Leveraging Natural Language Processing for Predictive Modeling

Abstract

Weather analysis is the process of examining parameters like atmospheric pressure, humidity, and wind speed, to make calculations and predictions regarding the current weather. Using the publicly available dataset made available by CrowdFlower on Kaggle, this research performs operations of natural language processing and predictive modeling on the provided data to determine the sentiment of a tweet, and predict the type of weather being mentioned in the tweet. We first perform the data-pre-processing operations of label engineering, feature extraction, and data visualization on the available data. Regression approaches including Gradient Boosting, Ridge, and Linear regressor, and classification approaches including Random Forest, Multinomial Naïve Bayes, Linear Support Vector, and Stochastic Gradient Descent classifier were then applied for data modeling. The results of the current research indicate that feature reduction has a positive effect on overall predictive accuracy. The results also indicate that predictive performance is higher when performing classification tasks as compared to regression tasks.