Transformers for Time Series Forecasting -- 3
Budget: ₹400 – ₹750 INR
The code for time series forecasting using transformers is available for prediction of rainfall data and the result shown is just the time series. The following things need to be added in the code:
1. How do the results compare? What is the average and standard deviation? What about the box plot comparison of the summary of predicted and true curves?
2. Every ML experiment needs a summary with standard measures such as accuracy, precision, recall, sensitivity, specificity, F1 measure, etc.
3. We need a comparison to methods different from Transformers to show that this method is state-of-the-art and better than previously known methods. You can use a simple RNN, perhaps with the same architecture that has been used (except without attention) and then compare.
4. Some experiments with varying the parameters. For example, 200 epochs is done. How worse are the results with 100 epochs? How much better are they with 1000 or 10,000 epochs? You can ask similar questions with other hyperparameters. What if you train with data on every alternative day over the same period? Are the results almost as good?
1. How do the results compare? What is the average and standard deviation? What about the box plot comparison of the summary of predicted and true curves?
2. Every ML experiment needs a summary with standard measures such as accuracy, precision, recall, sensitivity, specificity, F1 measure, etc.
3. We need a comparison to methods different from Transformers to show that this method is state-of-the-art and better than previously known methods. You can use a simple RNN, perhaps with the same architecture that has been used (except without attention) and then compare.
4. Some experiments with varying the parameters. For example, 200 epochs is done. How worse are the results with 100 epochs? How much better are they with 1000 or 10,000 epochs? You can ask similar questions with other hyperparameters. What if you train with data on every alternative day over the same period? Are the results almost as good?