Moving to a new country is like diving headfirst into a novel where you are barely fluent in the language, and everyone else has already read the plot. Each day, I find myself stumbling through moments of awe and confusion. Even the simplest things — directions, food labels, conversations — feel like riddles waiting to be solved. My comfort zone has become a distant memory, replaced by a world where each small interaction is a test of patience and adaptability. One month down, with possibly three years and eleven months to go. Yet I still brace myself every time I venture out to tackle the mysteries of grocery aisles or government offices. Even routine tasks like opening a bank account took nearly three weeks, leaving me to wonder if the Institute had anticipated these bureaucratic adventures in my fellowship. I would not call it entitlement, but as an MSCA Fellow, I have a right to require support for all the administrative procedures that come with being an early-stage researcher in a foreign country.

Then there is the research program and doctoral school — an entirely different kind of adventure. Long hours of data extraction, thanks to the computational complexity of implemented algorithm, have left me bleary-eyed. I have had a few presentations in this first month, each a fresh reminder of my own nerves and the gnawing doubt about whether I truly understand my own research. The language barrier adds yet another layer. Asking for technical help or sharing ideas with peers requires careful phrasing and humility. I second-guess my words, wondering if I have managed to communicate my thoughts clearly or if something essential has slipped through the cracks. It is a strange feeling, standing before an audience and realizing that while I am trying to bridge a language gap, I am also trying not to trip over my own words. Somehow, I will either find resilience in all this fumbling or surrender and go back to where I came from.

Each day, I find myself juggling between moments of optimism — trusting that I will eventually feel at home — and flashes of pessimism, wondering if I will always feel like a stranger. Let us see where this journey takes me, and decide by then whether there is still a reason, or even a glimmer of hope, to keep pursuing this doctorate degree.

The first year of a PhD is a whirlwind of ideas and ambitions, where every choice shapes the trajectory of your academic journey. Writing a review paper often seems like an early milestone—a chance to showcase your grasp of the field while establishing yourself as a knowledgeable voice. Yet, for those engrossed in research, the time and effort required for a strong review can feel like a diversion from achieving tangible results.

A wise alternative is to prioritize the research itself. By weaving literature insights directly into the project, one can stay informed while maintaining focus on outcomes that matter. Tracking key publications and even preprints ensures you remain updated without the added pressure of formalizing findings into a standalone review. This approach channels your time into meaningful progress rather than dispersing efforts, avoiding premature proposals that might later prove challenging.

Every PhD journey is unique, and while a review paper might be a pivotal step for some, it is not a universal necessity. Whether you choose to dive into a review early or let it emerge naturally within your research, the goal remains the same: to contribute meaningfully to your field and community. Focus on what propels your work forward, and the rest will follow.

(The associated codes and implementations in this project is located in the Jupyter notebook.)

Background and Objective

You are working as a data scientist at a local University. The university started offering online courses to reach a wider range of students. The university wants you to help them understand enrollment trends. They would like you to identify what contributes to higher enrollment. In particular, whether the course type (online or classroom) is a factor.

Executive Summary

To determine whether course type is a factor in enrollment, we analyze the enrollment counts and test our hypotheses. Furthermore, we provide a machine learning model to predict enrollment counts of future offered courses by the university. The report is divided into the following sections:

1. Introduction: Overview of enrollment trends in universities offering both online and onsite courses.
2. Data Preprocessing: Detailed explanation on data cleaning.
3. Hypothesis Testing: Test to determine whether difference in enrollment counts is significant.
4. Regression Model: Model to predict future enrollment trends.
5. Conclusion: Final remarks obtained from the model experiments.

Introduction

Enrollment trends in online courses accelerated dramatically during the COVID-19 pandemic. During the height of the pandemic, most classes moved to online-only instruction. While the effects of the pandemic has subsided, online learning remains popular, with many students choosing to take online courses or even complete entire degrees online. In hindsight, online courses offer flexibility and convenience that traditional classroom-based courses cannot. Students can take online courses at their own pace and on their own schedule, from anywhere in the world. This makes online learning a good option for students who are working full-time, have families, or live in remote areas. However, there are also some drawbacks such as the difficulty of staying motivated and engaged. Students also miss out on the social interaction and networking opportunities that are available in traditional classroom settings. It is important to note that online and onsite learning are not mutually exclusive. Many students choose to take a mix of online and onsite courses. This allows them to take advantage of the benefits of both types of learning.

Data Preprocessing

First, we import the data and clean the data using methods and functions available in the pandas library. For every column, we shall do the following:

1. Check the number of missing values.
2. Check the values whether they match the appropriate description (or data type).

The pandas are imported using the alias pd. The .read_csv method imports the dataset . The .head method allows observation of the first few specified rows of the pandas dataframe.

