22년 11월 진행된 Dacon 제주도 도로 교통량 예측 대회의 코드를 공유합니다.

(대회 정보 : https://dacon.io/competitions/official/235985/overview/description)

Import

import pandas as pd
import numpy as np

from workalendar.asia import SouthKorea
from datetime import datetime

from sklearn.model_selection import train_test_split, cross_val_score
from sklearn.preprocessing import OrdinalEncoder
from sklearn.model_selection import GridSearchCV

from xgboost import XGBRegressor

import matplotlib.pyplot as plt
import seaborn as sns
from tqdm import tqdm

import gc

import warnings
warnings.filterwarnings(action='ignore')

Data Load

def csv_to_parquet(csv_path, save_name):
    df = pd.read_csv(csv_path)
    df.to_parquet(f'../origin_data/{save_name}.parquet')
    del df
    gc.collect()
    print(save_name, 'Done.')

csv_to_parquet('../origin_data/train.csv', 'train')
csv_to_parquet('../origin_data/test.csv', 'test')

train Done.
test Done.

train = pd.read_parquet('../origin_data/train.parquet')
test = pd.read_parquet('../origin_data/test.parquet')

EDA

for i in train['maximum_speed_limit'].unique():
    idx = np.where(train['maximum_speed_limit']==i)
    sns.kdeplot(train.loc[idx, 'target'], label=i)
plt.legend()
plt.show()
  • Limit Speed 별로 target이 다른 분포를 보임.
gps = train[['start_longitude', 'end_longitude', 'start_latitude', 'end_latitude', 'target']]

gps_set = [gps['start_longitude'].min(), gps['start_longitude'].max(), gps['start_latitude'].min(), gps['start_latitude'].max()] # 지도 그림의 gps 좌표
gps_set

[126.182616549771, 126.930940973848, 33.2434317486804, 33.5560801767072]

vel_low_idx = gps.loc[gps['target']<15].index # 시내 교통 체증기준 10 km/h 미만
vel_high_idx = gps.loc[gps['target']>80].index # 고속도로 원활기준 80 km/h 초과
  • GPS 좌표를 기준으로 시각화

(참고자료 : https://towardsdatascience.com/simple-gps-data-visualization-using-python-and-open-street-maps-50f992e9b676)

f, ax = plt.subplots(figsize=(21,10))

ax.set_xlim(gps_set[0], gps_set[1])
ax.set_ylim(gps_set[2], gps_set[3])

image = plt.imread('map.png')
ax.imshow(image, zorder=0, extent=gps_set, aspect='equal')

for i in tqdm(vel_low_idx): # 교통 체증 도로 빨강
    x_1 = gps.loc[i,'start_longitude']
    x_2 = gps.loc[i,'end_longitude'] 
    y_1 = gps.loc[i,'start_latitude']
    y_2 = gps.loc[i,'end_latitude'] 
    ax.plot([x_1, x_2], [y_1, y_2], color='red')

for i in tqdm(vel_high_idx): # 교통 원활 도로 파랑
    x_1 = gps.loc[i,'start_longitude']
    x_2 = gps.loc[i,'end_longitude'] 
    y_1 = gps.loc[i,'start_latitude']
    y_2 = gps.loc[i,'end_latitude'] 
    ax.plot([x_1, x_2], [y_1, y_2], color='blue')

plt.show()

100%|██████████| 111725/111725 [00:45<00:00, 2475.80it/s]
100%|██████████| 26239/26239 [00:08<00:00, 2942.93it/s]
  • 가장 Importance가 큰 Max Speed Limit를 기준으로 Data Set을 분할
for i in range(30,90,10):
    if i==40: # Test Data에 max speed limit 40은 없으므로 생략.
        continue
    else:
        globals()[f'train_{i}'] = train.loc[train['maximum_speed_limit']==i]
  • Data set 별로 모델 Tuning. (Data set 별 Tuning 코드는 생략)
parameters = {'max_depth': np.arange(7,13,1),
                # 'n_estimators' : np.arange(10,101,10),
                # 'gamma' : np.arange(0,10,1),
              'random_state': [random_state]
              }

estimator = XGBRegressor()
xgb = GridSearchCV(estimator=estimator, 
                   param_grid=parameters, 
                   scoring='neg_mean_absolute_error', 
                   cv=5,
                   verbose=3)
xgb.fit(train_80.drop('target',axis=1), train_80['target'])
print(xgb.best_estimator_)

