This repository documents my approach for the GDG Kaggle competition during the 2024-25 recruitment process. I successfully advanced to the third round, where I tackled this challenge focused on predicting temperature changes along a rod.
The dataset provided temperature measurements at 50 discrete points on a rod, with these temperatures changing over time. Based on the initial temperature values, our task was to predict the temperature at each of these points in the subsequent time step. Here’s a summary of the key dataset features:
- L: Length of the rod
- nx: Number of divisions along the rod
- val_0 to val_49: Temperature values at each division point.
At first, I considered using multi-linear regression, where the rod could be represented by a line, with temperature predictions made based on the distance from the origin. However, the high dimensionality of the dataset made this approach inefficient and likely unoptimized.
I then explored using PCA to reduce dimensionality. However, with only 50 dimensions and over 1,000 values to predict, reducing further risked diminishing the model’s accuracy.
Given the sequential nature of the temperature data, I opted for a BiLSTM or BiGRU model, treating the temperature distribution along the rod as a sequence. This approach captured temporal dependencies more effectively, providing a robust solution to the prediction task.