Hey everyone,
I’m working on a stress detection project using tabular data from physiological signals — mainly HRV and EDA features (like statistical, time-domain, and frequency-domain features).
I was wondering what kind of model would work best for this kind of data. Should I stick with something like a Feedforward Neural Network or are there better options like TabNet, TabTransformer, or even some classical ML models?
I have also tried TabM model,but its acc is less than FFN model
Would love to hear what’s worked for others in similar use cases!
For tabular data, tree models seem to have good performance. And there are quite a few libraries for tree models, check kaggle to find out out more examples of such usages.
Thanks for the suggestion!
Yeah, I’ve seen tree models like XGBoost and LightGBM doing really well on tabular data. I’ll definitely check out some Kaggle notebooks for more ideas and examples. Appreciate the tip!
1 Like
Thanks for the suggestion!
Since I’m building a stress prediction model, I’ll eventually need to do calibration to improve confidence scores. With deep learning, I’ve seen that calibration can be done using simple code.
Just curious — is it possible to do similar calibration with tree-based machine learning models like XGBoost or Random Forest? If yes, are there any standard methods or tools you’d recommend?
You mean training the model with ground-truths, if so the process of training is similar!
Got it, thanks!
Yes, I meant calibrating the model outputs to better match the ground-truth probabilities. I was wondering more about post-training calibration techniques — like temperature scaling or Platt scaling — which are commonly used in deep learning.