Structure Learning Algorithms

CausalIQ Discovery aims to provide a wide range of different structure learning algorithm classes: exact or approximate score-based, sampling, constraint-based, hybrid, continuous optimisation and neural network-based for example. They will initially include algorithms supporting time-independent observational data, but will be extended to include interventional and time-series data.

The algorithms are either implemented internally or as wrappers to open source packages such as tetrad-jar, py-tetrad, causal-learn, gCastle, tigrmite, cdt, pcalg, BiDAG etc. In either case, the focus will be on adding features that allow knowledge to guide or constrain the structure learning, and to produce trace diagnostics which explain the structure learning process (these are the USPs of the CausalIQ Discovery package).

Supported Algorithms

Algorithm	Class	Package	Description
fges	score	Tetrad/causal-cmd	Fast Greedy Equivalence Search
gs	constraint	bnlearn	Grow-shrink local discovery
h2pc	hybrid	bnlearn	Parents & Children and hill-climbing
hc	score	CausalIQ	Hill-climbing
hc-stable	score	CausalIQ	Stable hill-climbing
iiamb	constraint	bnlearn	Interleaved IAMB local discovery
mmhc	hybrid	bnlearn	Markov Blankets and hill-climbing
pc-stable	constraint	bnlearn	Stable PC (Peters & Clark)
tabu	score	CausalIQ	Hill-climbing with tabu list
tabu-stable	score	CausalIQ	Stable hill-climbing with tabu list