Hedgehog

Landing Page
User Input
Model Output
Architecture of the Model

Hedgehog is an AI based loan approval service. Hedgehog eases loan seeking and allows for scalability. We aim to incorporate a reinforcement leaning based model that will learn from what we deem and “good” loans and “bad” loans. The model will consider factors such as user profiles, fraud detection, lending capacity and loan performance. Our approach will eliminate all human errors and reduce human resource cost.

Salient Features:

1) Applicants with no credit history can apply and get loans.

2) Our solution achieves a very high accuracy with less than 1% false positives

Feature Processing, Selection and extraction

1) New features like total income were inferred from the dataset

2) Logarithmic transformations were applied on left skewed features

3) LabelEncoder was used on the extracted categorical variables

4) BalancedRandomForest was used to find feature importance

5) Smote-Tomek links was used to resample the complex imbalance data

Implemented Solution Architecture:

=> The implemented solution uses a novel stacking ensemble ML-model, which uses several weak learner algorithms and a meta learner. The weak learners are less powerful algorithms that may underfit/overfit on the complex problem, but plays a crucial role when used to train a meta-learner.

=> The weak learners used in the implemented solution are: RandomForest, LogisticRegression, XGBClassifier, KNeighbourClassifier

=> The meta/power learner used is the state-of-the-art LGBMClassifier, which is a non-linear gradient boosting tree-based ML algorithm.

Web App

Hedgehog's web app was built using Typescript React as a frontend with a flask API backend. The frontend makes requests to the API via Axios and uses React Router for client-side routing. Our API was initially built using Express but we migrated to Flask in order to integrate our ml model.