Inspiration
After learning that suicide was a leading cause of death with over 800,000 instances per year worldwide we decided to investigate what are its causes and how we could make individual recommendations to ameliorate this problems
Challenges we ran into
The datasets on the servers provided in the competition were all aggregated data, which due to the Law of Large Numbers, makes them suitable for comparison but unsuitable for prediction. We needed to dive deeper!
What it does and how we built it
After searching the web for individualized suicide data, we found a dataset on Kaggle. As the 1st part of our project, we built a prediction system. We built a neural network using PyTorch and trained in on the dataset to predict whether an individual would commit suicide, with an accuracy of 83% on the validation data withheld from the model training, using 5-fold validation. In the 2nd part, we did the individual recommendations. On average, we found the average of 89% of probability of non-suicide.
Accomplishments that I'm proud of
After accounting for the error in our model, our recommendation system still produces a recommendation to on average reduce suicide by 74%. If everyone could provide all of the data and implement all of our recommendations, we could reduce global suicides by approximately 590,000 per year.
What I learned
We got experience with optimization using a Bayesian Optimizer framework
What's next for our project
Dig into a bigger dataset, such as the VA suicide dataset, to try to increase the accuracy even further. Also look at other ways of processing the columns that are currently in 0th Normal Form.
