Portfolio

🤖 Machine Learning

Summary: My implementation of classical ml-algorithms from scratch in NumPy.

Objective: To identify malicious(anomalous/fraud) transactions in a highly imbalanced data with a class ratio of 1000:17 (normal:malicious).
Designed a custom Deep Auto-Encoder model using TensorFlow which achieved:
1. A Recall score of 0.85 (successfully identifies 85 out of every 100 malicious transactions).
2. A 6.5% False Positive Rate (~6-7 transactions in every 100 transactions labeled as malicious are actually normal).

Objective: To label a news sample with one-of-four pre-defined News Topics(World, Sports, Business, Sci/Tech).
Designed custom LSTM and Bi-LSTM models which acheived performance scores of 91.44% and 92.34% Test Accuracies respectively.

Objective: To classify images into one-of-hundred pre-defined image classes using a small training dataset(500 images per class).
Implemented VGG16 and ResNet9 CNN(Convolutional Neural Network) models from scratch in PyTorch. ResNet9 performed better with a 74.32% Test Accuracy.
Applied Transfer Learning using EfficientNet which performed the best with 81.31% Test Accuracy. More details can be found below.

Objective: To perform a Statistical Analysis on the White Wine dataset, comprising 4898 wine samples with 12 chemical properties.
Conducted various hypothesis tests like Z-test, F-test, and T-test, which includes testing for legal limits, quality assurance, and business development.
Conducted Regression Analysis with - Model Adequacy Tests and Model Diagnostics. Finally reported the most important chemical properties contributing to the quality of the wine.