LoginSign Up
← Back to all sparks
A

anticlust

INFRA · APIS
Velocity0.0

Subset Partitioning via Anticlustering

github.com/m-py/anticlust ↗

Anticlustering keeps absorbing constraints — must-link, cannot-link, blocks, missing data

anticlusteringexperimental-designconstraint-solvingoptimizationstimulus-selection
◆Current state
anticlust partitions a set of items into groups that are as similar to each other as possible — assembling matched stimulus sets, balanced experimental conditions, comparable teaching groups. The current surface is one function, anticlustering(), with a growing list of things it will honour: must-link and cannot-link constraints, categorical variables as factors, NAs, blocking by a categorical level, and a choice among exact solvers and heuristics.
◆Where it's heading
Two years of releases have gone almost entirely into constraint handling and input tolerance rather than new objectives. The methods list grew by absorbing outside work — the three-phase search of Yang et al. contributed by an external author, a 2PML heuristic for must-link problems, a Gurobi backend for the exact formulations. Meanwhile categories_to_binary() has been rewritten twice, which is where the recent bugs have come from.
◆Prediction
The constraint types now interact combinatorially — blocks, must-link, cannot-link, missing values and each objective — and the last two releases were both regressions in the encoding layer underneath them. Consolidating that layer is the more likely next move than another solver.

◆Recent moves

  1. 4mo ago

    One-hot encoding regression from 0.8.13 reverted

    A correction to the internals that turn categorical variables into the binary columns the objectives operate on — broken across two prior releases. The encoding layer has now been the source of both of the last two fixes.

    View source ↗
  2. 8mo ago

    blocks argument applies anticlustering level by level

    A blocks argument runs anticlustering sequentially within the levels of a categorical variable, the standard blocked-design pattern, and categories_to_binary() stops erroring on single-level factors. Continues the pattern of teaching one function to accept more shapes of input.

    View source ↗
  3. 9mo ago

    Missing values, factor columns, and the three-phase search

    The broadest input-tolerance release in the window: anticlustering() handles NAs and takes factor columns directly, must-link constraints work with the kplus and variance objectives, and method='3phase' exposes the Yang et al. algorithm contributed by an outside author.

    View source ↗
  4. 1y ago

    Edge cases in the previous release's cannot-link handling

    A same-day follow-up patching corners of the vectorised cannot-link support shipped 35 minutes earlier.

    View source ↗
  5. 1y ago

    Vector cannot-link constraints skip the solver entirely

    cannot-link constraints can be given as a vector, and when they come from a grouping variable the problem is solved without invoking a solver at all. A constraint that was expensive to express becomes the cheap path.

    View source ↗
  6. 1y ago

    Uncaught test error flagged by CRAN additional checks

    A CRAN-driven patch to the test suite with no change to package behaviour.

    View source ↗