Australian Rain Predictor: Project Overview

Created a tool that predicts if there will be rainfall in Australia (Accuracy ~ .87; AUC ~ .87) to communicate weather patterns and tendencies
Cleaned over 150,000 weather documentations from various cities across Australian
Built Logistic Regression, Random Forest, and SVM Classifier models
Tuned models by testing the efficacity of Normalizing and Resampling variables; Normalizing improved model performance and Resampling was deleterious
Optimized Logistic Regression and Random Forest Classifiers using Pipeline and GridsearchCV to reach the best model

Code and Resources Used

Python Version: 3.9
Packages: pandas, numpy, sklearn, seaborn, pickle
Data Source: https://www.kaggle.com/jsphyg/weather-dataset-rattle-package

Began cleaning and feature selection through the following changes:

Calculated the percent of missing data values for each column; removed columns where nearly 40% of the data was missing
Dropped rows in which data values for the RainTomorrow column were missing because they were essential for model testing
Filled missing values in categorical columns with column averages; filled missing values in categorical columns with "Unknown"
Determined and removed outliers in Rainfall column
Parsed out month from Date column
Label Encoded Month, Location, WindGustDir, WindDir9am, WindDir3pm, RainToday, and RainTomorrow columns

I built histograms and bar charts to help understand our data and note any immediate trends. I also made a heatmap to make correlations limpid.

The first thing I did was split the data into train and test sets with a test size of 20%.

I used three different models and evaluated them with Accuracy and Area Under Curve. I trained the models on Normalized variables.

The models I tried:

Logistic Regression: Benchmark for the model
Random Forest Classifier: I thought this would be a good model to try because of the many categorical variables
Support Vector Machine: I wanted to try an SVM because this was a binary classification problem and frankly I was just curious because SVMs are such a classic classification algorithm

The Random Forest model performed the best on the test and validation sets.