I’m an ecologist using machine learning algorithms (using the gradient boosting approach for the first time 😊) to gain insight into land use patterns. My question is related to a machine learning challenge in my research, specifically in its use for predicting a (continuous) outcome.
Can you explain why a predictor might have a relative influence but no marginal effect? So, for example see image, ‘Distance from a town’ was a fairly important predictor of land use in an area (say 15% relative influence) but, in further analysis (using the GBM and partial dependence plots) we found no visible relationship between the two. On the graph, the x axis is distance from the town in km and on the y is the marginal effect on the extent of a particular land cover type in an area.
What does it mean when a feature has an influence but no marginal effect on an outcome? Is it that, by itself this feature does not explain land use, or perhaps that this feature may have an indirect influence?
I may be wrong, and it is just an assumption, but distance from a town not is the main predictor of land use, at least in your case-study area. Is it possible to try the other predictors to find some of them which will have tangible marginal effects?
E.g., distance from factories or green cores with an area more than "x" where x is yet to be detected...
The specified dependence belongs to the functions of the interval type. On a certain segment, the predictor "distance to the city" has less statistical significance. Other predictors gain significance. This function is a system of equations
+ The processes of soil formation continued for thousands of years. Remote anthropogenic loading may not have a clear impact.
Maybe this variable has a strong correlation with other input variables. This may mean that it is useless, therefore, having not much impact on the results.
It's better to work with variable with strong correlation between itself and the target output and discard input variables with strong correlation between themselves only.
Anatolii Dyvnych thank you for the insight! Quick q:
Are we then saying, we find the distance from town/city, statistically, has no significant effect on land use change? OR, because of its relative influence (15%), we can say that its effect most probably depends on other factors and cannot be easily isolated.
The structure of the soil is highly dependent on the anthropogenic load (that is, the distance to the city). However, this parameter has a limited effect. My advice is to find a threshold level for this effect. After crossing this threshold, the function changes, it will take on a different formalized form. Other parameters will be more significant. Good luck!
The function consists of several parametric equations (with restrictions on the parameter - "distance to the city")
Anatolii Dyvnych Thanks for helping me understand, makes more sense now :)
Have a look at this paper
Conference Paper CALCULATING THE DEPENDENCY OF COMPONENTS OF OBSERVABLE NONLI...
and try to calculate the partial derivatives.Maybe that might help.
Kind regards, Wolfram
- Badges
- Science topic
- Similar topics
- Mathematics