Model factory for clustered time series models

[LittleNeutrino]

Hi

I need to build a model factory for clustered time series models. I had a look at this accelerator, but I am still having trouble.

Specifically: how can I pass the clusters found by a clustering model as the segments for a segmented time series project?

thanks

Little

[matthias_kullowatz]

Hey @LittleNeutrino!

I've got some code here from our DataRobot Python API client to build a time series clustering project, select a clustering model from the leaderboard, and use it as the segmenter in a time series project. Note that the segment specification can be passed in the SegmentationTask.create() function as a time series clustering model package (as shown below), or as a feature name in the training dataset using the input user_defined_segment_id_column (see docs here).

#Import packages:
import datarobot as dr # recommended package version 3.1.0 or later
from datarobot import Dataset
from datarobot import Project
from datarobot import FeatureSettings
from datarobot import DatetimePartitioningSpecification
from datarobot.enums import UnsupervisedTypeEnum
from datarobot.enums import AUTOPILOT_MODE
from datarobot import Model
from datarobot.models import ModelPackage
from datarobot import SegmentationTask
from datarobot import PredictionServer
from datarobot import Deployment
from datarobot import BatchPredictionJob

#Create the time series clustering project
##here we create the project directly from an AI Catalog dataset, but of course it could be created through an existing data connection, from a local file, or with a pandas dataframe, as well.

cluster_project = dataset.create_project(project_name = "<project name>")
autopilot_cluster_list = [2,3,5,7,10]
cluster_featurelist = cluster_project.create_featurelist(
    "cluster", 
    [
        datetime_partition_column, 
        multiseries_id_column, 
        target
    ]
)
    
cluster_spec = DatetimePartitioningSpecification(
    use_time_series=True, 
    datetime_partition_column=datetime_partition_column, 
    multiseries_id_columns=[multiseries_id_column], 
    disable_holdout=True, 
    number_of_backtests=1)

cluster_project.analyze_and_model(
    partitioning_method=cluster_spec, 
    mode=AUTOPILOT_MODE.FULL_AUTO, 
    worker_count=-1, 
    unsupervised_mode=True, 
    unsupervised_type=UnsupervisedTypeEnum.CLUSTERING, 
    featurelist_id=cluster_featurelist.id, 
    autopilot_cluster_list=autopilot_cluster_list
)

#select a cluster blueprint from the leaderboard
cluster_models = cluster_project.get_models()
cluster_top_model = cluster_models[0]

#Create a model package from the top cluster model
cluster_model_package = ModelPackage.create(model_id=cluster_top_model.id)

#Create the segmented time series project
segment_project = dataset.create_project(project_name = "<project name>")

feature_settings = [FeatureSettings(feat_name, known_in_advance=True) for feat_name in kia_features]

segment_spec = dr.DatetimePartitioningSpecification(
    use_time_series=True,
    datetime_partition_column=datetime_partition_column,
    multiseries_id_columns=[multiseries_id_column],
    feature_settings=feature_settings
)
segment_task_results = SegmentationTask.create(
    project_id=segment_project.id,
    target=target,
    use_time_series=True,
    datetime_partition_column=datetime_partition_column,
    multiseries_id_columns=[multiseries_id_column],
    #include the cluster model package as the segmenter here!
    model_package_id=cluster_model_package.id
)
segment_task = segment_task_results['completedJobs'][0]

segment_project.analyze_and_model(
    target=target,
    partitioning_method=segment_spec,
    mode=AUTOPILOT_MODE.QUICK,
    worker_count=-1,
    segmentation_task_id=segment_task.id
)