Inspiration

Predicting purchase behaviour of customers is an important task for an insurance company and directly affects their business model. The dataset is fairly unclean and highly imbalanced. We decided to apply XGBoost, and efficient and powerful model for tabular data that has proven itself in many challenging competitions and datasets.

What it does

It is a propensity model that predicts the likelihood of buying a life or health insurance product within the next 3 months

How we built it

To build a xgboost model we decide to set up a 5-fold cross validation with shuffling, then loop through each fold of the cross validation. We used split_balance_and_pca to obtain training and testing sets and intialise XBboost on every set. We decided on using the AUC ROC as our evaluation metric due to accuracy being not very meaningful for highly imbalanced datasets.

Challenges we ran into

The target column of the dataset is rather imbalanced, with a majority of target column having negative result, it is difficult for us to build a meaningful model that do not simply overfit or give a high but unmeaningful accuracy rate.

Accomplishments that we're proud of

We managed to achieve a AUC ROC of 0.66 given the challenging nature of the dataset.

What we learned

The process of building an XGBoost model for highly unbalanced data likely provided valuable insights, including:

Understanding how the XGBoost model makes decisions and identifying influential features.
Analysing the impacts of different hyperparameters on model performance, especially in the context of imbalanced datasets. -Learning practical considerations for dealing with highly imbalanced data, including potential pitfalls and effective strategies.

What's next for DATATHON 2024 - CAT B team 217

Go to FINALS

|__ | __| /\ | \/ | |__ \/_ |____ | | | | |__ / \ | \ / | ) || | / / | | | | / /\ \ | |\/| | / / | | / /
| | | |__ / ____ | | | | / /_ | | / /
|| |__// __| |_| |_|||/_/