[QST] The model performed poorly when I used my own data.

[fang-tech]

What is your question?
I trained the model using my own data, but I found that the data was inherently biased. Approximately 90% of the apps had an acceleration ratio greater than 1x, while only 10% had an acceleration ratio less than 1. This led to poor prediction results for negative optimization scenarios by the model. How should I proceed to address this bias in the model?

[fang-tech]

Do we have any troubleshooting procedures for model training?

[leewyang]

@fang-tech we don't have any official troubleshooting procedures, but you can try the following:

• Modify the sample_weight configuration in the qualx-conf.yaml file.
  • Either manually assign weights to positive/negative classes, or
  • Use "auto" to let the system compute the weighting.
• If your training set is relatively small, e.g. less than 100 sqlIDs, try gathering more training examples (targeting >1000 sqlIDs).
• Try adding more negative examples to your training set.

[fang-tech]

However, based on the comments and variable names, the values of these features should all be the same for this node. I'm sorry, I didn't clearly explain my problem in the context. The suggestion you mentioned above I have already tried. Our training data contains over 1600 SQLs. To increase the number of negative examples, I have tried random noise replication to enhance the data.

[fang-tech]

My biggest concern is that data skewness is almost inevitable for these acceleration components (if not, this component would be of no use in most cases). My model can already accurately predict high-speed tasks, with an MAPE of 21%, but it is difficult to predict low-speed tasks. It is very prone to misclassifying low-speed tasks as high-speed tasks. And QXS is almost always causing trouble. How can I balance my model? Or is it simply impossible to achieve this? Because of the inherent data imbalance, are there any solutions to this?