timeseriesflattener

What timeseriesflattener is

timeseriesflattener is a Python package for turning irregular time-series observations into structured features for machine-learning predictors. It produces a dataframe with one row for each prediction time and columns containing constructed features, grouping observations by an ID column. Feature specifications define prediction times, lookbehind windows for predictors, and lookahead windows for outcomes. The documented aggregation choices include count, earliest, latest, maximum, mean, minimum, slope, sum, unique count, and variance. Its API describes temporal predictors, outcomes, boolean outcomes, static features, and time-delta features. A fallback value can be supplied when a relevant look window has no data. The package accepts pandas or Polars dataframes for prediction-time and static inputs. The documentation also demonstrates creating flattened predictors from pre-embedded text held in a dataframe with entity IDs, timestamps, and embedding columns. An introductory tutorial says n_workers can parallelize processing across multiple cores. Downloadable Jupyter notebook tutorials are available to run locally. The package is especially aimed at complicated, irregular time series used while training simple models. Installation is through pip, using the package name timeseriesflattener. It is deployed self-hosted, with bug reports and feature requests handled through GitHub Issues and general usage discussion through GitHub Discussions.

Who it suits

It suits Python users preparing irregular or complicated time-series data for machine-learning models, especially when working with pandas or Polars dataframes. Teams that need local notebooks, configurable time windows, and self-hosted processing can use its documented API.

Getting started

1. Install the package with pip install timeseriesflattener
2. Prepare prediction-time or static data in pandas or Polars dataframes
3. Define IDs, prediction times, and lookbehind or lookahead windows
4. Select feature types and aggregation functions
5. Run a downloadable Jupyter notebook tutorial locally

Verdict

Pick timeseriesflattener when you need configurable feature construction from irregular time series before training machine-learning models. Consider another option if you want a hosted service or a non-Python workflow rather than a self-hosted package installed with pip.