Blog#
Acceleration
Andrey 0.9.0, our biggest speed breakthrough: BOSS up to 8.1x faster, exact search 5.2x, and FCI 3.1x, with exactly the same graphs.
Earlier posts
Nonlinear dependence
In Andrey 0.8.0, PC, FCI, GFCI, and CD-NOD test independence with KCI, RCoT, CMIknn, and GPDC to detect nonlinear effects that partial correlation misses.
Permutation invariance
In Andrey 0.7.0, reordering your columns no longer changes the graph from PC, FCI, GFCI, CD-NOD, hill climbing, tabu search, BOSS, GRaSP, or GIN.
Mixed and categorical data
Andrey 0.6.0 learns causal graphs from categorical data and from tables that mix categories with numbers, in PC, GES, hill climbing, and seven more methods.
Temporal causal discovery
Temporal causal discovery in Andrey 0.5.0: six methods that find causes in data measured over time, from PCMCI to LongitudinalLiNGAM, are now supported.
Pairwise direction
Pairwise direction in Andrey 0.4.0: ANM and PNL test which of two variables causes the other, and both are now supported.
Score-based search
Score-based search in Andrey 0.3.0: Tabu search and XGES are new and supported, as is hill climbing, which now returns the same graphs faster.
Hidden common causes
Andrey's methods for data with hidden common causes, with RCD and BottomUpParceLiNGAM new and supported, and small fits several times faster.
Why Andrey is fast
The four ideas behind Andrey's speed, each shown as a short animation.
Introducing Andrey
Introducing Andrey, a very fast causal discovery package for Python.