Company Bankruptcy Prediction with Mlflow and deployment in EC2

1. Data Ingestion :

   • In Data Ingestion phase the data is first read as csv.
   • Then the data is split into training and testing and saved as csv file.

2. Data Transformation :

   • In this phase a ColumnTransformer Pipeline is created.
   • for Numeric Variables first SimpleImputer is applied with strategy median , then Standard Scaling is performed on numeric data.
   • for Categorical Variables SimpleImputer is applied with most frequent strategy, then ordinal encoding performed , after this data is scaled with Standard Scaler.
   • This preprocessor is saved as pickle file.

3. Model Training :

   • In this phase base model is tested . The best model found was catboost regressor.
   • After this hyperparameter tuning is performed on catboost and knn model.
   • A final VotingRegressor is created which will combine prediction of catboost, xgboost and knn models.
   • This model is saved as pickle file.

4. Prediction Pipeline :

   • This pipeline converts given data into dataframe and has various functions to load pickle files and predict the final results in python.

5. Flask App creation :

   • Flask app is created with User Interface to predict the gemstone prices inside a Web Application.

Exploratory Data Analysis Notebook

Link : EDA Notebook

Model Training Approach Notebook

Link : Model Training Notebook

Model Interpretation with LIME

Link : LIME Interpretation

Loan Approval Prediction - Random Forest Model and Deployed Web App

Project Purpose:

To demonstrate a full ML project from scratch to deployment to a web app.

Business Case:

To create a model that accurately predicts loan approval and automates the manual approval process.

Goal:

To accurately predict loan approvals without falsely approving too many loans that should be denied (minimize False Positives).

Deliverable:

A web app that end users can utilize to predict loan approvals using a Random Forest model on the backend.

1. Web App: https://jessramos2-loan-approval-random-forest-web-streamlit-app-47jl27.streamlitapp.com/

In Repository:

1. Data & data dictionary

2. Random Forest Model Creation (Loan Approval Model.ipynb)

3. Sample decision tree from random forest

4. Web App Python Code (streamlit_app.py)

5. Web app link and screenshot

Results and Model Evaluation:

From a business perspective, we want to avoid predicting a positve loan approval when it was actually denied (False Positives), so Precision will be our best measures for the model. This will ensure that we are accurately predicting loan approvals and that predicted approvals are actual approvals (precision).

Since this model produces very similar precision scores on both the training and testing data, it appears to be the best fit to maximize predictive power on the training dataset without overfitting and sacrificing predictability on the testing data.

Precision: Precision on the testing data is ~78%, which means that we don't have a large amount of False Positives. This is great, because as a business, we want to avoid predicting loans approvals that will have to be denied later.

Accuracy: Accuracy on the testing data is ~80% which means that the model correctly predicts 4/5 of the loans.

Recall: Recall on the testing data is 100% which means that the model accurately predicts all True Positives. This means that we will not miss out on any potential loan approvals (and revenue).

F1 Score: The F1 score on the testing data is ~88%, which is great since it takes into account both False Positives and False Negatives.

Business Impact:

End users will be able to use the web app built off of this model to predict loan approvals right in front of the borrower. There will be no missed revenue opportunities since the model captures all true approvals (recall is 100%), and only a small portion of borrowers predicted to be approved will actually be denied. This will speed up the manual approval process and allow the company to process more loans in less time, resulting in more clients and revenue.

Next Steps: Monitor performance and retrain model with more data as more data becomes available.