Predicting Depression based on Tweets

Inspiration

I was thinking about problems that I faced in my daily life, and I realized that many of my friends struggled with their mental health, and that mental health issues have become an epidemic recently. Also, I realized that people may not always realize they are struggling. I thought creating a tool that would help identify when people are struggling might be helpful in combating the mental health epidemic.

What it does

Using the BERT model, I performed sentiment analysis on a dataset of tweets labeled by whether the user has depression or not. At the end, there is a function where the user can input a sentence to try it out.

How we built it

I built it on Google Colab using the BERT NLP model.

Challenges we ran into

The model did not work very well. This is because people with depression are not their illness and therefore will not tweet "depressed-sounding" things all the time. It would have been better to analyze a person's entire twitter history rather than just one tweet, but when I tried this, the model memorized the data instead of learning it.

Accomplishments that we're proud of

I am proud of being able to build a somewhat-working machine learning model using the BERT model.

What we learned

I learned that models can have biases based on what data you feed it. My model was biased, because the data was not labeled in the way I wanted it to, so I had to work with what I had.

What's next for Predicting Depression based on Tweets

I would like to be able to use a larger dataset of tweets with better labeling. Possibly, instead of predicting depression, the model might predict whether a tweet is concerning or not concerning.

Built With

• ipynb
• python