SC1015 Data Science Project - Improving your Life Expectancy
❓ About
This is a Mini-Project for SC1015 (Introduction to Data Science and Artificial Intelligence) which focuses on countries' Life Expectancies from Kaggle's Life Expectancy Dataset
🧑🏽🤝🧑🏽 Contributors
- Chong Wei Kang: Problem Statement, Motivation, Data Preparation
- Joel Tan: Deep Learning with TensorFlow, Conclusion
- Sua Qi Rong: EDA, Multi-variate Linear Regression, Random Forest Regression
However, overall, we helped each other for the whole project and our contributions were not limited to the above.
🔎 Problem Definition
- What should we do to effectively increase the life expectancy of Singapore's population in today's context?
- What are some main areas of concern to prioritise and tackle in order to become the country with the highest life expectancy globally?
💪 Motivation
- The rate at which Singapore's Life Expectancy is increasing in the past decade has slowed significantly.
- Singapore is currently the 5th ranked country globally with a life expectancy of 84.07 years
- By 2025, we hope to see Singapore as the country with the highest life expectancy of 88 years!
🚀 Models used
- Multi-variate Linear Regression (SKLearn)
- Random Forest Regression (SKLearn)
- Deep Learning Neural Network (Tensorflow)
🚶 Steps we took
1. Data Preparation
- Dropping Data: Dropped "Year" and "Country" - Correcting Variable names:
Some were wrongly written and had weird spaces - Addressing NA Values:
Filling NA values with the median of the original data - Removing outliers:
Data points which are +- 1.5 IQR from Q1 and Q3 were considered to be
outliers - Encoding of Category Variables: Label Encoder to encode "Status"
into 0 and 1. 0 represents Developed Countries while 1 represents Developing
Countries
2. Data Visualisation & Exploratory Data Analysis
- Plotted the distribution of all variables using a boxplot, histogram and
violinplot for all 20 variables - Scatterplot of all predictor variables
against Life Expectancy - Ended off by plotting a Heatmap to show the
correlation of all variables - Generate Data-driven insights
3. SKLearn Multi-Variate Linear Regression
- Ran a train-test split of 80-20 - Fitted the model and plotted a scatter
plot of Life Expectancy against the predictors - Obtained an initial MSE
score of 13.49 - To deal with issues such as overfitting and selection bias,
we made use of 10-Fold Cross Validaton and obtained a final MSE score of
12.63.
4. SKLearn Random Forest Regression
- Again, we ran a train-test split of 80-20 - Made use of GridSearch to
determine the best number of estimators from 1 to 101. - Obtained best
number of estimators as 101. - Obtained an initial MSE score of 2.85 which
seemed too low and suggested the presence of overfitting. - Thus, we again
made use of 10-Fold Cross Validaton and obtained a final MSE score of 3.30.
5. Deep Learning Regression using TensorFlow
- To prepare the data, we had to remove all spaces and replace them with
dashes. - Ran a train-test-split of 80 to 20 and built the model by stacking
layers using rectified linear unit as the activation. - Used Adam’s
stochastic gradient descent as the optimizer, and MSE for our loss function.
- Implemented an early stopping function to ensure that our model do not
overfit with a patience of 2. - Used 512 epochs for our model so that
runtime will not be too long.
💡 Conclusion & Recommendations
From our 3 models, Random Forest regression has the lowest MSE of 3.30.
| Model | Minimum MSE (2 d.p) |
|---|---|
| SKLearn Multi-Variate Regression with Cross Validation | 12.63 |
| Random Forest Regression with Cross Validation | 3.30 |
| Multi-Variate Regression with TensorFlow | 6.45 |
Random Forest Regression maybe more effective due to the size of the dataset. For Deep Networks, a large dataset would make the model more accurate. However, this dataset only has 2578 rows of data, further split into a 80-20 train-test split. Hence Random Forest may have performed better than the Deep Learning Model due to this reason.
Furthermore, the Deep Learning Model used a non-linear regression model, and it performed better than the linear regression model from SKLearn, suggesting that the relation between X variables and Life Expectancy were not linear to begin with.
From the Random Forest Regression, we identified the most important factors that the decision tree used to sift the information. From the tree, we identified that Adult Mortality, Income Composition of Resources and Schooling were the most important factors.
With the 3 factors, we made comparisons with the countries with the highest life expectancies as shown below. .The blue line represents Singapore's statistics of the 3 variables over the years, while the red line represents the average of each variable for these top countries.
From Schooling, we can see that Singapore is very much under average of what the top countries had.
We inputted these variables and found that just by slightly increasing schooling, income composition of resources and decreasing adult mortality, there is improvement in life expectancy with schooling being the most significant.
Hence some of our recommendations for singapore are to invest more funds to subsidize higher education to increase schooling and provide better incentives for individuals to lead healthier lifestyle.
📚 New Content Learnt
- Deep Learning using TensorFlow and Keras library
- Random Forest Regression and determining best number of estimators
- Various ways to prevent the issue of overfitting (Cross Validation and EarlyStopping in Keras Callbacks)
- Feature Importance using Random Forest
- Git and Github Usage