LightGBMCV
Time series cross validation with LightGBM.
source
LightGBMCV
LightGBMCV (freq:Union[int,str], lags:Optional[Iterable[int]]=None, lag_t ransforms:Optional[Dict[int,List[Union[Callable,Tuple[Callabl e,Any]]]]]=None, date_features:Optional[Iterable[Union[str,Callable]]]=None, num_threads:int=1, target_transforms:Optional[List[Union[mlfo recast.target_transforms.BaseTargetTransform,mlforecast.targe t_transforms._BaseGroupedArrayTargetTransform]]]=None)
Create LightGBM CV object.
Type | Default | Details | |
---|---|---|---|
freq | Union | Pandas offset alias, e.g. ‘D’, ‘W-THU’ or integer denoting the frequency of the series. | |
lags | Optional | None | Lags of the target to use as features. |
lag_transforms | Optional | None | Mapping of target lags to their transformations. |
date_features | Optional | None | Features computed from the dates. Can be pandas date attributes or functions that will take the dates as input. |
num_threads | int | 1 | Number of threads to use when computing the features. |
target_transforms | Optional | None | Transformations that will be applied to the target before computing the features and restored after the forecasting step. |
Example
This shows an example with just 4 series of the M4 dataset. If you want to run it yourself on all of them, you can refer to this notebook.
import random
from datasetsforecast.m4 import M4, M4Info
from fastcore.test import test_eq, test_fail
from mlforecast.target_transforms import Differences
from nbdev import show_doc
from window_ops.ewm import ewm_mean
from window_ops.rolling import rolling_mean, seasonal_rolling_mean
group = 'Hourly'
await M4.async_download('data', group=group)
df, *_ = M4.load(directory='data', group=group)
df['ds'] = df['ds'].astype('int')
ids = df['unique_id'].unique()
random.seed(0)
sample_ids = random.choices(ids, k=4)
sample_df = df[df['unique_id'].isin(sample_ids)]
sample_df
unique_id | ds | y | |
---|---|---|---|
86796 | H196 | 1 | 11.8 |
86797 | H196 | 2 | 11.4 |
86798 | H196 | 3 | 11.1 |
86799 | H196 | 4 | 10.8 |
86800 | H196 | 5 | 10.6 |
… | … | … | … |
325235 | H413 | 1004 | 99.0 |
325236 | H413 | 1005 | 88.0 |
325237 | H413 | 1006 | 47.0 |
325238 | H413 | 1007 | 41.0 |
325239 | H413 | 1008 | 34.0 |
info = M4Info[group]
horizon = info.horizon
valid = sample_df.groupby('unique_id').tail(horizon)
train = sample_df.drop(valid.index)
train.shape, valid.shape
((3840, 3), (192, 3))
What LightGBMCV does is emulate LightGBM’s cv function where several Boosters are trained simultaneously on different partitions of the data, that is, one boosting iteration is performed on all of them at a time. This allows to have an estimate of the error by iteration, so if we combine this with early stopping we can find the best iteration to train a final model using all the data or even use these individual models’ predictions to compute an ensemble.
In order to have a good estimate of the forecasting performance of our
model we compute predictions for the whole test period and compute a
metric on that. Since this step can slow down training, there’s an
eval_every
parameter that can be used to control this, that is, if
eval_every=10
(the default) every 10 boosting iterations we’re going
to compute forecasts for the complete window and report the error.
We also have early stopping parameters:
early_stopping_evals
: how many evaluations of the full window should we go without improving to stop training?early_stopping_pct
: what’s the minimum percentage improvement we want in theseearly_stopping_evals
in order to keep training?
This makes the LightGBMCV class a good tool to quickly test different configurations of the model. Consider the following example, where we’re going to try to find out which features can improve the performance of our model. We start just using lags.
static_fit_config = dict(
n_windows=2,
h=horizon,
params={'verbose': -1},
compute_cv_preds=True,
)
cv = LightGBMCV(
freq=1,
lags=[24 * (i+1) for i in range(7)], # one week of lags
)
source
LightGBMCV.fit
LightGBMCV.fit (df:pandas.core.frame.DataFrame, n_windows:int, h:int, id_col:str='unique_id', time_col:str='ds', target_col:str='y', step_size:Optional[int]=None, num_iterations:int=100, params:Optional[Dict[str,Any]]=None, static_features:Optional[List[str]]=None, dropna:bool=True, keep_last_n:Optional[int]=None, eval_every:int=10, weights:Optional[Sequence[float]]=None, metric:Union[str,Callable]='mape', verbose_eval:bool=True, early_stopping_evals:int=2, early_stopping_pct:float=0.01, compute_cv_preds:bool=False, before_predict_callback:Optional[Callable]=None, after_predict_callback:Optional[Callable]=None, input_size:Optional[int]=None)
Train boosters simultaneously and assess their performance on the complete forecasting window.
