New LensKit-Auto brings TPE, DeepCAVE and new LensKit Core features

Disclaimer: I am also head of the Intelligent Systems Group, which developed LensKit-Auto

A new step for LensKit-Auto

The Intelligent Systems Group at the University of Siegen has announced a new update to LensKit-Auto, its open-source AutoRecSys library built on the Python LensKit recommender-system framework. The release adds compatibility with the newer LensKit architecture, support for Python 3.12 and 3.13, automated visualizations with DeepCAVE, a third optimization strategy based on TPE, improved time management for model search, and updated documentation.

LensKit-Auto is designed for a task that many recommender-system researchers know well: selecting a suitable algorithm, tuning its hyperparameters, and evaluating the resulting model. The library automates this process around LensKit models. In its standard use case, a developer can call get_best_recommender_model() and let LensKit-Auto search for a strong configuration for the given dataset.

The new LensKit foundation

A central part of the update is compatibility with the new LensKit architecture. LensKit 2025.1 introduced major API and design changes, including a flexible pipeline abstraction for wiring together components such as candidate selectors, personalized item scorers, and rankers. This matters for LensKit-Auto because automated model selection becomes more valuable when it can work with the current structure of the underlying recommender framework.

LensKit itself has a long history in the recommender-systems community. It was first released in 2010 as a Java framework and later re-envisioned as a Python package for recommender-systems research, education, and early-stage development. The Python version was presented by Michael D. Ekstrand at CIKM 2020. Today, LensKit describes itself as a set of Python tools for training, running, and evaluating recommender algorithms in a flexible way for research and education.

The project has been shaped by Michael D. Ekstrand and collaborators, with institutional roots in Boise State University and Drexel University, and earlier connections to the GroupLens research tradition. The current LensKit documentation credits Drexel University, Boise State University, and collaborators, and the LensKit paper explicitly frames the project as infrastructure for reproducible recommender-systems experiments.

More transparent optimization with DeepCAVE

One of the most useful additions is automated plot generation with DeepCAVE. DeepCAVE is a visualization and analysis tool for AutoML, with a focus on hyperparameter optimization. It provides interactive views that help users inspect optimization runs, understand performance patterns, and debug automated searches.

For LensKit-Auto, this means that the search process becomes easier to study. A researcher can look at how configurations performed, how the optimizer moved through the search space, and where promising regions appeared. This is valuable for recommender-systems experiments because model selection is part of the scientific result. Seeing the search is often as useful as seeing the final score.

TPE joins the optimization toolbox

The update also adds the Tree-structured Parzen Estimator, or TPE, as a third optimization technique. LensKit-Auto already supported random search and Bayesian optimization. TPE now gives users another established strategy for navigating hyperparameter spaces.

TPE is widely used in modern hyperparameter optimization. In Optuna’s implementation, the sampler models parameter values associated with better objective values separately from the remaining values and proposes new values by comparing these models. In practice, this gives LensKit-Auto a stronger and more flexible optimization toolbox for recommender-system experiments.

Better time management for model search

The release also improves time management for model search. This is a practical feature with real research value. Recommender-system evaluation can involve several algorithms, data splits, metrics, and parameter spaces. Time control helps users run searches that fit their available compute budget while still benefiting from automation.

This also fits the spirit of AutoRecSys. Automation should make recommender-system experimentation easier to run, easier to repeat, and easier to integrate into regular development workflows. A tool that manages search time well is more useful in classrooms, labs, and applied projects.

LensKit-Auto in the AutoRecSys landscape

AutoRecSys brings AutoML ideas into recommender systems. The general idea is to automate design decisions that otherwise require manual work: algorithm selection, hyperparameter optimization, component selection, pipeline construction, and model ensembling. Surveys on AutoML for recommender systems describe this area as a response to the growing complexity of recommender-system design, especially when many modeling choices interact with one another.

LensKit-Auto occupies a clear place in this landscape. It is an open-source AutoRecSys toolkit based on LensKit. It performs automated algorithm selection, hyperparameter optimization, and post-hoc model ensembling for LensKit algorithms. The project page at ISG Siegen describes it as a tool for warm-starting recommender-system development, producing a working recommender system with one line of code, and building RecSys pipelines with less engineering effort.

LensKit-Auto was introduced to the RecSys community in 2023 as a demo paper by Tobias Vente, Michael D. Ekstrand, and Joeran Beel. This gives the project a nice bridge between two parts of the community: the LensKit ecosystem around Michael Ekstrand and collaborators, and the AutoRecSys research line at the Intelligent Systems Group at the University of Siegen.

A useful update for RecSys experiments

The new LensKit-Auto release strengthens the connection between automated recommender-system development and the modern LensKit framework. The update brings the library into the current LensKit architecture, adds Python 3.12 and 3.13 support, improves documentation, gives users DeepCAVE visualizations, and expands the optimizer choices with TPE.

For the recommender-systems community, this is useful progress. LensKit has long served researchers and teachers who need clear tools for experimentation. LensKit-Auto builds on that tradition and adds automation where researchers often spend much of their time: choosing algorithms, tuning parameters, and managing search runs. The update makes that workflow more modern, more visible, and more convenient.