mthorrell
mthorrell
I'm new to Julia, so I may be doing something wrong. But the following code pulled basically straight from the example on Github doesn't run for me. ``` using GLM...
Problem: Both XGB and LightGBM boosting objects have explicit methods that update the Booster by exactly one round (see for example https://lightgbm.readthedocs.io/en/latest/pythonapi/lightgbm.Booster.html#lightgbm.Booster.update). These methods offer a constant-time way to update...
When we changed the numeric representation of datetimes to seconds (rather than the default from pandas of nanoseconds) performance actually dropped. Looking at the affected cases, it seems that nano-seconds...
True batch training will use new batches for each training round. This is the usual way NNs are trained. Unfortunately, one aspect of GB packages making them fast is prediction...
Is there a way to get the diagonal of the hessian that is more efficient than the current way?
Raw code to be optimized ```python import numpy as np import pandas as pd from gboost_module import xgbmodule def meas_df_to_long_df( meas_df: pd.DataFrame, resolution: int ): assert 'time' in meas_df assert...
Maybe just implement a quantile regression on top of predictions? similar to https://xgboost.readthedocs.io/en/latest/python/examples/quantile_regression.html
changepoints would meet prediction level feature parity with prophet. The remaining big feature to match prophet functionality would be confidence intervals.
CatModule (in experimental/catmodule.py), which relies on catboost as its gbdt backend, does not have performant updates and thus, for larger problems, can be computationally infeasible. It is also just generally...