Introduction

This is a (hopefully) exhaustive list of ongoing/future work for HDBSCAN. These have all been discussed and are considered wanted, but some still require thorough investigation (especially heuristic evaluations).

Priority List

The higher priority items appear earlier in this list.

• ENH Add np.float32 data support for HDBSCAN #26888
  • ENH Add float32 implementations for BallTree and KDTree  #25914
  • Finalize {KD, Ball}Tree API to avoid writing custom dispatcher
• Support np.nan in Cython implementation for sparse matrices
  • MNT Improve robustness of sparse test in HDBSCAN #26889
• Reintroduce Boruvka algorithm (removed in b7736ef)
• Implement PWD backend for weighted argkmin in medoid calculation
• Investigate PWD backend for mst_from_* functions in _linkage.pyx
• Investigate PWD backend for _reachability.pyx
• Benchmark KD vs Ball Tree efficiency
• Add consistent threading semantics to enable prange, e.g. in _reachability.pyx
• Improve partition strategy in _reachability.pyx (cf. CLN Cleaned cluster/_hdbscan/_reachability.pyx #24701 (comment))
• Add support for np.inf values when metric=='precomputed' and X is sparse.