[ML] YOUTUBE revenue prediction

youtube daily estimated revenue prediction with machine-learning (tree-based models)

• 프로젝트 레포트(Kaggle Notebook)
• 데이터셋

Introduction

• Interest & Inspirations
• Is it possible to predict a daily revenue of a certain youtuber with his/her channel's daily view, subscribers gained, average viewed duration, etc...?
• Which factor is most related with increasing daily Revenue?
• How much can Total Views of videos devote to actual Monetizing?
• If one has more videos posted on his/her channel, would she/he happen to earn more?

Environment & Skills

• Google Colab | Jupyter Notebook
• Python 3.9
  • Library for EDA & Data Preprocessing : Pandas, Numpy
  • Library for Statistics & Data Analysis : Scipy.stats, sklearn
  • Library for Data Visualization : Matplotlib, Seaborn
  • Library for ML Models : sklearn, xgboost, lightgbm
  • Library for Model Interpretation & Insight : eli5, pdpbox, shap
• Jamobi
• STATISTIC TEST -Anova (one-way) Test, Variance Inflation Factor Check, F-Statistics(p-value)
• ML MODELS
  • Linear-Models : LinearRegression, Ridge, Lasso, ElasticNet
  • Tree-based-Models : RandomForest Regressor, Xgboost Regressor, Lightgbm Regressor
  • TransformedTargetRegressor : Log Transformed Linear Regression
• Miri Canvas
• OBS

DATA

Data from : YouTube Studio (Youtuber - PoohinKorea) Data Size : (1132, 16) / (rows, columns)

Date	Column Names	Describtion
1	Average_views_per_viewer	시청자 당 평균 시청 횟수
2	Unique_viewers	순 시청자 수
3	Click_rate	노출 시 클릭율
4	Impressions	노출 횟수
5	Comments	댓글 수
6	Shares	공유 수
7	Dislikes	싫어요 수
8	Likes	좋아요 수
9	Average_viewed	평균 시청율
10	Videos_published	새 영상 수
11	Subscribers	구독자 순 증가수
12	Views	조회수
13	Watch_hours	하루 누적 총 시청시간
14	Average_view_sec	영상 평균 시청시간(초)
15	Revenue	일일 예상 수익

Data Analasys

Correlations with Target

F-statistic & P-Values

Features	F-statistics	p-Values
Unique_viewers	6222.843772	0.000000
Impressions	1876.512061	0.000000
Likes	3073.756391	0.000000
Views	4103.723499	0.000000
Watch_hours	4466.619683	0.000000
Subs_accumulated	662.696705	0.000000

Evaluations

MODELS	MAE	RMSE	R2
Log Transforemed Linear	$717.946	$1319.06	-2815.379
Log Transformed Ridge	$132.31	$177.49	-49.99
Log Transformed Lasso	$99.75	$135.58	-28.75
Log Transformed ElasticNet	$91.33	$121.91	-23.06
Log Transformed RF Regressor	$17.13	$20.68	0.307
Log Transformed XGB Regressor	$10.28	$13.51	0.704
Log Transformed LGBM Regressor	$16.93	$21.10	0.279
Combined Model(XGB:0.8, RF:0.1, LGBM:0.1)	$10.11	$13.15	0.72

Hyperparameters Tuning

• Log Transformed RF Regressor

DATA	MAE	RMSE	R2
TRAIN SET	$2.227	$4.509	0.967
TEST SET	$19.38	$22.19	0.20

• Log Transformed XGB Regressor

DATA	MAE	RMSE	R2
TRAIN SET	$0.39	$0.66	0.999
TEST SET	$10.01	$13.33	0.71

• Log Transformed LGBM Regressor

DATA	MAE	RMSE	R2
TRAIN SET	$1.03	$5.01	0.959
TEST SET	$17.01	$21.05	0.28

• Tuned Combined Models (XGB:0.8, RF:0.1, LGBM:0.1)

DATA	MAE	RMSE	R2
TEST SET	$9.59	$12.94	0.728

Top Coef & Feature Importance

Top Permutation Importance

Results

• SHAP values