feat(karthik_model): neural network training on NBA CSV data

[Kravi001]

What I changed

• Added karthik_model.py in src/mini_nn/ to train a neural network on NBA CSV data.
• Loads data from Data/PlayerStatistics.csv, preprocesses features, handles NaNs, and manages class imbalance.
• Trains a model to predict whether a player scores ≥20 points and saves predictions (gitignored).

Why I changed it

• Provides a fully working neural network pipeline for NBA CSV data.
• Allows generating predictions without connecting to a database.

How did I test it

• Ran the script locally on the full dataset (1,655,736 rows, 19 features).
• Observed class imbalance: 0 → 0.8699, 1 → 0.1301; positive class weight = 6.69.
• Training metrics:
  • Epoch 10 | Train Loss: 0.1343 | Val F1: 0.9847
  • Epoch 20 | Train Loss: 0.0942 | Val F1: 0.9400
  • Epoch 30 | Train Loss: 0.0690 | Val F1: 0.9919
  • Epoch 40 | Train Loss: 0.0624 | Val F1: 0.9946
  • Epoch 50 | Train Loss: 0.0418 | Val F1: 0.9990
  • Epoch 60 | Train Loss: 0.0330 | Val F1: 0.9997
  • Epoch 70 | Train Loss: 0.0264 | Val F1: 1.0000
• Early stopping triggered at epoch 72.
• Test results:
  • Best Epoch: 57
  • Test Accuracy: 0.9999
  • Test F1 Score: 0.9997
  • Test ROC-AUC: 1.0000

[JonathanPLev]

can you add your results? curious to know how well this model did.

edit: sorry i see it actually. how are you getting 100% test accuracy? I think you are having leaking or overfit data.

[JonathanPLev]

you are definitely having leaking data here because you are having random split

[JonathanPLev]

youre also not dropping any of the lines that leak data to the outcome. like youre not removing any stats like points or anything thats current game.