Type | Default | Details | |
---|---|---|---|
df | DataFrame | Series data in long format. | |
n_windows | int | Number of windows to evaluate. | |
h | int | Forecast horizon. | |
id_col | str | unique_id | Column that identifies each serie. |
time_col | str | ds | Column that identifies each timestep, its values can be timestamps or integers. |
target_col | str | y | Column that contains the target. |
step_size | Optional | None | Step size between each cross validation window. If None it will be equal to h . |
num_iterations | int | 100 | Maximum number of boosting iterations to run. |
params | Optional | None | Parameters to be passed to the LightGBM Boosters. |
static_features | Optional | None | Names of the features that are static and will be repeated when forecasting. |
dropna | bool | True | Drop rows with missing values produced by the transformations. |
keep_last_n | Optional | None | Keep only these many records from each serie for the forecasting step. Can save time and memory if your features allow it. |
eval_every | int | 10 | Number of boosting iterations to train before evaluating on the whole forecast window. |
weights | Optional | None | Weights to multiply the metric of each window. If None, all windows have the same weight. |
metric | Union | mape | Metric used to assess the performance of the models and perform early stopping. |
verbose_eval | bool | True | Print the metrics of each evaluation. |
early_stopping_evals | int | 2 | Maximum number of evaluations to run without improvement. |
early_stopping_pct | float | 0.01 | Minimum percentage improvement in metric value in early_stopping_evals evaluations. |
compute_cv_preds | bool | False | Compute predictions for each window after finding the best iteration. |
before_predict_callback | Optional | None | Function to call on the features before computing the predictions. This function will take the input dataframe that will be passed to the model for predicting and should return a dataframe with the same structure. The series identifier is on the index. |
after_predict_callback | Optional | None | Function to call on the predictions before updating the targets. This function will take a pandas Series with the predictions and should return another one with the same structure. The series identifier is on the index. |
input_size | Optional | None | Maximum training samples per serie in each window. If None, will use an expanding window. |
Returns | List | List of (boosting rounds, metric value) tuples. |
hist = cv.fit(train, **static_fit_config)
[LightGBM] [Info] Start training from score 51.745632
[10] mape: 0.590690
[20] mape: 0.251093
[30] mape: 0.143643
[40] mape: 0.109723
[50] mape: 0.102099
[60] mape: 0.099448
[70] mape: 0.098349
[80] mape: 0.098006
[90] mape: 0.098718
Early stopping at round 90
Using best iteration: 80
By setting compute_cv_preds
we get the predictions from each model on
their corresponding validation fold.
cv.cv_preds_
unique_id | ds | y | Booster | window | |
---|---|---|---|---|---|
0 | H196 | 865 | 15.5 | 15.522924 | 0 |
1 | H196 | 866 | 15.1 | 14.985832 | 0 |
2 | H196 | 867 | 14.8 | 14.667901 | 0 |
3 | H196 | 868 | 14.4 | 14.514592 | 0 |
4 | H196 | 869 | 14.2 | 14.035793 | 0 |
… | … | … | … | … | … |
187 | H413 | 956 | 59.0 | 77.227905 | 1 |
188 | H413 | 957 | 58.0 | 80.589641 | 1 |
189 | H413 | 958 | 53.0 | 53.986834 | 1 |
190 | H413 | 959 | 38.0 | 36.749786 | 1 |
191 | H413 | 960 | 46.0 | 36.281225 | 1 |
The individual models we trained are saved, so calling predict
returns
the predictions from every model trained.
source
LightGBMCV.predict
LightGBMCV.predict (h:int, before_predict_callback:Optional[Callable]=None, after_predict_callback:Optional[Callable]=None, X_df:Optional[pandas.core.frame.DataFrame]=None)
Compute predictions with each of the trained boosters.
