NUS 104 Submission

Inspiration

What it does

Predicts whether a customer is likely to remain a customer or convert

How we built it

We first used various methods such as downsampling and SMOTE to rebalance the data, as well as convert some data to be more useful (eg. convert date of birth to age). We then built a neural network with 3 layers using relu and sigmoid.

Challenges we ran into

It was difficult to achieve a good f1-score because the dataset was imbalanced, and when the rebalanced train set was applied on the imbalanced test set, we could not yield a good result

Accomplishments that we're proud of

The accuracy of our model was consistently high, and our f1 score did improve after multiple rounds of refinement

What we learned

How to deal with imbalanced dataset, and the techniques to build a neural network

What's next for NUS 104

We intend to improve our model by exploring other suitable ML models that may be more complex, such as deep learning

Built With

• python
• pytorch
• scikit-learn
• tensorflow