Awarded 16 analytical reports with Python at 16GB Kernel of Kaggle.

About the Author

• Name: Satoru Shibata / 柴田 怜
• Job: Sr. Data Scientist
• Titles:
  • 3x Kaggle Expert
    • Retired from Kaggle at 2021 to focus on business as a data scientist.
  • 4x Certified Professional
    • IBM Data Science Professional Certificate at 2022.
    • IBM Applied Data Science Specialization Certificate at 2023.
    • SAS Statistical Business Analyst Professional Certificate at 2024.
    • IBM Generative AI for Data Scientists Specialization Certificate at 2024.

Score Table

Departments	Top Levels	Highest Rank	Awarded Medals
Code	0.2%	317/161,898	3 Silver 13 Bronze
Discussion	0.3%	588/188,433	100 Bronze
Datasets	1%	354/34,643	3 Bronze
Competitions	20-30%

Abstracts

3 Silver Medals

1. Optimized LightGBM with Optuna adding SAKT Model
   • Lead sentences
     • Submitted code using two Ensemble Learning Methods.
     • Used 100 million rows of training data for prediction on a 16GB Kernel removing unnecessary objects.
   • Issue
     • Algorithms for TOEIC Learning Applications
   • Significance
     • Predict percentage of correct answers based on user's behavioral history.
     • User's percentage of correct answers will increase with the number of problems solved.
   • Purpose
     • Optimize Binary Classification for AUC.
   • Methodology:
     • Ensemble Learning of LightGBM and SAKT.
     • Hyperparameter Optimization with Optuna.
   • Results
     • Score: AUC = 0.781.
     • Code: 31 Points.
   • Considerations
     • Obsessed with Models, Feature Engineering Remains a Challenge.
     • Systematizing Multiple Models will also be a challenge in the future.
   • Conclusion
     • Code Silver Medal
2. LightGBM Classifier and Logistic Regression Report
   • Lead sentences
     • Optimized Classification of Anonymized Raw Data from Stock Market on 16GB Kernel.
     • Contributed code that systematizes Ensemble Learning and Logistic Regression.
   • Issue
     • Utility Function Optimization of Supply and Demand Forecasting in Securities Markets.
   • Significance
     • Calculate based on indicators of absence or degree of stock returns.
     • Optimize the behavior of whether to trade or not.
   • Purpose
     • AI Dev for Profit Maximization.
   • Methodology
     • Optimal classification of LightGBM.
     • Logit Transformation of Purpose Variables Based on Probability Distributions.
   • Results
     • Score: 3741.118 (Outside of Medal Zone).
     • Code: 33 Points.
   • Considerations
     • Utility function was not fully deciphered
     • Which left some issues for paper surveys.
     • The report was appreciated by other Kagglers.
   • Conclusion
     • Code Silver Medal
3. Optimize LightGBM HyperParameter with Optuna and GPU
   • Lead sentences
     • Unprecedented LightGBM Hyperparameter Optimization on GPU.
     • Procedure was annexed and highly evaluated.
   • Issues
     • Preliminaries of “LightGBM Classifier and Logistic Regression Report“ .
   • Significance
     • High Parameter Optimization.
     • There were few precedents for LightGBM.
   • Purpose
     • Code submission for optimizing LightGBM Hyperparameter on GPU.
   • Methodology
     • A survey of prior case studies using Optuna for LightGBM.
     • Procedure of ssubmissions.
   • Results
     • Run: 953.9s on GPU
     • Code: 31 Points.
   • Consideration
     • Available hyperparameter optimization of futures.
   • Conclusion
     • Code Silver Medal.

