Objectives
-
Become familiarized with classification problems in machine learning
-
Build the model from scratch, ensuring a profound understanding of the architecture
-
Implement ten-fold cross-validation
-
Not allowed to use any libraries with addidional ML tools
-
First Classification Task: Batteries - defective or not?
-
Second Classification Task: Does the patient have chronic kidney disease (CKD) or not?
Methods/Features/Results
-
Normalized the values of the features in the dataset by z-score
-
Experimented with different learning rates, and decay
-
No test set - accuracies only compared with ten-fold cross-validation
-
Experimented with adding fictitious features via quadratic and cubic terms
-
Experimented with removing features that were measurably less significant
-
Accuracy on cross-validation could not exceed 89% on the CKD dataset, and 70% on battery dataset
-
Perhaps a more complex model with a nonlinear decision boundary may fare better
-
This work is associated to a project submitted for the McGill course: ECSE 551