aa 10% 12%) 16% image source: YouTube Mario Tomic (https://www.youtube.com/c/MarioTomicOfficial) Body Fat Level Predictions - Machine Learning vs. US Navy Formula instead of the criginal cateset, we are gonna use the data generated with my first notebook (https://www.kaggle.com/code/alexandrepetit881234/bodyfat-visualization) in adddition to the criginel cata, 2 few more column were created: BMI, 2 column with the individual Body Mass Index LeenPercent, the percent of lean tissues FatFreeMass, the lean mass in ko FFM, the fat-free mass index, Similar to BMI, but for the lean mass Overweight, BMI over 25 Obese, BMI over 30 HighFat, body-fat above 22 percentIn [1] import numpy as np # linear algebra import pandas as pd # data processing, CSV file 1/0 (e.g. pd.read_csv) import os for dirname, ., filenames in os.walk(' /kaggle/input') for filename in filenames: join(dirname, filename) ) import matplotlib.pyplot as plt import seaborn as sns ‘matplotlib inline /kaggle/input /body-fat-prediction-dataset /bodyfat .csv /kaggle/input /bodyfat-prediction-cleaned-extended/BodyFatCleaned.cs v 1.Data Preparation In [2] df = pd.read_csv(' /kaggle/input /bodyfat-prediction-cleaned-extended/BodyFatClean ed.csv') df.drop(‘Unnamed: 8°, axis df .head() , inplace=True) out [2] Density BocyFat Age Weight Height Neck Chest Abconen Hip Thigh 1 (1.0853 61 22 78.8 1.8 38.5 93.6 832.0 98.7 58.7 2 10414 25.3 22 70.0 1.7 34.0 95.8 87.9 99.2 58.6 3 1.0751 10.4 26 84.0 1.8 37.4 101.8 864 101.2 60.1 4 1.0340 28.7 24 82.8 1.8 34.4 97.3 100.0 101.8 63.2 es , 5 rows x 22 columnslet's add @ new column for people under 10% body-fat, it will be used to compare the density besec body-fat percent with the US navy formula fat percent df[‘LowFat'] = @ lf = 18 df[‘LowFat'][df['BodyFat'] < 1f] = 1 /opt/conda/1ib/python3 .7/site-packages/ipykernel_launcher.py:5: Set tingWithCopyWarning: A value is trying to be set on a copy of a slice from a DataFrame See the caveats in the documentation: https: //pandas.pydata.org/pan das-docs/stable/user_guide/indexing.html#returning-a-view-versus-a- copy This formula is used by the US navy to estimate the body-fat percent, let's see how it compare with the density based bocy-fat calculation %BF = 495 | ( 1.0324 - 0.18077 iagi0( waist - neck } + 0.15456 log‘0( height } ) - 450 In [4 df['NavyFat'] = 495 / (1.0324 - 6.19077 * np.logi@(df['Abdomen’] - df['Neck']) + @.15456 * np.logi@(df['Height’] * 190) - 450In [5] df['NavyFat'] = round(df['NavyFat'], 1) df.head() out[5] Density BoayFat Age Weight Heigrt Neck Chest Abcomen Hip Thigh © 1.0708 123 23 70.1 17 36.2 93.1 85.2 94.5 58.0 1 1.0853 6.1 22 78.8 1.8 385 93.6 83.0 98,7 58.7 wo 2 10414 25.3 22 70.0 1.7 34.0 95.8 87.9 98,2 58.6 we 3 1.0751 10.4 26 84.0 1.8 37.4 101.8 864 101.2 60.1 we 4 1.0340 28.7 24 83.8 1.8 34.4 97.3 100.0 101.8 63.2 we Po , 5 rows x 24 columns .use('ggplot') ts(1,3, figsize=(12, 4), sharey-True) sns.scatterplot(ax=ax[@], data=df, x='BodyFat', y='NavyFat', hue='LowFat') (ax=ax[1], date=df, BMI', y="BodyFat', hue='LowFat') (ax=ax[2], data=df, x='BMI', y='NavyFat', ‘LowFat' ) plt.ylabel('FatPercent') odyFat an ew!In [7] df[[‘BodyFat', ‘NavyFat', ‘Abdomen']][df['BodyFat'] < 10]out [7] BcdyFat 61 41 ma 78 37 78 37 a8 87 77 56 4.0 66 8.0 6.3 3.8 6.3 88 85 88 8.3 85 9.0 9.6 9.4 53 94 94 3.0 9g 86 66 6.0 96 7 75 5.2 NavyFat 14.1 14.0 13.0 178 137 105 114 176 16.3 158 145 88 18.1 16.6 14.1 118 12.8 18.5 18.7 121 16.8 15.0 178 21.3 WW? nS 13.8 185 16.1 148 14.0 10.7 19.2 168 18.1 14.2 18.3 Abcomen 83.0 82.5 83.6 90.9 79.7 748 73.9 835 84.5 76.0 79.5 70.4 778 82.0 79.8 778 78,4 82.8 82.9 788 86.0 83,1 88.8 916 74 765 778 812 819 79.4 83,7 78.0 89.1 83.9 79.4 815 82.8In [8 df[df['NavyFat'] < 18] out [a Dersity BooyFet_ Age Weight Height Neck Chest Abdomen Hip Thigh 47 1,0903 4.0 47° 580 17 340 834 70.4 87.2 506 150 1.0758 101 27 664 18 347 885 728 911 536 4 > with the navy body