This is the Gitlab page for my work in Introduction to Data Mining.

My name is Qixiao Zhu and you can call me Jimmy. I am double majoring in Computer Science and Music with a piano performance concentration. I love to code and learn about computer science. At the same time, I do a lot of projects related to augmented reality, audio processing, and music software development.

I wish to learn a lot about data science in this course, and I am really looking forward to learning different models of machine learning. I am extremely passionate about machine learning and data science and want to get more familiar with them.

Final Report: https://github.com/JimmyZQX/csci349_2023sp/blob/main/project/Final_Report.ipynb

Presentation: https://youtu.be/w0oFfmKs6rs

The final project uses the dataset "Glass Identification Data Set" from the UCI Machine Learning Repository: http://archive.ics.uci.edu/ml/datasets/Glass+Identification to hone and test my skills in data mining and machine learning through predicting the type of glass each instance represents using neural network.

This dataset is hard to deal with because it has imbalanced classes and it does not have a lot of samples. Because of this characteristic, I tried multiple methods to train decent neural network classification models.

In my later experiments with different methods of coping with this imbalanced dataset, I have found that this dataset was unexpectedly hard to fit a classification neural network. Different classes had different numbers of samples. The largest class had 76 elements, while the smallest class only had 9 elements. At first, I chose to simply use the weighted f1 score as the performance metric, but it only had a score of 70% when using cross validation. I then tried to use an oversampling technique called SMOTE, which generates samples for the smaller classes so that each class had the same number of samples, and I used accuracy for the performance metric after oversampling. It resulted in much better scores, with about 90% accuracy when doing cross validation. After I tuned the hyperparamenters, I tried to split the original data, train with oversampled data, and test with the portion that is not oversampled. Train test split test resulted in about 75% accuracy. I had to adjust the number of epochs a little to prevent overfitting on the oversampled training data. The result was not a particularly satisfying one, but it was the first time I have dealt with imbalanced data and the progress from not even close to accurate to somewhat accurate predictions was satisfying for me.