13 Bronze Medals

1. Optimized Logit LightGBM Classifier and CNN Models
   • Lead sentences
     • Submitted a simulation of Multiple Model Systematization.
     • Based on this failure, I was able to concentrate on LightGBM Optimization and Inference.
   • Issue
     • Exploring Optimization Models
   • Significance
     • Simulation iterations of Optimization Model.
   • Purpose
     • Optimize Utility Function by systematizing Multiple Models.
   • Methodology
     • Applying the Logit Transform to LightGBM.
     • Explore combining with CNN.
   • Results
     • Score: 3344.738 (Outside of Medal Zone).
     • Code: 15 Points.
   • Considerations
     • This code does LightGBM and CNN at the same time, which was prone to overflow.
     • From now on, I will focus on one Model Optimization.
   • Conclusion
     • Code Bronze Medal
2. Optimized LightGBM with Optuna
   • Lead sentences
     • Dev Baseline Model for Code Competition to process 100 million rows of training data.
     • The minimum performance was predicted to be 16GB.
   • Issue
     • 100 million rows of training data must be predicted on a 16GB Kernel.
   • Significance
     • This is the cornerstone of the final submission model.
     • Preprocess and Feature Engineering were adjusted for further optimization.
   • Purpose
     • Baseline Model Dev
   • Methodology
     • Binary Classification by LightGBM Optimization.
   • Results
     • Score: AUC = 0.774.
     • Code: 12 Points.
   • Considerations
     • Policy of additional development to Baseline Model.
     • The improvement of AUC by the additional development was only 0.07, which left some issues.
   • Conclusion
     • Code Bronze Medal
3. LightGBM on GPU with Feature Engineering, Optuna, and Visualization
   • Lead sentence
     • Code Bronze Medal for first attempt at submitting code.
   • Issue
     • This was my first real effort at Kaggle.
   • Significance
     • Visualize in a timely manner, and features were studied.
     • Optuna was also used for the first time and applied later.
   • Purpose
     • Work on Feature Engineering.
   • Methodology
     • I read and referred to posted code by Kaggle Grandmaster.
   • Results
     • Code: 11 Points.
   • Consideration
     • I could gain experiences in implementing LightGBM with Optuna on GPU.
   • Conclusion
     • Code Bronze Medal.
4. LightGBM with the Inference and Empirical Analysis
   • Lead sentences
     • In the first scored submission code, AUC = 0.76.
     • The challenges were used as the cornerstone of development experiences.
   • Issue
     • Scoring by developing additions to the submitted code for my first challenge.
   • Significance
     • A single process was limited to Model Object Generation.
   • Purpose
     • To further improve the performance of Prediction Model.
   • Methodology
     • Inference was added to improve Score.
     • Empirical Analysis between raw data and predicted results.
     • Detected significant differences in Gaussian Distribution.
   • Results
     • Score: AUC = 0.76.
     • Code: 12 Points.
   • Consideration
     • This submitted code left insufficient understanding of inference as an issue.
   • Conclusion
     • Code Bronze Medal.
5. Submission and the Inference of LightGBM
   • Lead sentences
     • My first scoring submission code prototype
     • Few examples of Empirical Analysis, I won Code Bronze Medal.
   • Issue
     • Prototype version of submission code for first scoring.
   • Significance
     • Implementing the scoring submission code.
   • Purpose
     • Gaining development experiences.
   • Methodology
     • Model objects were coded for scoring.
     • Empirical Analysis detected a significant difference in Gaussian Distribution.
   • Result
     • Code: 7 Points.
   • Considerations
     • Actual scoring submission code became a separate file.
     • This was an opportunity for me to feel the challenge of coding.
     • Focused on its afterwards.
   • Conclusion
     • Code Bronze Medal.
6. Market Prediction XGBoost with GPU Modified
   • Lead sentences
     • Performance comparison with LightGBM by XGBoost Optimization.
     • LightGBM takes the cake.
   • Issue
     • I seen good results with XGBoost sometimes.
   • Significance
     • Simulate on Models other than LightGBM and search for Optimized Model.
   • Purpose
     • Score improvement by XGBoost.
   • Methodology
     • GPU Implementation into XGBoost Optimization.
   • Results
     • Score: 3308.824 (Outside of Medal Zone).
     • Code: 8 Points.
   • Considerations
     • XGBoost is easy to implement due to its many precedents.
     • LightGBM is superior in performance comparison, which led me to focus on LightGBM.
   • Conclusion
     • Code Bronze Medal.
7. Cassava Leaf Disease Best Keras CNN Tuning
   • Lead sentences
     • I also participated in a competition on image analysis, challenging myself with raw data of various properties.
     • I was left with some issues on the theoretical side, which gave me an opportunity to work from theoretical books.
   • Issue
     • I would like to try my hand at image analysis and find out what I am good at.
   • Significance
     • I want to gain experience in Keras implementation.
     • Deepen my understanding CNN.
   • Purpose
     • I learn to understand and implement acoustic analysis and image analysis.
   • Methodology
     • I complemented the advanced submission code.
   • Results
     • Score: Accuracy = 0.885.
     • Code: 18 Points.
   • Considerations
     • Theoretical aspects of acoustic analysis and image analysis remained a challenge.
     • An opportunity to raise awareness to need to start with a survey of theoretical papers.
