Hello everyone, we have encountered some problems in data completion. I would like to ask you if there is a better and more efficient data completion method, thank you
What data completion method are you using right now?
Also, what libraries are you working with?
Hi. Please what do you mean by data completion?
I think the TS meant the method he filled NAs.
I mean how to deal with nan data ，excuse me do you have any recommended method
Yes I have problem how to deal with nan data ，so Do you have any recommended method
We are not good at geology, we are still using data completion methods commonly used in machine learning, such as linear completion，so Do you have any recommendations
You could try ffill/bfill and check for the score in your validation set. Given the sequential nature of the data, it should give good results.
One common strategy used in log analysis is to train either a multi-linear regression or random forest to predict what missing log values should be. This works well for logs like the density (rhob), neuton (nphi), and sonic (dt). It’s not going to work for things like caliper, etc.
You could also try to use something like the Gardener-relation, which is an empirical model between RhoB and DT that allows you to guestimate one from another.
A final track would be to take a mineral inversion approach to predict the missing curves. How this works is you take a matrix of potential minerals and their properties and the data that is present, and make a prediction on the mineral volumes (via inverting the matrices)… usually you will be in a quite underdetermined state, but that is what it is. Then, using the mineral proportions you just calculated, and the mineral properties, you can predict the values of the missing logs. For example, if you are missing the Photoelectric factor curve (PE) you can predict it via something like:
Vol. of mineral1 * PE mineral1 + Vol. mineral2* PEmineral2…
This process is outline in Doveton’s excellent book, “Mathematical Petrophysics”.
Thank you very much for your patience, we will try the method you provided