prop_unique = (X[obj_cols].nunique() / len(df)).sort_values(
ascending=False) # in order of most unique to least
unique = pd.concat([prop_unique, nunique], axis=1)
unique.columns = [
"proportion",
"nunique",
]
print(unique)
ENCODE = True
if ENCODE:
X = similarity_encode(
X,
encode_columns=[
"Subtitle",
],
n_prototypes=4,
preran=False,
drop_original=True,
)
LENGTH_ENCODE = True
if LENGTH_ENCODE:
len_encode = ["URL"]
for col in len_encode:
X[f"{col}_len"] = X[col].apply(len)
X = X.drop(col, axis=1)
CATEGORIZE = False
if CATEGORIZE:
X[obj_cols] = X[obj_cols].astype("category")
import lightgbm as lgb
from helpers import similarity_encode
import pandas as pd
from pathlib import Path
import matplotlib.pyplot as plt
df = pd.read_csv(
r"data\forestfires.csv",
parse_dates=[],
index_col=[],
)
X, y = similarity_encode(df, encode=False, categorize=True, preran=False)
X = X.drop("rain", axis=1)
d = lgb.Dataset(X, y, silent=True)
# rmse: 98.18188205858038
NUM_BOOST_ROUND = 455
params = {
"objective": "rmse",
"metric": "rmse",
"verbose": -1,
"n_jobs": 6,
"learning_rate": 0.004090619790710353,
"feature_pre_filter": False,
"lambda_l1": 6.99239231800302e-08,
"lambda_l2": 9.330959145992983,
"num_leaves": 9,
"feature_fraction": 0.8999999999999999,
"bagging_fraction": 1.0,
"bagging_freq": 0,
"min_child_samples": 20,
obj_cols = X.select_dtypes("object").columns
nunique = X[obj_cols].nunique()
prop_unique = (X[obj_cols].nunique() / len(df)).sort_values(
ascending=False) # in order of most unique to least
unique = pd.concat([prop_unique, nunique], axis=1)
unique.columns = [
"proportion",
"nunique",
]
unique
if ENCODE:
X = similarity_encode(
X,
encode_columns=[],
n_prototypes=5,
train=True,
drop_original=False,
)
X[obj_cols] = X[obj_cols].astype("category")
sns.kdeplot(y)
plt.title("KDE distribution")
plt.show()
SEED = 0
SAMPLE_SIZE = 10000
Xt, Xv, yt, yv = train_test_split(
X, y, random_state=SEED) # split into train and validation set
obj_cols = X.select_dtypes("object").columns
nunique = X[obj_cols].nunique()
prop_unique = (X[obj_cols].nunique() / len(df)).sort_values(
ascending=False
) # in order of most unique to least
unique = pd.concat([prop_unique, nunique], axis=1)
unique.columns = [
"proportion",
"nunique",
]
unique
X = similarity_encode(
X,
encode_columns=["Name", "MeetName", "Division", "Federation"],
n_prototypes=5,
train=True,
drop_original=False,
)
X[obj_cols] = X[obj_cols].astype('category')
sns.kdeplot(y)
plt.title("KDE distribution")
plt.show()
SEED = 0
SAMPLE_SIZE = 10000
Xt, Xv, yt, yv = train_test_split(
X, y, random_state=SEED