Unexpected gain on `check_params`

[jjbrophy47]

Hello, I like this package a lot and i've been trying to use it with my own models lately, but I have been encountering this assertion: assert abs(expect_gain - parent['gain']) < 1.e-2 a number of times.

I have been training an xgboost model with 1000 trees, and this assertion gets thrown on about a dozen of those trees on average. I was just wondering if it's something I am doing, or if it's a problem with the model that XGBoost dumps out in text form not always being correct.

Thanks for any input!

[gameofdimension]

this function ensure you passed the params (lambda and eta) exactly as you trained the xgboost model, if you are sure about this you can safely comment out this line.
for the specific error, i can give nothing useful without details.

[eranmosh3]

Here's 2 errors I get for no reason:
(using on private data set which is solid. I can train an xgboost model on it)

line 56, in model2table
k,v = p.split('=')
ValueError: not enough values to unpack (expected 2, got 1)

Process finished with exit code 1

line 74, in model2table
node_lst[node_idx] = d
IndexError: list assignment index out of range

Process finished with exit code 1

[2g-XzenG]

Hi @hanooka,
I encounter similar errors, which is caused by the white spaces within my feature names, like "feature 1".
Changing the feature name to "feature_1" fix the error.
Hope it helps,
Thanks

[charlie9526]

Hi @ALL

1. I am trying this explainer on XGB regressor. The problem is that, the sum of the contribution is not same as the prediction of the XGB regressor model. There is gap of 0.5 between those two values.
2. In the medium article of the R package of this library has explained that the contribution is log odds. But in regression the contribution is a continuous value which the collection is same to the prediction-0.5. Here also in the medium article, the graph shows the 0.5 line, but not mentioned in the calculations about that 0.5.