[python][scikit-learn] KBinsDiscretizer 이용해서 Discretization 하기

fig1. title

---

scikit-learn.preprocessing 에서 KBinsDiscretizer 를 이용해서 Discretization 하기

continuous variable을 grouping 할 때 유용함

---

continuous 단일 column을 이용해서 grouping

bins=2일때 세가지 방법 쉽게 이해하기

#1. strategy = 'uniform'

#2. strategy = 'quantile'

#3. strategy = 'kmeans'

---

purpose : y_train_pd 라는 데이터프레임의 pred-true라는 column으로 grouping

 

#0. data

train_pt = pd.DataFrame(y_train_pd['pred-true'])

 

#1. strategy = 'uniform' : mean 기준

#discretizer strategy = uniform

est_uni = KBinsDiscretizer(n_bins=2, encode='ordinal', strategy='uniform')
est_uni.fit(train_pt)

Xt_uni=est_uni.transform(train_pt)
print("edges : ", est_uni.bin_edges_[0])

unique, counts = np.unique(Xt_uni, return_counts=True)
print(dict(zip(unique, counts)))

plt.hist(Xt_uni, bins='auto')
plt.xticks([0,0.5,1])
plt.rcParams['figure.figsize']=(5.5,6)
plt.show

fig2. output of #1

 

#2. strategy = 'quantile' : median 기준

#discretizer strategy = quantile

est_quan = KBinsDiscretizer(n_bins=2, encode='ordinal', strategy='quantile')
est_quan.fit(train_pt)

Xt_quan=est_quan.transform(train_pt)
print("edges : ", est_quan.bin_edges_[0])

unique_q, counts_q = np.unique(Xt_quan, return_counts=True)
print(dict(zip(unique_q, counts_q)))

plt.hist(Xt_quan, bins='auto')
plt.xticks([0,0.5,1])
plt.rcParams['figure.figsize']=(5.5,6)
plt.show

fig3. output of #2

 

#3. strategy = 'kmeans' : normal distribution이 두개 생성되게

#discretizer strategy = kmeans

est_km = KBinsDiscretizer(n_bins=2, encode='ordinal', strategy='kmeans')
est_km.fit(train_pt)

Xt_km=est_km.transform(train_pt)
print("edges : ", est_km.bin_edges_[0])

unique_k, counts_k = np.unique(Xt_km, return_counts=True)
print(dict(zip(unique_k, counts_k)))

plt.hist(Xt_km, bins='auto')
plt.xticks([0,0.5,1])
plt.rcParams['figure.figsize']=(5.5,6)
plt.show()

 

fig4. output of #3

Detail : https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.KBinsDiscretizer.html