Design Goals and Principles

LensKit is intended to be a flexible toolkit for a wide range of teaching and research purposes. The LensKit paper [Eks20] documents several of these principles in detail; as we have gained more experience, particularly with the changes leading to LensKit 2025.1 (see Migrating from LensKit 0.x).

These principles capture the broad goals that guide LensKit development; they are sometimes in tension with each other or other engineering concerns, and are aspirational rather than fully-realized, but serious divergences are usually bugs or work-in-progress.

Make best practices the default (or at least easy)

It should be easy to do research using currently-understood best practices.

Use the pieces you want

LensKit is a toolkit of individual pieces that do not have hard dependencies on each other. If you have your own recommendation pipelines and want to use LensKit’s metrics to evaluate recommendations you have saved in Parquet or CSV files, it works (and we encourage this use — standardizing metric implementations is important for research comparability, and using LensKit’s metrics don’t lock you in to the other parts of LensKit). If you want to use the pipeline abstraction for components that work on complete different data sructures, that works too (we do this in POPROX). If you want to use some of the LensKit components but not others, that works too.

The goal of LensKit is to support research using whatever computational toolkits you want.

Implement the pieces you need

LensKit components are designed to be sufficiently configurable and re-composable to enable you to experiment with new recommendation ideas by only implementing the new pieces that are essential to your idea.

Bridge to other tools

It should be easy to implement bridges to use scoring models from other libraries, or new scoring models using arbitrary computational libraries, in LensKit experiments (and, where appropriate, LensKit pipelines).

Use standard data structures

This principle has evolved from its original version [Eks20], when we used Pandas data frames as the interchange format between components, but LensKit data is still based on standard arrays (supporting both NumPy and PyTorch), sparse matrices, etc.

Support flexible data and paradigms

LensKit intends to support a wide range of data and recommendation paradigms, including collaborative filters, content-based recommendation, session-based and session-aware recommendation, etc.

Note

This principle is more aspirational than the others at present. For historical reasons, most of the models are collaborative filters, and they have the best support, but the 2025.1 (in progress) changes and planned imminent work are partially designed to improve the for recommendation beyond user-item-matrix collaborative filtering.