코딩팡 2호선 전처리 종합 28번 중고 카메라 매입가 문제 최적화 질문드립니다.

안녕하세요.

코딩팡 2호선 노선도 문제에서 28번 중고 카메라 매입가 문제를 풀었는데, 어떻게 해도 RMSE 값이 낮아지지 않아서 질문드립니다. 현재 아래와 같은 코드로 했을 때 가장 정확도가 높았고, 'shutter_count'를 중앙값이나 평균 등으로 채울 경우 정확도가 더 떨어졌었습니다.

해당 코드에서 어떤 부분을 수정해야 정확도가 높아질 수 있는지 모르겠어서 질문드립니다.

감사합니다.

import pandas as pd
pd.set_option('display.max_column', None)

train = pd.read_csv("data/camera_train.csv")
test = pd.read_csv("data/camera_test.csv")

# EDA
# print(train.shape, test.shape) # (2500, 11) (800, 10)
# print(train.info(), test.info()) # non-null, 대부분 obj, item_id drop, shutter_count '?' 유의
# print(train.describe(include='object'), test.describe(include='object')) # different
# print(train.isnull().sum())

# Pre - price
y_train = train.pop('price')
train = train.drop(['item_id'], axis=1)
test = test.drop(['item_id'], axis=1)
# encoding
import numpy as np
train['unknown'] = np.where(train['shutter_count']=='?', 1, 0)
test['unknown'] = np.where(test['shutter_count']=='?', 1, 0)
train['shutter_count'] = train['shutter_count'].replace({'?':'-1'}).astype(int)
test['shutter_count'] = test['shutter_count'].replace({'?':'-1'}).astype(int)

train['megapixel'] = train['megapixel'].str.replace('MP', '', regex=False).astype(float)
test['megapixel'] = test['megapixel'].str.replace('MP', '', regex=False).astype(float)
from sklearn.preprocessing import LabelEncoder,RobustScaler
le = LabelEncoder()
cols = train.select_dtypes(include='object').columns
tmp = pd.concat([train, test], axis=0)
for col in cols:
    tmp[col] = le.fit_transform(tmp[col])
train = tmp.iloc[:len(train)].copy()
test = tmp.iloc[len(train):].copy()
sc = RobustScaler()
train['shutter_count'] = sc.fit_transform(train[['shutter_count']])
test['shutter_count'] = sc.transform(test[['shutter_count']])
# print(train.head(10))
# Valid
from sklearn.model_selection import train_test_split
X_tr, X_val, y_tr, y_val = train_test_split(train, y_train, test_size=0.2, random_state=0)

# Modeling - RMSE
from sklearn.metrics import mean_squared_error
def val(name, pred):
    print(f"[{name}] RMSE = {(mean_squared_error(y_val, pred)) ** 0.5}")
# RF
from sklearn.ensemble import RandomForestRegressor
rf = RandomForestRegressor(random_state=0, n_estimators=600)
rf.fit(X_tr, y_tr)
pred_rf = rf.predict(X_val)
val('RF n_estimators=600', pred_rf)
# XGB
from xgboost import XGBRegressor
xgb = XGBRegressor(random_state=0, max_depth=5)
xgb.fit(X_tr, y_tr)
pred_xgb = xgb.predict(X_val)
val('XGB max_depth=5', pred_xgb)
# LGBM
from lightgbm import LGBMRegressor
lgb = LGBMRegressor(random_state=0, n_estimators=400, learning_rate=0.03, verbose=-1, max_depth=7)
lgb.fit(X_tr, y_tr)
pred_lgb = lgb.predict(X_val)
val('LGB max_depth=7 n_estimators=400', pred_lgb)

# [RF def] RMSE = 129.19054561590218
# [XGB def] RMSE = 85.62395927743444
# [LGB def] RMSE = 104.9641856432373

# [RF n_estimators=600] RMSE = 125.20707171872648
# [XGB max_depth=5] RMSE = 79.00022250759804
# [LGB max_depth=7 n_estimators=400] RMSE = 102.49297269651322

# Submit
xgb.fit(train, y_train)
pred = xgb.predict(test)
pd.DataFrame({'pred':pred}).to_csv('result.csv', index=False)
res = pd.read_csv('result.csv')
print(res.shape, res.head())