search

M-Nauta / TCDF

Temporal Causal Discovery Framework (PyTorch): discovering causal relationships between time series
GNU General Public License v3.0
482 stars 106 forks source link

Scaling affecting results #14

Closed kapkic closed 1 year ago

kapkic commented 1 year ago

Hello, Thank you for the paper and the work. I have a question regarding how TCDF interprets scaling.

Shouldn't TCDF be scale invariant, as the causes are related to the relationship between variables and values, not to their scaling? For example, when we take one of the Financial dataset files (let's assume random-rels_20_1D_returns30007000_header in this example) and multiply every value by 100, ideally we should be getting the same causal relationships since we only change the scale of the data.

But in practice, the raw data returns 35 Causal predictions (X causes Y in T timesteps), and the raw data multiplied by 100 returns 51 predictions. I have noticed that in AD-DSTCN, a softmax function is applied, but I am unsure why even with that the scale creates a difference.

Thanks!

M-Nauta commented 1 year ago

Interesting finding. We didn't look into scaling in much detail. Softmax is not scale-invariant so together with the significance that might be related to the issue. If custom datasets have a different scaling than the datasets we used, I would recommend to scale the datasets accordingly.