Binary Classification Neural Network (From Scratch)

This project aimed to build a foundational understanding of neural networks by implementing a binary classifier from scratch without using deep learning frameworks. The dataset contained customer attributes such as age and affordability score, with the target variable indicating whether the customer purchased insurance.

While libraries such as TensorFlow or PyTorch provide high-level abstractions, building a network manually was necessary to fully understand the underlying mechanics of forward propagation, gradient computation, and parameter updates.

I designed and implemented a custom neural network class in Python that takes two input features and predicts a binary output. The model uses a sigmoid activation function for probability estimation and optimizes its weights and bias through gradient descent while minimizing log-loss as the objective function. To prevent unnecessary training once convergence was reached, I implemented an early stopping mechanism triggered when the improvement in loss became negligible.