Skip to content

v0.23.0

Latest

Choose a tag to compare

@OnlyDeniko OnlyDeniko released this 12 Nov 12:08
· 22 commits to main since this release

RePlay 0.23.0 Release notes

Release date: 12.11.2026 · Version: 0.23.0 · Type: Minor

Highlights

This release adds PR-Entmax, a sparse prior-regularized loss designed to
mitigate popularity bias in recommender systems while preserving
recommendation relevance. PR-Entmax requires no architectural changes and
uses item-frequency information only during training.

New features

  • Added PREntmax and PREntmaxSampled loss functions. PR-Entmax incorporates
    an item-frequency prior directly into the sparse Entmax mapping, allowing
    contextual relevance and item frequency to jointly determine which items
    receive non-zero probabilities and gradients. The strength of the prior
    correction is controlled by gamma, while temperature controls logit
    scaling. Both loss functions accept a one-dimensional item_frequency tensor aligned
    with encoded item IDs. If item_frequency is not provided, they operate as
    standard Entmax-1.5 losses.

    Added a complete
    SasRec example
    demonstrating how to calculate the item-frequency prior, configure
    PREntmax, train SasRec with CE and PR-Entmax losses, and compare relevance,
    catalog coverage, and novelty metrics.

    The method is described in
    Mitigating Popularity Bias in Sequential Recommendations with Sparse Prior-Regularized Objective.