One successful Udacity Capstone project was developed by a student named Sarah for the Data Analyst Nanodegree. For her project, Sarah analyzed publicly available data on Airbnb listings in New York City to help potential hosts understand the Airbnb market and how they can maximize their profits. She obtained listing data for over 20,000 Airbnb listings in NYC from Inside Airbnb. She then cleaned the data, performed exploratory data analysis, and developed regression models to understand the key drivers of nightly rates and overall reviews for listings.
Some of her key findings included that neighborhood, number of bedrooms, number of bathrooms, and amenities like washing machines were significant predictors of nightly rates. She also found that number of beds, superhost status, and neighborhood significantly impacted overall review scores. To make recommendations to potential hosts, she built interactive maps and graphs that allowed a user to explore predicted rates and reviews based on listing attributes. She included detailed explanations of her data cleaning, exploration, and modeling process in an Jupyter Notebook. Her work provided valuable insights into the NYC Airbnb market and actionable recommendations for hosts.
Another successful Udacity Capstone project was completed by a student named John for the Machine Learning Engineer Nanodegree. For his project, John chose to tackle the problem of detecting toxic online comments. He obtained a large dataset of Wikipedia comments that were labeled as ‘toxic’ or ‘non-toxic’ by human evaluators. His goal was to develop machine learning models that could accurately detect toxic comments to help moderate online discussions.
He started by preprocessing the text data using techniques like removing punctuation, stopwords, stemming, and lemmatization. He then engineered various features from the text like bag-of-words, n-grams, TF-IDF, etc. He evaluated several classifiers like logistic regression, gradient boosting, and recurrent neural networks on this multi-class text classification problem. Through rigorous experimentation, he found that a bidirectional LSTM model achieved the best performance of over 90% accuracy on the held-out test set for detecting toxic comments.
He then explored model explanations techniques like LIME to gain insights into what factors most influenced each model’s predictions. He also discussed limitations of the current approach and ideas for future work like handling new or modified forms of toxic language. He developed aFlask API to deploy his best model and allow users to submit comments for prediction. His thorough end-to-end process and focus on real-world applicability of detecting online toxicity made his a standout capstone project.
Another impressive Udacity capstone project was completed by a student named Melissa for the Self-Driving Car Nanodegree. For her project, she worked to develop a path planning strategy for navigating complex intersections. She first analyzed real-world traffic data from various cities to understand intersection usage patterns and common safety issues. She then modeled intersections as graphs with nodes representing lanes and edges denoting possible vehicle movements between lanes.
She designed a graph search algorithm that incorporated traffic rules, turn restrictions, vehicle dynamics constraints, and aCost function prioritizing safety and smooth driving. She implemented this algorithm using a CARLA simulator for her self-driving car. Through rigorous testing in various simulated intersection scenarios, she fine-tuned her path planning strategy and cost function weights. Her approach demonstrated safe navigation through complex four-way intersections with turns, merges and lane changes.
To evaluate her solution, she recorded metrics like completion time, maximum acceleration/braking, and number of collisions over hundreds of trials. She found her approach safely navigated intersections over 97% of times, often performing comparably or better than human drivers based on metrics. She provided detailed documentation of her intersection modeling approach, path planning algorithm design, simulation setup and results. By focusing on a real-world self-driving challenge and thorough evaluation, her project served as an excellent capstone that could have applicability for autonomous vehicles.
These three student capstone projects demonstrate the high caliber of work that is often produced for Udacity Nanodegree programs. Each project focused on solving a meaningful real-world problem through end-to-end data analysis, machine learning modeling or technical application development cycles. The students exhibited strong programming and analysis skills through Python/ML code, rigorous testing and reporting of results. Their work also incorporated important considerations like safety, ethics and real-world deployment factors. Through ambitious yet executable scopes, these projects exemplify the applied, hands-on learning that the Udacity Capstone project is intended to assess.