Megan Doiron

Motivated and detail-oriented Data Scientist with a Master's Degree in Data Science and a solid foundation in statistical analysis, machine learning, and data visualization. Proficient in Python, SQL, and R with experience in applying supervised and unsupervised learning models to real world data. Knowledgeable in Natural Language Processing (NLP), with hands on experience using BERT models. Strong mathematical background in Statistics, calculus, and data-driven problem solving. A quick learner with excellent analytical thinking, time management, and team work skills.

Zero-Shot Text Classification
January 2022

• Conducted a research project to explore the workings of the zero-shot transformers pipeline on text data.
• Analyzed output probabilities indicating the relevance of text to specific keywords and compared the results against a baseline to evaluate accuracy.
• Developed proficiency in Python programming and utilized the Transformers package for implementation.

Anomaly Detection
January 2022

• Employed the XGBoost machine learning model to predict drops in value within a unique time series dataset.
• Conducted feature extraction (e.g., rolling standard deviation, rolling mean) due to the minimal data points available, enhancing the model's predictive capabilities.

Senior Capstone
January 2022

• Analyzed text data using zero-shot text classification, k-means text clustering, and masked language modeling.
• Gained a comprehensive understanding of the BERT model and the mathematics behind the skip-gram algorithm, which is foundational to BERT.
• Delivered a 30-minute public lecture on the mathematical principles underlying the various models.

Master’s Thesis
January 2023

• Classified text as cyberbullying or non-cyberbullying using zero-shot and few-shot learning techniques.
• Compared results against traditional supervised machine learning models to assess performance across different methodologies.
• Although zero-shot learning showed lower accuracy than supervised models, the results demonstrated impressive capabilities for a blind prediction approach.