Computational project · Summer 2026
Forest cover type classification.
A CS 7641 Machine Learning project comparing classification, optimization, and representation-learning approaches on forest cover data. The summary below is based on my supplied CV; the full experiment report is not available publically but can be provided on request.
- Project type
- Academic computational project
- My role
- Benchmarking, preprocessing & evaluation
- Status
- Completed
Learning to distinguish forest cover types.
This project addresses forest cover type classification. It is included as computational preparation alongside my research interests in human-centered AI and NLP.
I used a stratified 20,000-record Covertype sample with 54 features and seven classes. The dataset citation, version, and sampling seed will be added with the report.
Make the experiment inspectable.
The experimental setup reported in my CV is summarized below. Exact hyperparameters, random seeds, environment details, and the full evaluation record will accompany the report.
| Component | Details to document |
|---|---|
| Data & split | Stratified 20,000-record sample; 54 features; seven classes; 64/16/20 train/validation/test split. |
| Preprocessing | Preprocessing fitted exclusively on training data to prevent leakage. Individual transforms await the report. |
| Algorithms | Decision Trees, kNN, SVMs, and neural networks; Decision Trees had the strongest supervised-learning performance.> |
| Evaluation | Macro-F1, balanced accuracy, and confusion matrices examined performance under class imbalance. |
| Reproducibility | Report and code has not been publically released; environment, seeds, and execution steps can be provided upon request. |
Attribution matters.
I benchmarked classifiers, fitted training-only preprocessing, and evaluated performance. I compared SGD, Adam-family optimizers, and randomized final-layer optimization under controlled compute budgets, examining learning rate, momentum, and regularization. I also evaluated K-Means and Gaussian Mixture Models alongside PCA, ICA, and randomized projections, measuring cluster quality, label alignment, and downstream neural-network performance.
Evidence before numbers.
The comparison supports the importance of regularization in this experiment.
Report & code.
Available upon request.