Irregation need Competation

Turning tabular data into high-performing deep learning solutions was the real challenge.

In this competition, we were restricted to using only Deep Learning models on a tabular dataset — a limitation that pushed us to engineer a more precise and optimized architecture to surpass our 96% target score.

Our approach combined advanced preprocessing with a custom Deep Tabular Model design:

• Categorical Embedding Layers for features such as Crop Type and Zone

• Periodic (Sinusoidal) Embeddings to capture seasonal and cyclical numerical patterns

• A stacked Dense Neural Network with ReLU activation, Dropout regularization, and a Softmax classification head for robust performance and generalization

This project was a strong example of how carefully designed architectures and feature representations can significantly enhance deep learning performance on structured data.