Question about feat_static_cat

[NielsRogge]

In the extended tutorial, feat_static_cat is explained as: "static (over time) categorical features, list with dimension equal to the number of features"

What are examples of static categorical features? Are these features to identify individual time series, in case one wants to perform multivariate time series forecasting?

[timoschowski]

Example: If you have 5 different categorical features, the dimension needs to be 5.

Typically, you can have a categorical feature to identify groups of time series. E.g., all time series with categorical features A have a certain time of seasonality or a high volume or something, whereas time series with another feature behave differently.

[alexhallam]

Hi @timoschowski, I was just searching for feat_static_cat guidance also. If you don't mind I have a simple reproducible example that would clarify this for me.

At the moment this trains fine, but it is not using key as a categorical feature. Could you show me how to modify this code to make that happen?

from gluonts.dataset import common
from gluonts.model import deepar
from gluonts.model.deepar import DeepAREstimator
from gluonts.distribution.neg_binomial import NegativeBinomialOutput
from gluonts.trainer import Trainer

import pandas as pd
data = "https://gist.githubusercontent.com/alexhallam/f0410e546fde69e8bda92e89878386fe/raw/b582d71840de595e8329efec899ec3bfe17db32a/car_train.csv"
df = pd.read_csv(data)
df_cats = pd.get_dummies(df['key'], prefix='part')
data = common.ListDataset([{
    "start": df.date[0],
    "target": df.value,
    'feat_static_cat': df_cats
}],freq="d")
estimator = DeepAREstimator(
    prediction_length=6,
    freq="D",
    distr_output = NegativeBinomialOutput(),
    #use_feat_static_cat=True,
    trainer=Trainer(
        epochs=5
    )
)
predictor = estimator.train(training_data=data)
prediction = next(predictor.predict(data))
print(prediction.mean)
prediction.plot(output_file='graph.png')

[NielsRogge]

@alexhallam is this a single or multivariate time series?

[alexhallam]

Well, I know that I can model each of these series individually, but I also wanted to compare those results with using each time series key as a feature. A pooled vs non-pooled set up. What I am trying to do above is use each time series as a categorical feature.

[StatMixedML]

@alexhallam You might want to try the following for creating your dataset (using the code-snippet from above)

from gluonts.dataset.common import ListDataset
from gluonts.dataset.field_names import FieldName

data_list = ListDataset(
    [
        {FieldName.START: pd.Timestamp(df_cats.iloc[:, col_idx].index[0], freq = "d"),
         FieldName.TARGET: df_cats.iloc[:, col_idx].values,
         FieldName.ITEM_ID: df_cats.columns[col_idx],
         FieldName.FEAT_STATIC_CAT: np.array([col_idx]),
        }
        for col_idx in range(df_cats.shape[1])
    ],
    freq="d"
)

It creates the dataset that can be used for model training. You can check out some of the entries of gluonts.dataset.common.ListDataset

data_list.list_data[0]

showing

{'start': Timestamp('1970-01-01 00:00:00', freq='D'),
 'target': array([0, 0, 0, ..., 0, 0, 0], dtype=uint8),
 'item_id': 'part_10055165',
 'feat_static_cat': array([0])}

The categorical feature feat_static_cat corresponds to the numeric index of the loop, item_id is the column name. DeepAR learns an embedding for each categorical feature feat_static_cat.

[jonbeckman]

Hey @StatMixedML, I would appreciate if you could clarify something for me. In the example you provided, supplying df_cats.shape[1] or n number of time-series will result in n number of future time-series predicted (N:N)?

The reason I ask is because I am trying to understand the difference between your code snippet and something like this:

data_list = ListDataset(
    [
        {
            FieldName.START: pd.Timestamp(df_cats.iloc[:, 0].index[0], freq = "d"),
            FieldName.TARGET: df_cats.iloc[:, 0].values,
            FieldName.FEAT_DYNAMIC_REAL: [
                df[:, col_idx].values for col_idx in range(df_cats.shape[1]) if col_idx != 0
            ],
        }
    ],
    freq="d"
)

Where supplying n time-series will predict only one time-series (N:1)

[alexhallam]

This answers the question for me. What do you think @NielsRogge?

[ShirleyMgit]

Hi can use_feat_static_cat be used in deepVAR? it seems that the model does not recognize it

[mbohlkeschneider]

Hi @ShirleyMgit,

DeepVAR does not use feat_static_cat at the moment. It is a bit odd in the interface that you can still set it (and probably should be removed) but the network itself does not use it. In the multivariate case, it also only make sense if you have different multivariate time series that represent different categories. Most people simply use one, big multivariate dataset that includes all time series and there the concept of a categorical variable does not make sense (because all the time series are present at each forward pass).

Hope that answers your question.

[ShirleyMgit]

Yes, in my problem I do have multiple multivariate timeseries that represent different categories. Is there any network that support it for multivariate timeseries? Is it possible for me to extand deepVAR to include it? Can I see the full source code? Thanks

[mbohlkeschneider]

Hi @ShirleyMgit,

I see. I'm not aware of a model that does that in GluonTS. The DeepVAR model code is here: https://github.com/awslabs/gluonts/tree/dev/src/gluonts/mx/model/deepvar. If you add that, feel free to send us a PR!

[ShirleyMgit]

By the way, if there are no categorical variables why deepVAR gets 'cardinality' and 'embedding dimensions' as parameters?

[ShirleyMgit]

Hi,
I am trying to extend the network with an embedding layer for static features (similarly to deepAR).
Therefore, I am trying to understand the code.
Can you please explain what is the role of 'self.embed' and repeated_index_embeddings?
thanks

[ShirleyMgit]

also it seems that the code does not deal with the observed target in calculating the loss and it is not documented