The data set was put together through a self-administered questionnaire meant to obtain data on young people’s health behaviour and protective factors related to the leading causes of morbidity and mortality. The variables were analyzed by their correlation with one another and the number of clusters were determined with a scree plot.

The above image is a correlation heatmap where the largest correlation between each of the variables are boxed in green.

Analysis on an anonymized sleep study with the goal to determine whether the difference in drugs taken can cause a difference in sleep patterns for each patient in the study. After determining the data is not normally distributed, the Wald-Wolfowitz Runs Test and the Mann-Whitney-Wilcoxon Test were ran at a significance level of 5%, and it was concluded that the difference in drugs taken do not cause a difference in sleep patterns for their patients.

Analysis on the difference of effects on weight for each feed type given to chicks. Tests used for analysis are the Chi-Squared Test for homogeneity, the 𝐺2 Test, and the Fisher’s Exact Test.

Prediction of how many likes a movie will receive. The Movies data set was used for this analysis and originated from the UCI Machine Learning Repository, which contains information on 231 Movies from the years 2014-2015 and variables such as Year, Ratings, Gross, Budget, Views, Likes, Dislikes, and Comments. The data was split into a 50:50 training and test set split. Model selection was carried out with forward selection, transformations, Ridge, and Lasso Regression.

The coding and creation of a theremin (an electronic musical instrument controlled without physical contact by the performer). The code was made with Arduino IDE and included circuitry basics. I enjoy applying my knowledge of math and coding to physical projects and learning more about science in the process.