LinearRegression.py

import pandas as pd
teams = pd.read_csv("teams.csv")
#print(teams)
#Data Exploration
teams = teams[["team", "country", "year", "athletes", "age", "prev_medals", "medals"]] 
print(teams)

teams_numeric = teams[["athletes", "age", "prev_medals", "medals"]] #only ints because .corr won't work with string values like the year, team, and country

print(teams_numeric.corr()["medals"])

import seaborn as sns #run in interactive session using jupyter notebook
#sns.lmplot(x = "athletes", y = "medals", data = teams, fit_reg = True, ci = None) #fits regression line over data, ci = None gets rid of confidence interval.
#sns.lmplot(x="age", y="medals", data = teams, fit_reg = True, ci =None)
#teams.plot.hist(y= "medals") #look at column you are trying to predict to see how balanced it is

#Data Cleaning
print(teams[teams.isnull().any(axis = 1)]) #finds any rows w missing values

teams = teams.dropna() #removes all empty rows from data
print(teams)

train = teams[teams["year"]<2012].copy()
test = teams[teams["year"]>= 2012].copy()
print(train.shape) 
print(test.shape) #abt 80/20 split in training vs test data so its valid

#Training the Model
#using mean absolute error to evaluate metric
import sklearn
from sklearn.linear_model import LinearRegression
reg = LinearRegression() #training a LR model
predictors = ["athletes", "prev_medals"] #using these columns
target = "medals" #to predict this
print(reg.fit(train[predictors], train["medals"]))

predictions = reg.predict(test[predictors])
test["predictions"] = predictions

#print(predictions) #returns very long numpy array (not rounded)
#countries can only predict a whole number of medals and some numbers are negative which is incorrect

#rescale data
test.loc[test["predictions"] < 0, "predictions"] = 0 #if predictions < 0, turned into a zero
test["predictions"] = test["predictions"].round() #rounds values to nearest whole number

from sklearn.metrics import mean_absolute_error
error = mean_absolute_error(test["medals"], test["predictions"])

print(error) #3.299

print(teams.describe()["medals"]) #error is far below SD


print(test[test["team"]=="USA"]) #can view predictions by country
print(test[test["team"]=="IND"]) #enters less athletes 

errors = (test["medals"] - test["predictions"]).abs()
print(errors) #difference between predicted number of medals and actual number of medals

#group error rate by team
errors_by_team = errors.groupby(test["team"]).mean()
print(errors_by_team) #prints how many medals off we were for each country


medals_by_team = test["medals"].groupby(test["team"]).mean() #finds how many medals each team won on average
print(medals_by_team)

#Clean results 
error_ratio = errors_by_team / medals_by_team 
print(error_ratio) #has many NaN because a lot of team's have an average number won = 0

error_ratio = error_ratio[~pd.isnull(error_ratio)] #there is also some countries with infinity
print(error_ratio)

import numpy as np
error_ratio = error_ratio[np.isfinite(error_ratio)] #deals with inf
print(error_ratio)

error_ratio.plot.hist()

error_ratio = error_ratio.sort_values()
print(error_ratio)

#Add in more predictors
#try different ML models (neural network/computer vision)
#build a specific athlete level model