Using the .info method, the data types of the features are as follows:

• course_type: integer
• year: integer
• enrollment_count: string
• pre_score: string
• post_score: floating-point number
• pre_requirement: string
• department: string

For the missing values, the post_score and pre_requirement columns have 185 and 89 missing values respectively. These values are obtained using the .isna and .sum methods.
Now, we correct the data types and fill-in the missing values. The missing scores are assumed to be equal to zero (0) while the missing pre-requisitees are assumed to be empty or 'None'. Note that some values of the pre-scores are marked as a dash. We also replace these values as 0. Afterwards, the pre_score column is converted as a list of floating-point numbers, similar to the post_score, using the .astype method. Lastly, the 'Math' and 'Mathematics' values both exist in the department column. To ensure consistency, 'Math' values are replaced by 'Mathematics'.
The course ID column contain unique values. This means that the number of courses is one-thousand eight hundred fifty (1850). Using the .value_counts method, one thousand three hundred seventy-five (1375) are online while four hundred seventy-five (475) are classroom-based.
To provide an overview of the number of courses of each type, refer to the countplot shown below. Online courses are dominant for both the departments and the university. The Technology department offered the most online courses. For succeeding visualization, we shall mainly use the Seaborn package.

Hypothesis Testing

Recall that we intend to compare the difference between the enrollment counts in each type of course. To start with the hypothesis testing, first observe the the distribution of the enrollment counts. Many courses attracts more than two-hundred fifty (250) enrollees. Moreover, no courses are enrolled with more than two hundred (200) or greater than two hundred twenty (220) students. Lastly, the t-test is not advised to utilize since the enrollment counts are not normally distributed.

We provide more insights about the enrollment. Consider the boxplot showing the enrollment counts across the course types. Students tend to enroll more in online courses than in classroom-taught courses. The range and the median of enrollment counts for online courses are farther and greater than that of the classroom-taught courses. As a remark, no anomalies in the enrollment counts are present for both course types.

We proceed with the hypothesis testing using a Mann-Whitney U Test. This test is a non-parametric version of the t-test suitable for non-normally distributed data. For implementation in Python, the mannwhitneyu function is imported in the scipy.stats library. By index slicing, the enrollment counts of classroom-based courses is separated from that of the online courses. We choose a significance value of 0.01 and a null hypothesis stating that there is no significant difference between the course type. A p-value of a power of negative two hundred thirty-six (-236) is obtained. Thus, the null hypothesis is rejected and say that there is a significant difference between the course type in the enrollment.

Regression Model

Before choosing a regression model, we look at the correlation coefficients of the features. Highly-correlated features need to be reduced as they cause bias to the model. A heatmap showing the correlation coefficients of features are shown below.

We employ a regression model to predict the enrollment count of a course offering. For a baseline model, a simple Linear Regression model is implemented using the scikit-learn library. Linear regression is a good baseline model due to its simplicity and computing efficiency. It is also robust to overfitting. For a comparison model, we shall implement the Elastic Net model, which is also less prone to overfitting. In this model, the loss function puts more weight on errors is utilized. Additionally, an elastic net model is able to select important features and handle correlated features.
For feature training, the target variable enrollment_count is dropped from the set of features. The categorical variables are one-hot encoded to using the .get_dummies method of pandas.

We choose the learning rate of the elastic net equal to 0.0001. After predicting the enrollment counts of a testing subset, the root mean-squared error (RMSE) on the testing set from the baseline model is 0.31489320802667325 while the RMSE for the comparison model is 0.3241052888644588. In this case, we conclude that the two models perform equally the same.

Alternatively, we use the automated machine library TPOT to find a suitable model. This process may be beneficial especially when there are no time constraints in releasing a pipeline. Using the TPOTRegressor from TPOT, the best pipeline is as follows:

1. SelectPercentile: Keep eighty-four percent (84%) of the features based on percentile scores
2. StandardScaler: Standardize features by removing the mean and scaling to unit variance.
3. Binarizer: Binarize features with 0.55 as threshold.
4. RobustScaler: Scale features that are robust to outliers.
5. MaxAbsScaler: Scale each feature by its maximum absolute value.
6. LassoLarsCV: Cross-validated L1 regularization using the least-angle regression algorithm

The root mean square error of the pipeline is 0.3135939840712398, which performs the same as the initial models.

Conclusion

Therefore, we choose any model for deployment and observing more differences once new data are obtained. Note that any model is a good choice since the predictions on the holdout set roughly off by one (1) enrollee.

(The associated codes and implementations in this project is located in the Jupyter notebook.)

Background and Objective

A start-up company wants to automate loan approvals by building a classifier to predict whether a loan will be paid back. In this situation, it is more important to accurately predict whether a loan will not be paid back rather than if a loan is paid back. Your manager will want to know how you accounted for this in training and evaluation your model. As a machine learning scientist, we need to build the classifier and prepare a report accessible to a broad audience.

Executive Summary

To determine whether a loan will be paid back, we analyze the important features of a borrower, such as the credit score, the number of public derogatory records, and others. In this scenario, we provide a machine learning model for the classification. The report is divided into the following sections:

1. Introduction: Overview of loan repayment and factors affecting the failure of non-repayment.
2. Data Preprocessing: Detailed explanation on data cleaning.
3. Feature Engineering: Extract or select features to train in the classifier.
4. Classifier: Model to classify loan repayments.
5. Conclusion: Final remarks obtained from the model experiments.

Introduction

The repayment rate of loans is influenced by a variety of factors, including the borrower's creditworthiness, the type of loan, and the economic climate. Studies found that borrowers with higher credit scores are more likely to repay their loans on time and in full. Furthermorele, student loans are typically repaid at a lower rate than other types of loans, such as personal loans or auto loansLastinally, the economic climate can also affect the loan repayment rate. During economic downturns, borrowers may be more likely to default on their loans due to job loss or other financial difficult of defaults.

Data Preprocessing