SC1015_Project Suicide Prediction

File Descriptions

SC1015_Project_SuicideRisk.ipynb

• A Jupyter Notebook containing the source code for the project.

SC1015 Presentation.pptx

• Presentation Slides

README.txt (this file)

• Basic information about the project repository.

SC1015_Transcript.pdf

• A PDF copy of the transcript for the video presentation

continents2.csv

• Country mapping Dataset

master.csv

• Suicide Rates from 1985-2021 Dataset

About

Our project aims to explore key indicators of suicide and develop models that can help identify and predict those at risk of suicide. This Readme serves as a overview of the project, more details can be found in presentation and notebook.

Package requirements for running notebook

• geopandas
• graphviz

Contributions

• Zj12248 - Data extraction, Data Preprocessing, Data visualisation, Random Forest, XGBoost
• Vithun - Classification Tree, Presentation Structure, Problem Ideation

Introduction

Datasets from

• https://www.kaggle.com/datasets/omkargowda/suicide-rates-overview-1985-to-2021 (Suicide Data)
• https://www.kaggle.com/datasets/andradaolteanu/country-mapping-iso-continent-region (Country-region Mapping)

Problem Definintion

• Are we able to predict if a person is of high or low suicide risk?
• Which model would be the best to predict it?

Data Preprocessing

• Exploratory Data analysis performed on numeric and categorical varables against suicide/100k
• Data cleaning(), normalisation(MinMaxScaler), transformation performed

Methodology

Models Used

1. Decision Tree
2. Random Forest
3. XGBoost

Train-Test Split - train model using Train data, backtest against Test data.

Experiments

Baseline: Decision Tree
Details in presentation.

Performance Measurement through classfication accuracy - score() function, True Positive Rate and False Positive Rate

Conclusion

• Random Forest is the optimal model among the 3 to predict suicide risk - approximately 93% accuracy, with 95% True Positive Rate, 9%False Postive Rate, thus is a good predictor of suicide risk.

• NOTE: Lower False positive is better to minimize the waste of resources, but it is NOT more important than a person of suicidal risk missing out on aid(TPR), thus rule out XGBoost for this problem.

• GDP per capita have extremely weak linear correlation values with suicide rates(contrary to instinctual belief), however random forest model does list it as the most important feature

• Europeans, Males, and older people are characteristics of those of the highest suicide risk.

Possible Improvements

• Hyperparmeters matters - XGBoost should not have performed worse than random forest if hyperparmeter tuning was performed as XGBoost has been proven to be better than RF.
• Perform Data Augmentation(Oversampling/Undersampling) to balance imbalanced target data(suicide_risk)

What did we learn from this project

• one hot encoding, min-max scaling
• importance features
• Random Forest, XGBoost - machine learning models
• geopandas , graphviz

References

• https://machinelearningmastery.com/why-one-hot-encode-data-in-machine-learning/
• https://towardsdatascience.com/xgboost-theory-and-practice-fb8912930ad6
• https://medium.com/@gabrieltseng/gradient-boosting-and-xgboost-c306c1bcfaf5
• https://towardsdatascience.com/beginners-guide-to-xgboost-for-classification-problems-50f75aac5390
• https://towardsdatascience.com/understanding-random-forest-58381e0602d2
• https://machinelearningmastery.com/implement-random-forest-scratch-python/
• https://towardsdatascience.com/everything-you-need-to-know-about-min-max-normalization-in-python-b79592732b79
• https://towardsdatascience.com/having-an-imbalanced-dataset-here-is-how-you-can-solve-it-1640568947eb