ocpodariu / udacity-mlnd

Machine Learning Engineer Nanodegree projects

Machine Learning Engineer Nanodegree

Projects developed as part of Udacity's Machine Learning Engineer Nanodegree.

Projects

0. Titanic Survival Exploration

• View Jupyter Notebook or Go to project directory
• Explored the dataset to identify which features best predict a passenger's survival
• Used those features to create decision functions to predict the survival of the passengers

1. Predicting Boston Housing Prices

• View Jupyter Notebook or Go to project directory
• Measured the linear correlation between features and selling price using Pearson's r
• Observed the effect of different training and testing splits on model performance
• Used learning and complexity curves to detect underfitting and overfitting
• Combined Grid Search with cross-validation to find the optimal maximum depth for a decision tree regressor
• Final model obtained an R^2 score of 0.77
• Discussed the model's applicability in a real-world scenario

2. SMS Spam Classification

• View Jupyter Notebook or Go to project directory
• Transformed the SMS messages using the Bag-of-Words model
• Generated features based on the frequency of each word
• Classified SMS messages as spam or not spam with a Naive Bayes model

3. Finding Donors for CharityML

• View Jupyter Notebook or Go to project directory
• Evaluated the performance of different supervised algorithms in identifying individuals making more than $50,000
• Preprocessed the data by scaling numerical features, one-hot encoding categorical features and applying logarithmic transformations on features with skewed distribution
• Built a pipeline to quickly evaluate the performance of different algorithms
• Analyzed AdaBoost's performance in relation to the maximum number of estimators
• Identified the top 5 most important features and analyzed the effects of feature selection on AdaBoost's performance

4. Creating Customer Segments

• View Jupyter Notebook or Go to project directory
• Applied PCA to identify customer spending patterns and to reduce the dimensionality of the data
• Compared K-means and Gaussian Mixture Model to decide which is more suitable for grouping customers into segments
• Performed a silhouette analysis to determine optimal number of components for Gaussian Mixture Model
• Designed an A/B test to measure the effect of a delivery service change on each customer segment
• Trained a classifier on customer segment data to label new customers based on their estimated spendings