Add Multi-Horizon Time Series Classification Support

[eetulauri]

Darts has established itself as a premier time series forecasting library. Adding multi-horizon time series classification support would solidify its position and significantly benefit researchers and practitioners alike. This feature aligns with previous discussions and interest, as seen in issue #1473.

Use Case Example:

Traffic congestion prediction: Classifying whether traffic congestion is likely to occur over multiple upcoming time intervals (e.g., next 15 minutes, next hour) based on historical traffic patterns.

Desired Functionality

• Input: Time series data and desired classification horizon.
• Output: Class labels for each time series segment across the specified horizon.

Potential Approaches

• Incorporate classifiers like LightGBMClassifier, enabling them to produce class label predictions for multiple future time steps within the Darts framework.

[madtoinou]

Hi @eetulauri,

This is on Darts' roadmap but we would like to add conformal prediction before tackling the classification features as it would require some work for the deep learning models.

As described in the linked issue, you can easily use "Classifier" variant of supported model by creating a new class and overwritting the "model" attribute.

You can also use classifier models from sklearn thanks to the RegressionModel class;

import numpy as np
from darts import TimeSeries
from darts.models import RegressionModel
from sklearn.ensemble import RandomForestClassifier

ts = TimeSeries.from_values(np.random.randint(0,10,100))
model = RegressionModel(
    lags=4,
    model=RandomForestClassifier()
)
model.fit(ts)
model.predict(10)