temporian 0.9.0

Temporian is a Python package for feature engineering of temporal data, focusing on preventing common modeling errors and providing a simple and powerful API, a first-class iterative development experience, and efficient and well-tested implementations of common and not-so-common temporal data preprocessing functions.

Temporian is a library for safe, simple and efficient preprocessing and feature engineering of temporal data in Python. Temporian supports multivariate time-series, multivariate time-sequences, event logs, and cross-source event streams.

Temporian is to temporal data what Pandas is to tabular data.

Key features

• Supports most types of temporal data 📈: Handles both uniformly sampled and non-uniformly sampled data, both single-variate and multivariate data, both flat and multi-index data, and both mono-source and multi-source non-synchronized events.

• Optimized for Temporal data 🔥: Temporian's core computation is implemented in C++ and optimized for temporal data. Temporian can be more than 1,000x faster than off-the-shelf data processing libraries when operating on temporal data.

• Easy to integrate into an existing ML ecosystem: Temporian does not perform any ML model training - instead it integrates seamlessly with any ML library, such as PyTorch, Scikit-Learn, Jax, TensorFlow, XGBoost, or Yggdrasil Decision Forests.

• Prevents unwanted future leakage 😰: Unless explicitly specified with tp.leak, feature computation cannot depend on future data, thereby preventing unwanted, hard-to-debug, and potentially costly future leakage.

Quickstart

Installation

Install Temporian from PyPI with pip:

pip install temporian -U

Temporian is currently available for Linux and MacOS (ARM and Intel). Windows support is under development.

Minimal example

Consider sale records that contain contain the timestamp, store, and revenue of individual sales.

$ cat sales.csv
timestamp,store,revenue
2023-12-04 21:21:05,STORE_31,5071
2023-11-08 17:14:38,STORE_4,1571
2023-11-29 21:44:46,STORE_49,6101
2023-12-20 18:17:14,STORE_18,4499
2023-12-15 10:55:09,STORE_2,6666
...

Our goal is to compute the sum of revenue for each store at 11 pm every weekday (excluding weekends).

First, we load the data and list the workdays.

import temporian as tp

# Load sale transactions
sales = tp.from_csv("sales.csv")

# Index sales per store
sales_per_store = sales.add_index("store")

# List work days
days = sales_per_store.tick_calendar(hour=22)
work_days = (days.calendar_day_of_week() <= 5).filter()

work_days.plot(max_num_plots=1)

Then, we sum the daily revenue for each workday and each store.

# Aggregate revenue per store and per work day
daily_revenue = sales_per_store["revenue"].moving_sum(tp.duration.days(1), sampling=work_days).rename("daily_revenue")

# Plot the results
daily_revenue.plot(max_num_plots=3)

Finally, we can export the result as a Pandas DataFrame for further processing or for consumption by other libraries.

tp.to_pandas(daily_revenue)

Check the Getting Started tutorial to find out more!

Next steps

New users should refer to the Getting Started guide, which provides a quick overview of the key concepts and operations of Temporian.

After that, visit the User Guide for a deep dive into the major concepts, operators, conventions, and practices of Temporian. For a hands-on learning experience, work through the Tutorials or refer to the API reference.

If you need help, have a question, want to contribute, or just want to be a part of the Temporian community, we encourage you to join our Discord server! 🤝🏼

Documentation

The documentation 📚 is available at temporian.readthedocs.io. The Getting Started guide is the best way to start.

Contributing

Contributions to Temporian are welcome! Check out the Contributing guide to get started.

Credits

Temporian is developed in collaboration between Google and Tryolabs.