Exceeded maximum of 40 nodes using cross-sectional functions

General
1

Encountered this error message testing a new ranking system. Is this limit still necessary or can it be raised? And is the limit on the number of nodes using FRank, etc calls or on the number of calls to FRank, etc. within the system?

Thanks,
Feldy

2

Checking on the limits.

However, using FRank in a ranking system seems redundant. What’s the use-case please?

3

Thanks Marco.

I’m looking at the interplay of two features, so instead of just two separate factor nodes for A and B, I create a formula node for FRank(A)*FRank(B) to magnify cases when both are high (or low).

That’s also why I asked if the limit was on function calls vs formula nodes, as I didn’t have 40+ formula nodes containing cross sectionals, to my knowledge at least. But I had over 40 calls to FRank in the system as this pattern yielded two FRank calls per node where it was used.

Thanks.

4

I can confirm there’s a hard limit of 40 “formula” functions in Ranking. They do take up resources so we set conservative limits. It’s the first time I think someone has hit the limit.

The use-case is interesting. For example if two factors rank 60/60 then they average 60; the same as if one ranked 100 and the other 20. Doing a product does favor 60/60 quite a bit more than 100/20.

Maybe we can offer an additional option at the Composite node level which currently offers: Higher better, Lower better, and Summation. We could add Product as well (which is then normalized back to a score from 100-0). Would be the same thing you are doing with FRank but with considerable less compute effort.

1 Like
5

Thanks, your Composite product proposal sounds like an elegant solution to handle this use case.

6

Great Idea. Maybe you can add at the same time the following composite options as well:
max (returns the highest rank value among all nodes of the composite)
min
median

7

Multiplying factors is done in multiple regression to find interaction of variables.

Agee, interesting.

Edit: is multiplying z-scores more or less computer intestive? Better results? Or worse?

Jim

8

Good question re: z-scores. Had considered it but but avoided it in this use case in favor of FRank to avoid sign-flipping issues, e.g. ZScore(-A)*ZScore(-B) mapping to the same value as ZScore(+A)*ZScore(+B). But it could be interesting depending on the range and meaning of the underlying factors.

9

I do not know if z-score will work well or not. But if you find it promising using eval could possibly mitigate some of the sign-flipping issues.

E.g., Eval(A <=0 and B<=0, -1AB,A*B)

Frank() should work well and might be better as it is non-parametric (less sensitive to outliers).

1 Like