Trello: https://trello.com/b/kGRYd7Zx/weatherml
Dataset: https://docs.google.com/spreadsheets/d/1wjoOM3OyRlOUdET7_jU2uoCyOraQOWY8XGOw8LVWK3A
One unique thing about me is that I love my black Adidas jacket and I wear it as much as I can. My default style is to leave it unzipped with rolled-up sleeves. However, the weather can change that. If it's chilly, I may roll down the sleeves or zip up. I may even need to wear a coat. If it's too hot, it can be uncomfortable to have it on at all.
The idea for this project started when I was a first-year college student. As someone living on-campus, I had to walk to my classes. In addition, as a student at Virginia Tech, I would be crossing the Drillfield every time, exposing me to the elements. As a result, I got in the habit of checking the weather every time when I had to go somewhere (I even ended up adding a temperature widget to my lock screen since I checked the weather so many times). The problem was that predicting my choice of outerwear wasn't straightforward. There would be days when I could go out with just a T-shirt when it was only 65. On others, I would need a coat even though it was 73.
Obviously, temperature is not the only factor in comfort. Wind, humidity, and whether the sun is shining can all affect how warm or cold I feel. However, it's very hard for me to make a prediction while juggling all the different factors. As a result, I thought about potentially coding some sort of AI that could read the weather and give me a prediction on what I should wear outside. Thus, the idea for WeatherML was born.
At the start, I didn't have any experience or knowledge to code a ML model. However, I knew I would need a lot of data, so I started collecting data right away. I included as many factors that I thought might be relevant, and tried to collect several times a day (morning, noon, evening). My source of data was just the Weather app on my iPhone. Although I've heard that it's not the most accurate source for weather information, it was the most convenient for me. Besides, I used the Weather app anyways to check the weather so I trusted that the ML model would be able to handle the inaccuracies. I record my data in this Google sheet and then download it as a CSV when I'm ready for another round of training.
I learned the basic concepts of AI models in a course I took at Virginia Tech: CS4804: Intro to Artificial Intelligence. However, I learned how to actually code them via LinkedIn Learning courses (check out my LinkedInLearning repository for notes and exercise files from courses I've taken. Most of the ML courses are under the Python folder).
I use pandas to import data and scikit-learn for dataset preparation and
training. The model I use is Multiclass Linear Regression. As of current (with
v1 of the dataset), I have a training accuracy of 73.47% and test accuracy of
81.82%. It's not terrible, but I definitely want to improve it (my goal is ~90%
for both is possible). The model itself will need to be tweaked, but the biggest
thing I'll need to do is collect a lot more data.
I don't want to just train my data, get a good model, and call it a day. This project has piqued my interest in different ways I can experiment with my data. Some of the questions I hope to answer soon include:
- How important is each data feature? Temperature obviously is a big factor, but what about smaller things like humidity and wind speed?
- Does the time of day matter, or is pure weather data sufficient for predicting what to wear
- How do the average values for each data feature differ between labels? What
was the average temp, humidity, wind, etc. for each label (
none,jacket-long,jacket-short,coat)?
Check out the experiment log for more info on what experiments I'm trying out with my data!