Skip to content

README.md

Latest commit

 

History

History
47 lines (36 loc) · 3.28 KB

README.md

File metadata and controls

47 lines (36 loc) · 3.28 KB
Raw

Lending Club Default Risk

  • main goals: predict loan acceptance, default status, and interest rate; deploy model

  • skills learned:

    • handling tables > 5GB, 30 Mio. rows, 160 features.
    • correlation strength and feature importances
    • model selection for classification and linear regression
    • exploring recursive feature elimination
    • exercise model deployment (Docker, Google Cloud Platform)

  • most important modules used:

    • dask, fastparquet (memory saving processing and storing)
    • pandas, numpy (standard processing)
    • plotly, seaborn, matplotlib (visualization)
    • pingouin (statistical analysis)
    • scikit-learn, imblearn, lightgbm, xgboost, yellowbrick, pickle (modelling and performance)
    • eli5, shap (explainability)
    • fastapi, pydantic, uvicorn (deployment)

    Feedback from Peer Reviewer and from Senior Team Lead (STL, data science expert currently active in the industry)

  • peer: 96/100 points

    • Positives:
      • The analysis presents a comprehensive and informative Exploratory Data Analysis (EDA) with interactive Plotly visualizations, effectively conveying the story of the data.
      • The model evaluation process is detailed, covering various aspects, and a wide range of models and methods have been tested.
    • Areas for Improvement:
      1. For classification problems, consider splitting the training and test sets by date, similar to what was done for interest rate prediction. This can help account for time-related changes in the data.
      2. Improve preprocessing strategies by tailoring them to the specific model type.
        • A. For tree-based models, consider treating missing values as a separate category or using a distinct numeric value rather than median/most frequent imputation. Additionally, skip scaling numeric features, which is unnecessary for tree-based models and can affect results when dealing with extreme outliers.
        • B. For linear models, emphasize the importance of scaling and address outliers as a preprocessing step in pipelines involving linear models.
      3. Provide explanations for the choice of scoring metrics in hyperparameter tuning for classification tasks. This can help readers understand the decision-making process.
      4. Be cautious about filtering too many features solely based on direct relationships during EDA, as this may overlook conditional relationships that are valuable for tree-based models.
      5. Avoid setting a strict cut-off point for low variance features, as it may lead to the removal of rare but crucial features that are essential in specific cases.
      6. I could not evaluate the deployment of the model at the time.

  • STL (after incorporating several of peer's suggestions): 96/100 points

    • Great work! I really love that you took effort to understand the domain of the problem you were solving. Everything from dataset cleaning to model selection was completed with care.

    • What I would like you to try out next time is:

      • Do not be afraid of adding more text within your notebooks. Now it is not easy to understand origin of some plots of code cells without your vocal comments.
      • Always have a baseline to compare your model to.
      • Some visualisations are hard to understand. Try to look at each plot and decide if you can understand it without information about how it was generated.