Massachusetts Education Machine Learning Project
This project explores how qualities of Massachusetts' school districts impact students' access to higher education with visualizing and analyzing the correlations between variables, and testing and tuning multiple machine learning regression models on the data.
The setup
Our project explores access to higher education for students in Massachusetts public schools. Specifically, we
want to
better understand what qualities might make a school district have higher rates of students who go on to
attend
college or university.
The insights from this project can be useful for further research. If these decision makers understand the
characteristics that most impact access to higher education, they can form regulations and distribute
resources in a way most likely to create more equitable public education systems across the state.
Data was acquired from Mass DoE and combined and cleaned. When working with
our data and creating the project, we selected hypothesis tests about data behavior to see correlations and
better
decide on the final features. Further illustrating data relationships, we looked at heatmaps and scatterplots.
We built Ridge, Lasso, k-Nearest Neighbor, SVM, and Multiple Linear
Regression models. With our top 3 models, we fine-tuned our parameters to find the best performance. For
Ridge regression we tested different alphas, or complexities, of the model and for k-nearest neighbors we
could
test different distance metrics and the number of n_neighbors.
Results
Multiple Linear Regression
R-squared: 0.662030
Mean Squared Error: 0.008933
Ridge
R-squared: 0.664494
Mean Squared Error: 0.008868>
k-Nearest Neighbor
R-squared: 0.710751
Mean Squared Error: 0.007645
Although we continued with Multiple Linear, Ridge, and k-Nearest Neighbors models, the algorithm we recommend to be used for our predictive model is the Ridge Algorithm. In the end, hyperparameter tuning ended up most positively affecting the k-Nearest Neighbor algorithm but couldn't overcome the noticable overfitting to the training data. Our Ridge regression had a minor improvement to the composition of the training and testing scores, meaning that hypertuning our Ridge model is for the goal of avoiding overfitting to the training data.
Reflections
As a beginning Data Science project, this covers the fundamentals for working with models. I appreciate this project because, as we progressed, it showed clear growth in our understanding of the mechanics and uses of machine learning algorithms. Were we to go back, my group would be able to choose an even more useful question to ask in the scope of the available data.