Type | Default | Details | |
---|---|---|---|
h | int | Forecast horizon. | |
before_predict_callback | Optional | None | Function to call on the features before computing the predictions. This function will take the input dataframe that will be passed to the model for predicting and should return a dataframe with the same structure. The series identifier is on the index. |
after_predict_callback | Optional | None | Function to call on the predictions before updating the targets. This function will take a pandas Series with the predictions and should return another one with the same structure. The series identifier is on the index. |
X_df | Optional | None | Dataframe with the future exogenous features. Should have the id column and the time column. |
Returns | DataFrame | Predictions for each serie and timestep, with one column per window. |
preds = cv.predict(horizon)
preds
unique_id | ds | Booster0 | Booster1 | |
---|---|---|---|---|
0 | H196 | 961 | 15.670252 | 15.848888 |
1 | H196 | 962 | 15.522924 | 15.697399 |
2 | H196 | 963 | 14.985832 | 15.166213 |
3 | H196 | 964 | 14.985832 | 14.723238 |
4 | H196 | 965 | 14.562152 | 14.451092 |
… | … | … | … | … |
187 | H413 | 1004 | 70.695242 | 65.917620 |
188 | H413 | 1005 | 66.216580 | 62.615788 |
189 | H413 | 1006 | 63.896573 | 67.848598 |
190 | H413 | 1007 | 46.922797 | 50.981950 |
191 | H413 | 1008 | 45.006541 | 42.752819 |
We can average these predictions and evaluate them.
def evaluate_on_valid(preds):
preds = preds.copy()
preds['final_prediction'] = preds.drop(columns=['unique_id', 'ds']).mean(1)
merged = preds.merge(valid, on=['unique_id', 'ds'])
merged['abs_err'] = abs(merged['final_prediction'] - merged['y']) / merged['y']
return merged.groupby('unique_id')['abs_err'].mean().mean()
eval1 = evaluate_on_valid(preds)
eval1
0.11036194712311806
Now, since these series are hourly, maybe we can try to remove the daily seasonality by taking the 168th (24 * 7) difference, that is, substract the value at the same hour from one week ago, thus our target will be . The features will be computed from this target and when we predict they will be automatically re-applied.
cv2 = LightGBMCV(
freq=1,
target_transforms=[Differences([24 * 7])],
lags=[24 * (i+1) for i in range(7)],
)
hist2 = cv2.fit(train, **static_fit_config)
[LightGBM] [Info] Start training from score 0.519010
[10] mape: 0.089024
[20] mape: 0.090683
[30] mape: 0.092316
Early stopping at round 30
Using best iteration: 10
assert hist2[-1][1] < hist[-1][1]
Nice! We achieve a better score in less iterations. Let’s see if this improvement translates to the validation set as well.
preds2 = cv2.predict(horizon)
eval2 = evaluate_on_valid(preds2)
eval2
0.08956665504570135
assert eval2 < eval1
Great! Maybe we can try some lag transforms now. We’ll try the seasonal
rolling mean that averages the values “every season”, that is, if we set
season_length=24
and window_size=7
then we’ll average the value at
the same hour for every day of the week.
cv3 = LightGBMCV(
freq=1,
target_transforms=[Differences([24 * 7])],
lags=[24 * (i+1) for i in range(7)],
lag_transforms={
48: [(seasonal_rolling_mean, 24, 7)],
},
)
hist3 = cv3.fit(train, **static_fit_config)
[LightGBM] [Info] Start training from score 0.273641
[10] mape: 0.086724
[20] mape: 0.088466
[30] mape: 0.090536
Early stopping at round 30
Using best iteration: 10
Seems like this is helping as well!
assert hist3[-1][1] < hist2[-1][1]
Does this reflect on the validation set?
preds3 = cv3.predict(horizon)
eval3 = evaluate_on_valid(preds3)
eval3
0.08961279023129345
Nice! mlforecast also supports date features, but in this case our time column is made from integers so there aren’t many possibilites here. As you can see this allows you to iterate faster and get better estimates of the forecasting performance you can expect from your model.
If you’re doing hyperparameter tuning it’s useful to be able to run a couple of iterations, assess the performance, and determine if this particular configuration isn’t promising and should be discarded. For example, optuna has pruners that you can call with your current score and it decides if the trial should be discarded. We’ll now show how to do that.
