Body Level Classification

Data Set:

Quick Statistics:

• 1477 data samples.
• 16 attributes (+1 for classes) [8 categorical and 8 numerical].
• 4 output classes.

Split:

• The data was split into 70% Training, 20% Validation & 10% Test. (1033, 296 & 148 data samples respectively).
• They are available in the data directory in the train.csv, val.csv and test.csv files respectively.

Attributes Description:

Feature	Description
Gender	Male or female.
Age	Numeric value.
Height	Numeric value (in meters).
Weight	Numeric value (in kilograms).
Fam_Hist	Does the family have a history with obesity?
H_Cal_Consump	High caloric food consumption.
Veg_Consump	Frequency of vegetables consumption.
Meal_Count	Average number of meals per day.
Food_Between_Meals	Frequency of eating between meals.
Smoking	Is the person smoking?
Water_Consump	Frequency of water consumption.
H_Cal_Burn	Does the body have high calories burn rate?
Phys_Act	How often does the person do physical activities?
Time_E_Dev	How much time does person spend on electronic devices.
Alcohol_Consump	Frequency of alcohols consumption.
Transport	Which transports does the person usually use?
Body_Level	Class of human body level.

Basic Analysis & Visualization:

• Basic Data Analysis was performed in analyze.ipynb. Also, detected some outliers and used a baseline measure: ZeroR which resulted in 47.297% accuracy and macro F1 of 0.1605.
• Used multiple ways to visualize the data in visualize_data.ipynb as bar charts, histograms, scatter, boxplots and a heatmap.

Class Imbalance Analysis

• Analyzed the class imbalance and found that Body Levels 1 -> 4 are 0.137464, 0.130687, 0.274927, and 0.456922.
• Analyzed the features & their correlation/connection to the Body_Level and found that all the features are at least statistically related to Body_Level and after further investigation, the Weight feature the most correlated (0.85) to the Body_Level, followed by Fam_Hist, and Age.
• Attempted to solve the class Imbalance by both over-sampling & under-sampling.
• Attempted to solve the class Imbalance using synthetic data generation once without using the relationship between features & once while taking them into account.

Preprocssing:

• Converted all the categorical data to numerical data by converting them to one hot encoding vectors manually in all_numerical_data.ipynb.
• To use it, simply write %run "../all_numerical_data.ipynb" at the beginning of your notebook and then directly use x_train, y_train, x_val and y_val afterwards.

Models Trained:

SVM

Model	Accuracy	Weighted F1-Score
Linear SVM	69.90 %	63.00%
Linear SVM with C=0.1	76.35%	74.00%
Linear SVM with C=0.03	80.00%	80.00%
Linear SVM with C=0.04	80.00%	80.00%
Linear SVM with C=0.03 , tol=0.001	80.04%	78.00%
Non Linear SVM with kernel =rbf	76.35%	75.00%
Non Linear SVM with kernel =rbf, gamma= 0.008	80.07%	79.00%
Non Linear SVM with kernel =rbf, gamma= 0.05	90.08%	90.00%
Non Linear SVM with kernel =rbf, gamma= 0.001	77.02%	76.00%
Non Linear SVM with kernel =rbf, gamma= 0.05 , C=2	92.22%	92.00%
Non Linear SVM with kernel =rbf, gamma= 0.05 , C=3	93.24%	93.00%
Non Linear SVM with kernel =linear	91.89%	92.00%
Non Linear SVM with kernel =linear with C=2	93.58%	93.00%
Non Linear SVM with kernel =linear with C=2, gamma=0.05	93.58%	93.00%
Non Linear SVM with kernel = Sigmoid	25.33%	19.00%
Non Linear SVM with kernel = Poly	72.97%	72.00%
Non Linear SVM with kernel = Poly, gamma=0.05	97.29%	97.00%
Non Linear SVM with kernel = Poly, gamma=0.05, C=2	97.29%	97.00%

Logistic Regression

Penalty	Solver	C	Accuracy	Macro F1-Score
'l2'	'liblinear'	1.0 (default)	84.79%	78.34%
'l2'	'lbfgs'	1.0 (default)	83.78%	77.44%
'l2'	'newton-cg'	1.0 (default)	84.12%	77.71%
'l2'	'sag'	1.0 (default)	83.11%	77.17%
'l2'	'saga'	1.0 (default)	81.76%	74.96%
'l1'	'liblinear'	1.0 (default)	88.51%	83.16%
'l1'	'liblinear'	100000.0	90.54%	86.10%
'l2'	'liblinear'	0.0001	51.01%	24.12%
'l2'	'liblinear'	100000.0	90.54%	86.24%

Gaussian Mixture

n_components	covariance_type	reg_covar	n_init	Accuracy	F1-Score
1	'full’	0.0001	1	86.00%	86.00%
1	‘full’	0.0001	4	86.00%	86.00%
7	‘full’	0.0001	1	72.00%	69.00%
2	‘full’	0.0001	1	72.00%	67.00%
3	‘full’	0.0001	1	75.00%	71.00%
1	'full’	0.0001	10	86.00%	86.00%
1	‘tied’	0.0001	1	86.00%	86.00%
1	‘diag’	0.0001	1	67.00%	62.00%
1	‘spherical’	0.0001	1	47.00%	30.00%

Multi-class Random Forest

n_estimators	max_features	criterion	Accuracy	Weighted F1-Score
100	‘sqrt’	‘gini’	94.59%	95.00%
1000	None	‘gini’	96.95%	97.00%
1000	None	"entropy"	97.29%	97.00%

Multinomial Naïve Bayes

Data Type	Accuracy	Weighted F1-Score
Numerical Data to Categorical in 10 bins	63.17%	62.00%
Numerical Data to Categorical in 16 bins	64.52%	65.00%
Numerical Data to Categorical in 30 bins	64.18%	65.00%
All numerical data (1 hot coded vectors)	70.61%	70.00%
All data as is	70.95%	70.00%

K Nearest Neighbors

Dataset Type	Accuracy
Original Dataset	77%
Undersampled	74%
Oversampled	77%
Smote	77%
Smote with Importance	77%

Bagged Decision Trees

• we did learning curve analysis and bias variance analysis

Boosting Decision Trees ( Best Model )

It gets 99% accuracy & 99% Weighted F1-Score

• Best max_depth: 5
• Best n_estimators: 100
• Best learning_rate: 1.0

Collaborators

Abdelrahman Jamal

Iten Elhak

Radwa Ahmed

Ahmed Jamal