Fitting 5 folds for each of 6 candidates, totalling 30 fits
[CV 1/5] END .....max_depth=7, random_state=42;, score=-2.919 total time=  12.1s
[CV 2/5] END .....max_depth=7, random_state=42;, score=-2.939 total time=  12.2s
[CV 3/5] END .....max_depth=7, random_state=42;, score=-2.935 total time=  12.1s
[CV 4/5] END .....max_depth=7, random_state=42;, score=-2.931 total time=  12.0s
[CV 5/5] END .....max_depth=7, random_state=42;, score=-2.932 total time=  12.0s
[CV 1/5] END .....max_depth=8, random_state=42;, score=-2.865 total time=  14.1s
[CV 2/5] END .....max_depth=8, random_state=42;, score=-2.881 total time=  14.1s
[CV 3/5] END .....max_depth=8, random_state=42;, score=-2.871 total time=  14.1s
[CV 4/5] END .....max_depth=8, random_state=42;, score=-2.871 total time=  14.1s
[CV 5/5] END .....max_depth=8, random_state=42;, score=-2.875 total time=  14.0s
[CV 1/5] END .....max_depth=9, random_state=42;, score=-2.833 total time=  16.3s
[CV 2/5] END .....max_depth=9, random_state=42;, score=-2.846 total time=  16.4s
[CV 3/5] END .....max_depth=9, random_state=42;, score=-2.837 total time=  16.3s
[CV 4/5] END .....max_depth=9, random_state=42;, score=-2.832 total time=  16.2s
[CV 5/5] END .....max_depth=9, random_state=42;, score=-2.838 total time=  16.2s
[CV 1/5] END ....max_depth=10, random_state=42;, score=-2.819 total time=  18.8s
[CV 2/5] END ....max_depth=10, random_state=42;, score=-2.833 total time=  18.6s
[CV 3/5] END ....max_depth=10, random_state=42;, score=-2.819 total time=  18.5s
[CV 4/5] END ....max_depth=10, random_state=42;, score=-2.822 total time=  18.6s
[CV 5/5] END ....max_depth=10, random_state=42;, score=-2.817 total time=  18.6s
[CV 1/5] END ....max_depth=11, random_state=42;, score=-2.827 total time=  21.2s
[CV 2/5] END ....max_depth=11, random_state=42;, score=-2.844 total time=  21.2s
[CV 3/5] END ....max_depth=11, random_state=42;, score=-2.827 total time=  22.0s
[CV 4/5] END ....max_depth=11, random_state=42;, score=-2.830 total time=  21.2s
[CV 5/5] END ....max_depth=11, random_state=42;, score=-2.827 total time=  21.4s
[CV 1/5] END ....max_depth=12, random_state=42;, score=-2.858 total time=  23.9s
[CV 2/5] END ....max_depth=12, random_state=42;, score=-2.868 total time=  24.2s
[CV 3/5] END ....max_depth=12, random_state=42;, score=-2.857 total time=  24.3s
[CV 4/5] END ....max_depth=12, random_state=42;, score=-2.864 total time=  23.9s
[CV 5/5] END ....max_depth=12, random_state=42;, score=-2.857 total time=  23.9s
XGBRegressor(base_score=0.5, booster='gbtree', callbacks=None,
             colsample_bylevel=1, colsample_bynode=1, colsample_bytree=1,
             early_stopping_rounds=None, enable_categorical=False,
             eval_metric=None, gamma=0, gpu_id=-1, grow_policy='depthwise',
             importance_type=None, interaction_constraints='',
             learning_rate=0.300000012, max_bin=256, max_cat_to_onehot=4,
             max_delta_step=0, max_depth=10, max_leaves=0, min_child_weight=1,
             missing=nan, monotone_constraints='()', n_estimators=100, n_jobs=0,
             num_parallel_tree=1, predictor='auto', random_state=42,
             reg_alpha=0, reg_lambda=1, ...)