Since the CV requires a bit of setup, like the LightGBM datasets and the
internal features, we have this setup
method.
source
LightGBMCV.setup
LightGBMCV.setup (df:pandas.core.frame.DataFrame, n_windows:int, h:int, id_col:str='unique_id', time_col:str='ds', target_col:str='y', step_size:Optional[int]=None, params:Optional[Dict[str,Any]]=None, static_features:Optional[List[str]]=None, dropna:bool=True, keep_last_n:Optional[int]=None, weights:Optional[Sequence[float]]=None, metric:Union[str,Callable]='mape', input_size:Optional[int]=None)
Initialize internal data structures to iteratively train the boosters. Use this before calling partial_fit.
Type | Default | Details | |
---|---|---|---|
df | DataFrame | Series data in long format. | |
n_windows | int | Number of windows to evaluate. | |
h | int | Forecast horizon. | |
id_col | str | unique_id | Column that identifies each serie. |
time_col | str | ds | Column that identifies each timestep, its values can be timestamps or integers. |
target_col | str | y | Column that contains the target. |
step_size | Optional | None | Step size between each cross validation window. If None it will be equal to h . |
params | Optional | None | Parameters to be passed to the LightGBM Boosters. |
static_features | Optional | None | Names of the features that are static and will be repeated when forecasting. |
dropna | bool | True | Drop rows with missing values produced by the transformations. |
keep_last_n | Optional | None | Keep only these many records from each serie for the forecasting step. Can save time and memory if your features allow it. |
weights | Optional | None | Weights to multiply the metric of each window. If None, all windows have the same weight. |
metric | Union | mape | Metric used to assess the performance of the models and perform early stopping. |
input_size | Optional | None | Maximum training samples per serie in each window. If None, will use an expanding window. |
Returns | LightGBMCV | CV object with internal data structures for partial_fit. |
cv4 = LightGBMCV(
freq=1,
lags=[24 * (i+1) for i in range(7)],
)
cv4.setup(
train,
n_windows=2,
h=horizon,
params={'verbose': -1},
)
LightGBMCV(freq=1, lag_features=['lag24', 'lag48', 'lag72', 'lag96', 'lag120', 'lag144', 'lag168'], date_features=[], num_threads=1, bst_threads=8)
Once we have this we can call partial_fit
to only train for some
iterations and return the score of the forecast window.
source
LightGBMCV.partial_fit
LightGBMCV.partial_fit (num_iterations:int, before_predict_callback:Optional[Callable]=None, after_predict_callback:Optional[Callable]=None)
Train the boosters for some iterations.
Type | Default | Details | |
---|---|---|---|
num_iterations | int | Number of boosting iterations to run | |
before_predict_callback | Optional | None | Function to call on the features before computing the predictions. This function will take the input dataframe that will be passed to the model for predicting and should return a dataframe with the same structure. The series identifier is on the index. |
after_predict_callback | Optional | None | Function to call on the predictions before updating the targets. This function will take a pandas Series with the predictions and should return another one with the same structure. The series identifier is on the index. |
Returns | float | Weighted metric after training for num_iterations. |
score = cv4.partial_fit(10)
score
[LightGBM] [Info] Start training from score 51.745632
0.5906900462828166
This is equal to the first evaluation from our first example.
assert hist[0][1] == score
We can now use this score to decide if this configuration is promising. If we want to we can train some more iterations.
score2 = cv4.partial_fit(20)
This is now equal to our third metric from the first example, since this time we trained for 20 iterations.
assert hist[2][1] == score2
Using a custom metric
The built-in metrics are MAPE and RMSE, which are computed by serie and then averaged across all series. If you want to do something different or use a different metric entirely, you can define your own metric like the following:
def weighted_mape(
y_true: pd.Series,
y_pred: pd.Series,
ids: pd.Series,
dates: pd.Series,
):
"""Weighs the MAPE by the magnitude of the series values"""
abs_pct_err = abs(y_true - y_pred) / abs(y_true)
mape_by_serie = abs_pct_err.groupby(ids).mean()
totals_per_serie = y_pred.groupby(ids).sum()
series_weights = totals_per_serie / totals_per_serie.sum()
return (mape_by_serie * series_weights).sum()
_ = LightGBMCV(
freq=1,
lags=[24 * (i+1) for i in range(7)],
).fit(
train,
n_windows=2,
h=horizon,
params={'verbose': -1},
metric=weighted_mape,
)
[LightGBM] [Info] Start training from score 51.745632
[10] weighted_mape: 0.480353
[20] weighted_mape: 0.218670
[30] weighted_mape: 0.161706
[40] weighted_mape: 0.149992
[50] weighted_mape: 0.149024
[60] weighted_mape: 0.148496
Early stopping at round 60
Using best iteration: 60