all. Currently I am working in a regressor to predict the median house value using 1990 census as a dataset (the same on Aurélien Géron used on his book).
My best result so far has been a score of ~0.96 for the training set and ~0.80 and for the test set with a bagging regressor. Here's how I had been doing:
precisao_treino = clf.score(dados_treino_feat,dados_treino_targ) #training score
precisao_valid = clf.score(dados_valid_feat,dados_valid_targ) #test score
(0.9698880465469455, 0.8087685894971024)
But here is where I get confused. That very high training score indicate that there is a potential overfit on my model. However, the test score is maintained a relative good score. So I don't know if my model is overfitting or if doing well at predicting. Could someone explain to me whether this a bad or good model?
Here's how my final dataset looks like, with one-hot-encoding:
['longitude',
'latitude',
'housing_median_age',
'total_rooms',
'total_bedrooms',
'population',
'households',
'median_income',
'median_house_value',
'<1H OCEAN',
'INLAND',
'ISLAND',
'NEAR BAY',
'NEAR OCEAN']
My solution to removing outliers:
df = df[(np.abs(zscore(df)) < 6).all(axis=1)]
My model:
clf = ensemble.BaggingRegressor(n_estimators=20)