Prediction

for i in range(30,90,10):
    if i==40:
        continue
    else:
        globals()[f'X_train_{i}'] = X_train.loc[X_train['maximum_speed_limit']==i]

# 튜팅된 Parameter를 각 모델에 부여
for i in range(30,90,10):
    if i==40:
        continue
    elif i==30:
        n_train = eval(f'X_train_{i}')
        n_y_train = y_train[eval(f'X_train_{i}').index]
        globals()[f'xgb_{i}'] = XGBRegressor(random_state=random_state, max_depth=9).fit(n_train, n_y_train)
    elif i==50:
        n_train = eval(f'X_train_{i}')
        n_y_train = y_train[eval(f'X_train_{i}').index]
        globals()[f'xgb_{i}'] = XGBRegressor(random_state=random_state, max_depth=12).fit(n_train, n_y_train)
    elif i==60:
        n_train = eval(f'X_train_{i}')
        n_y_train = y_train[eval(f'X_train_{i}').index]
        globals()[f'xgb_{i}'] = XGBRegressor(random_state=random_state, max_depth=12).fit(n_train, n_y_train)
    elif i==70:
        n_train = eval(f'X_train_{i}')
        n_y_train = y_train[eval(f'X_train_{i}').index]
        globals()[f'xgb_{i}'] = XGBRegressor(random_state=random_state, max_depth=11).fit(n_train, n_y_train)
    elif i==80:
        n_train = eval(f'X_train_{i}')
        n_y_train = y_train[eval(f'X_train_{i}').index]
        globals()[f'xgb_{i}'] = XGBRegressor(random_state=random_state, max_depth=10).fit(n_train, n_y_train)

# Limit Speed 별 모델로 test data set의 target 예측
for i in range(30,90,10):
    if i==40:
        continue
    else:
        globals()[f'pred_xgb_{i}'] = eval(f'xgb_{i}').predict(test.loc[test['maximum_speed_limit']==i])

# 모델 예측값을 각각 대입
for i in range(30,90,10):
    if i==40:
        continue
    else:
        test.loc[test['maximum_speed_limit']==i, 'pred'] = eval(f'pred_xgb_{i}')

pred = test['pred']

sample_submission = pd.read_csv('../origin_data/sample_submission.csv')

pred.reset_index(inplace=True, drop=True)

Submission

# Submission 파일 생성
sample_submission['target'] = pred
today= datetime.now().date().strftime('%m%d')
sample_submission.to_csv(f"./99.submission/submit_{today}_1.csv", index = False)

sample_submission
id target
0 TEST_000000 26.753897
1 TEST_000001 42.798801
2 TEST_000002 67.241051
3 TEST_000003 37.343075
4 TEST_000004 43.202198
... ... ...
291236 TEST_291236 47.885696
291237 TEST_291237 51.416042
291238 TEST_291238 23.034536
291239 TEST_291239 23.455151
291240 TEST_291240 48.867462

291241 rows × 2 columns

댓글남기기

참고

ChatGPT를 위한 최적의 프롬프트 만들기

AI 언어 모델인 ChatGPT와 같은 모델이 인기를 얻으면서 AI와 효과적으로 상호 작용하는 것이 점점 중요해지고 있습니다. ChatGPT에서 가장 정확하고 유용한 결과를 얻으려면 신중하게 고안된 프롬프트를 사용해야 합니다. 이 게시물에서는 ChatGPT를 위한 이상적인 프롬프트...

Crafting the Ultimate Prompts for ChatGPT

As AI language models like ChatGPT continue gaining popularity, effectively interacting with them has become increasingly important. To get the most accurate...

머신러닝 vs 딥러닝

머신 러닝과 딥 러닝은 인공 지능(AI) 분야에서 두 가지 중요한 개념입니다. 종종 서로 바꿔 사용되지만, 두 개념의 원리와 차이점을 이해하는 것이 중요합니다. 이 글에서는 머신 러닝과 딥 러닝의 핵심 아이디어, 강점 및 약점을 탐구하여 둘의 차이점을 완전히 이해해보겠습니다.