Smart Factory-LGaimers Hackathon

Defection Multiclass Classification, Prediction, Industrial AI Modeling

• LG Aimers Hackathon - Smart Factory Classification

• code, csv files, related materials

• 02/17/2023(FRI):

  • preprocessing + GradientBoostingRegressor + GradientBoostingClassifier (X)
  • NO preprocessing + LGBM

• 02/18/2023(SAT):

  • preprocessing: fillna with feature's column mean
  • append data for solving data imbalance
  • further Ensemble modeling is needed
  • Smart Factory Controling System: company's focus

• 02/20/2023(MON):

  • preprocessing: fillna(0) + SMOTE oversampling
  • train_test set split
  • f1_score macro avg
  • Y_Quality ---> Y_Class prediction : f1_score not good
    • directly predicts Y_Class
  • DL: TabNet techniques
    • TabNetCalssifier modeling gets in stucked [ON GOING]
  • Ensemble Modeling
    • 2 level stacking

• 02/21/2023(TUE):

  • params tuning: OPTUNA
  • XGBClassifier (predict directly Y_Class)

• 02/24/2023(FRI): SCORE 0.72765!!!!!

  • DL: Sequential model + catBoostClassifier + optuna params optimization
  • 03/07 code & PPT submission

• 02/25/2023(SAT): best output is not recalled

  • sequential model:
    • input layer, output layer(#neuron=#label)
    • num of hidden layers = 0 --> only capable of representing linear separable functions or decisions
      • when features are linearly correlated, it can be done by using ML.
      • but even in linear correlation, if neural networks are used, there is no need for any hidden layer

• 02/26/2023(SUN): Tensorflow.Keras NN MLP (Sequential model) is not reproductable

  • sequential model + optuna
    • Tensorflow.keras: random --> familiar BUT not recalled
    • Torch: unfamiliar AND output is expected to be same
  • catBoost + GBC + voting classifier + optuna

• 02/27/2023(MON): focusing on ML ensemble

  • catBoost + randomizedSearch + stacking ensemble
  • tried sequential model finally (CPU setting + seed fixing) but not recallable

• 02/28/2023(TUE): FINISHED

  • VotingClassifier
  • based on each basic model's performance

• RESULT

  • overfitting
  • Cross-Validation is VERY IMPORTANT

• 03/01/2023(WED): feature distribution (DATA ANALYSIS), MLP in torch

  • feature distribution: imbalanced --> MinMaxScaler
  • MLP: code in torch
    • BUT still results are random
  • FEEDBACK:
    • tensorBoard --> automatical logging tool for parameters experiments etc.
    • decrease # neuron slowly
    • deep layers to make training hard --> prevent OVERFITTING

• 03/04~03/06: PPT

• 03/07: CODE & PPT Final Check

• 03/22/2023(WED): DL modeling

  • LSTM, Sequential Model

• 03/23/2023(THU): cross-validation + feature Importance

  • cross-validation score 0.550.6 is good (if cross validation > 0.70.8, then it seems like overfitting)
  • use ONLY import features(SHAP) with catBoostClassifier