fat formule, only two people have an estimated bofy fat level under 10%. Much less than the people under 10% body-fat calculated with the density 2. Machine Learning Model we will use the measurements of the different bocy pert to estimate the body-fat percent of en individual, Because the density is used to calculate the BodyFat column, we will not keep it with the training date In [9 df.columns Out [9. Index(['Density', ‘BodyFat’, ‘Age’, ‘Weight’, ‘Height’, ‘Neck", ‘Ch est’, "Abdomen', ‘Hip’, ‘Thigh’, ‘Knee’, ‘Ankle’, ‘Biceps’, ‘Forea rm ‘Wrist’, ‘BMI’, ‘LeanPercent', ‘FatFreeMass’, ‘FFMI', ‘Overw eight’, “Obese”, ‘HighFat', ‘LowFat’, ‘NavyFat'], dtype=' object’) Data Selection and SeparationIn [10] x df[['Age', ‘Weight’, ‘Height’, ‘Neck’, ‘Chest’ ‘abdomen’, ‘Hip’, ‘Thigh’, ‘Knee’, ‘Ankle’, ‘Biceps’, ‘Forearm’ ‘Wrist’ ]]) y-b = df[‘BodyFat’] y_n = df[‘NavyFat’] In [11] from sklearn.model_selection import train_test_split X_train, X test, y-train_b, y_test_b, y_train_n, y_test_n = train_test_split(x, y-b, yn, test_ size=0.25, rando m_state=42) In [12] int(X_train. print (y_train_t (186, 13) (62, 13) (186,) (62,) (186,) (62,) For regression problem, we cannot use the accuracy score. We are going to use the mean squared error In [13] from sklearn.metrics import mean_squared_error from time import perf_counter Linear RegressionIn [14 from sklearn.linear_model import LinearRegression start = perf_counter() reg = LinearRegression( ) reg.fit(X_train, y_train_t) y-pred_b = reg.p’ (X_test) reg_mse_b = mean_squared_error(y_test_b, y_pred_b) print(f'the mean squared error for linear regression on density body-fat is {rou nd(reg_mse_t, 2) }") reg.fit(X_train, y_train_n) y-pred_n = reg.predict(X_test) reg_mse_n = mean_squared_error(y_test_n, y_pred_n) print(f'the mean squared error for linear regression on navy body-fat is {round( reg_mse_n, 2)}") end = perf_counter() elapsed_time = round(end - start, 3) print(f'it took {elapsed_time} second for predicting the body fat percent using linear regression’) the mean squared error for linear regression on density body-fat is 19.45 the mean squared error for linear regression on navy body-fat is @ 15 it took @.029 second for predicting the body fat percent using line ar regression Decision TreeIn [15] from sklearn.tree import DecisionTreeRegressor start = perf_counter() dtr = DecisionTreeRegressor() dtr.fit(X_trair, y_train_b) y-pred_b = dtr.p’ (X_test) tree_mse_b = mean_squared_error(y_test_b, y_pred_b) print(f'the mean squared error for decision tree on density body-fat is {round(t ree_mse_b, 2) }') dtr.fit(X_trair, y_train_b) y-pred_n = dtr.predict(X_test) tree_mse_n = mean_squared_error(y_test_n, y_pred_n) print(f'the mean squared error for decision tree on navy body-fat is {round(tree _mse_n, 2)}") end = perf_counter() elapsed_time = round(end - start, 3) print(f'it took {elapsed_time} second to predict the body fat percent using a de cision tree’) ‘the mean squared error for decision tree on density body-fat is 40. A2 the mean squared error for decision tree on navy body-fat is 33.97 it took 6.814 second to predict the body fat percent using a decisi on tree Lineer regression gives us presictions closer to the velue we celculete before for both the density based body-fat anc the navy formula based body-fat. Decision Tree also took a little more time. For some problems, we don't need a complex modelIn [16 mean_squared_error(y_b, y_n) nse_navy_fat mse_navy_fat = round(mse_navy_fat, 2) print(f'The mean squared error for the Navy bod-fat formula is {mse_navy_fat}’) The mean squared error for the Navy bod-fat formula is 27.4 Wait, was machine learning necessary for predicting the body-fat? we saw just before that the navy body-fat formula was giving us a gcod estimation of the body-fat level. The mean squared error for the formule is even smaller than the mean squared error of the cecision tree