   • Conclusion
     • Code Bronze Medal
8. RFCX Residual Network with TPU Customized
   • Lead sentences
     • I also participated in a competition for acoustic analysis, and tried my hand at raw data of various properties.
     • I was left with some issues on the theoretical side, which gave me an opportunity to work from theoretical books.
   • Issue
     • I would like to try my hand at acoustic analysis and find out what I am good at.
   • Significance
     • I want to gain experience in Keras implementation.
     • Deepen my understanding CNN.
   • Purpose
     • I learn to understand and implement acoustic analysis and image analysis.
   • Methodology
     • I complemented the advanced submission code.
   • Results
     • Score: 0.772.
     • Code: 12 Points.
   • Considerations
     • Theoretical aspects of acoustic analysis and image analysis remained a challenge.
     • An opportunity to raise awareness to need to start with a survey of theoretical papers.
   • Conclusion
     • Code Bronze Medal.
9. Research with Customized Sharp Weighted
   • Lead sentences
     • Work on Custom Metrics Clarification and systematization of Hyperparameters Optimization in LightGBM.
     • An each milestone optimization object generation is still important.
   • Issue
     • Private Custom Metrics were used as an Evaluation Function.
   • Significance
     • Improve prediction accuracy by elucidating Private Custom Metrics.
     • Reproducibility will be determined based on the Evaluation Function.
   • Purpose
     • Custom Metrics Clarification.
   • Methodology
     • LightGMB High Parameter Optimization.
     • Systematization with Custom Metrics Decoding Examples.
   • Results
     • Generate each Parameter Optimization Object.
     • Code: 6 Points.
   • Consideration
     • Importance of each milestone optimization object generation was reaffirmed.
   • Conclusion
     • Code Bronze Medal.
10. Optimize CatBoost HyperParameter with Optuna and GPU
    • Lead sentences
      • Performance comparison was performed on optimized Ensemble Learning.
      • LightGBM won the prediction accuracy.
    • Issue
      • I was new to CatBoost and wanted to compare performance with LightGBM.
    • Significance
      • Performance comparison of Ensemble Learning: LightGBM, XGBoost, CatBoost, etc.
    • Purpose
      • Algorithm selection for Prediction Models.
    • Methodology
      • Hyper-parameter optimization.
      • CatBoost implementation.
    • Results
      • Score: AUC = 0.500.
      • Code: 17 Points.
    • Consideration
      • At the baseline model stage, I gave the edge to LightGBM.
    • Conclusion
      • Code Bronze Medal.
11. LightGBM on Lyft Tabular Data added Inference and Tuning
    • Lead sentences
      • Regression Prediction of LightGBM with Grid Search and Multiple Evaluation Functions.
      • A harvest that uncovered all sorts of challenges!.
    • Issue
      • Regression Problem for Table Data Related to Automated Driving.
    • Significance
      • I want to work on Regression Prediction with LightGBM.
      • Gain further development experiences.
      • Implement multiple evaluation functions to improve accuracy.
    • Purpose
      • Improving accuracy of Regression Prediction.
    • Methodology
      • Set evaluation functions of LightGBM in MSE and RMSE.
      • Parameter search by grid search.
    • Results
      • Score: 356.084.
      • Code: 10 Points.
    • Considerations
      • Grid search shown that hyperparameter optimization is inefficient.
      • I reaffirmed the need to use feature engineering and inference.
    • Conclusion
      • Code Bronze Medal
12. COVID-19 with H2OAutoML Baseline Model
    • Lead sentences
      • Experimented with AutoML performance, but found the original to be more powerful.
      • This led to the original development of the LightGBM optimization.
    • Issue
      • COVID-19 infection explosion and new global challenges.
    • Significance
      • Improvement of coding techniques for anonymized Table data.
      • Accumulate experiences using AutoML.
    • Purpose
      • Optimization Regression Prediction with AutoML.
    • Methodology
      • Set RMSLE as evaluation function for Regression Prediction with H2O.
      • Extract the optimized Regression Prediction Models: Deep Learning, XGBoost, GLM, GBM, etc.
    • Results
      • Score: RMSLE = 0.086.
      • Code: 6 Points.
    • Considerations
      • Original development was more powerful than H2OAutoML.
      • Opportunity to work on Optimized Regression Prediction with LightGBM.
    • Conclusion
      • Code Bronze Medal.
13. Optimized Predictive Model with H2OAutoML
    • Lead sentences
      • Even in Binary Classification, AutoML was found to be inferior to proprietary.
      • It is thought that the difference was due to Preprocess and Feature Engineering.
    • Issue
      • Regression Prediction by H2OAutoML was inferior to original development.
    • Significance
      • It was unclear whether results would be similar to Regression Prediction.
    • Purpose
      • Experiment on H2OAutoML in Binary Classification.
    • Methodology
      • Set RMSLE as the evaluation function for Binary Classification with H2O.
      • Extract the Optimized Binary Classification Models: Deep Learning, XGBoost, GLM, GBM. ey/tc.
    • Results
      • Score: AUC = 0.850.
      • Code: 5 Points.
    • Considerations
      • The performance was higher than that of Regression Prediction Case.
      • Process and Feature Engineering itself is not automated.
      • It has to be developed independently.
    • Conclusion
      • Code Bronze Medal.