Third-party Model

PaddleTS provides make_ml_model interface allow users to build time series models based on scikit-learn for time series forecasting, pyod for time series anomaly detection, respectively. With this ability, users only need to develop a small piece of code to verify the feasibility and performance of their time-series-related ideas, which significantly improve the efficiency.

1. Third-party Model Integration for Time Series Forecasting

1.1 Minimal Example

Below is an example of how to make models based on sklearn.neighbors.KNeighborsRegressor for time series forecasting.

from paddlets.datasets.repository import get_dataset
from paddlets.models.ml_model_wrapper import make_ml_model

from sklearn.neighbors import KNeighborsRegressor

# prepare data
tsdataset = get_dataset("UNI_WTH")

# make model based on sklearn.neighbors.KNeighborsRegressor
model = make_ml_model(
    in_chunk_len=3,
    out_chunk_len=1,
    model_class=KNeighborsRegressor
)

# fit
model.fit(train_data=tsdataset)

# predict
predicted_ds = model.predict(tsdataset)
#             WetBulbCelsius
# 2014-01-01           -1.72

1.2 Convert MLDataLoader to Trainable / Predictable ndarray

The third-party library such as scikit-learn usually accepts numpy.ndarray data as fit and predict method inputs, while PaddleTS uses paddlets.models.forecasting.ml.adapter.ml_dataloader.MLDataLoader to represent trainable / predictable time series data. Thus, make_ml_model provides 2 optional arguments udf_ml_dataloader_to_fit_ndarray and udf_ml_dataloader_to_predict_ndarray allow users to convert MLDataLoader to an numpy.ndarray object.

By default, make_ml_model uses default_sklearn_ml_dataloader_to_fit_ndarray and default_sklearn_ml_dataloader_to_predict_ndarray to convert MLDataLoader to numpy.ndarray for fit and predict method, respectively. Also, users are able to develop user-defined convert functions to get expected trainable / predictable output.

from paddlets.datasets.repository import get_dataset
from paddlets.models.forecasting.ml.adapter.ml_dataloader import MLDataLoader
from paddlets.models.ml_model_wrapper import make_ml_model

from sklearn.neighbors import KNeighborsRegressor

# prepare data
tsdataset = get_dataset("UNI_WTH")

# develop user-defined convert functions
def udf_ml_dataloader_to_fit_ndarray(
    ml_dataloader: MLDataLoader,
    model_init_params: Dict[str, Any],
    in_chunk_len: int,
    skip_chunk_len: int,
    out_chunk_len: int
):
    # build and return converted numpy.ndarray object that sklearn model fit method accepts.
    pass

def udf_ml_dataloader_to_predict_ndarray(
    ml_dataloader: MLDataLoader,
    model_init_params: Dict[str, Any],
    in_chunk_len: int,
    skip_chunk_len: int,
    out_chunk_len: int
):
    # build and return converted numpy.ndarray object that sklearn model predict method accepts.
    pass

# pass the above 2 udf arguments to make_ml_model
model = make_ml_model(
    in_chunk_len=3,
    out_chunk_len=1,
    model_class=KNeighborsRegressor,
    udf_ml_dataloader_to_fit_ndarray=udf_ml_dataloader_to_fit_ndarray,
    udf_ml_dataloader_to_fit_ndarray=udf_ml_dataloader_to_predict_ndarray
)

# fit
model.fit(train_data=tsdataset)

# predict
predicted_ds = model.predict(tsdataset)

1.3 Multi-step forecasting

The time series models also support multi-timestep forecasting by calling recursive_predict .

from paddlets.datasets.repository import get_dataset
from paddlets.models.forecasting.ml.ml_model_wrapper import make_ml_model

# prepare data
tsdataset = get_dataset("UNI_WTH")

# make model
model = make_ml_model(
    in_chunk_len=3,
    out_chunk_len=1,
    model_class=KNeighborsRegressor
)

# fit
model.fit(train_data=tsdataset)

# recursively predict
recursively_predicted_ds = model.recursive_predict(tsdataset=tsdataset, predict_length=4)
#                      WetBulbCelsius
# 2014-01-01 00:00:00           -1.72
# 2014-01-01 01:00:00           -1.88
# 2014-01-01 02:00:00           -2.18
# 2014-01-01 03:00:00           -2.44

2 Third-party Model Integration for Time Series Anomaly Detection

2.1 Minimal Example

Below is an example of how to make models based on pyod.models.KNN for time series anomaly detection. As it has the unified interface as making time series forecasting models, you may refer to above section 1.1 to know how to define udf functions to customize the input ndarray of the built model.

from paddlets.datasets.repository import get_dataset
from paddlets.models.ml_model_wrapper import make_ml_model

from pyod.models.knn import KNN

# prepare data
tsdataset = get_dataset("WTH")

# make model based on pyod.models.knn.KNN
model = make_ml_model(
    in_chunk_len=3,
    model_class=KNN
)

# fit
model.fit(train_data=tsdataset)

# predict
predicted_ds = model.predict(tsdataset)
#                      WetBulbCelsius
# date
# 2010-01-01 02:00:00               0
# 2010-01-01 03:00:00               0
# 2010-01-01 04:00:00               1
# 2010-01-01 05:00:00               0
# 2010-01-01 06:00:00               0
# ...                             ...
# 2013-12-31 19:00:00               1
# 2013-12-31 20:00:00               1
# 2013-12-31 21:00:00               1
# 2013-12-31 22:00:00               1
# 2013-12-31 23:00:00               1

# [35062 rows x 1 columns]