Comparison To Synthpop¶

SeqTrees is inspired by the same broad sequential synthesis idea used by Synthpop, but it makes a different design choice at the data boundary.

Synthpop¶

Synthpop¹ is an R package for creating synthetic microdata. Its documentation describes variables as categorical or continuous and states that variables are synthesised one by one using sequential modelling. Replacements are drawn from conditional distributions fitted to the original data using parametric models or classification and regression trees.

Synthpop is designed to work directly with R data frames and their data types. It includes many synthesis methods, utility/disclosure tools, and data-type aware behavior. For example, low-cardinality numeric variables can be changed to factors depending on settings such as minnumlevels.

SeqTrees¶

SeqTrees is deliberately narrower. It accepts pandas DataFrames and does no internal preprocessing by default. Input must contain no missing values and every variable must be one of:

• continuous: represented as float;
• discrete: represented as int category codes.

SeqTrees does not accept one-hot encoding as the intended workflow. Binary variables are simply discrete integer-coded variables with values such as 0 and 1.

Users should declare the modelling role of every variable:

model = SequentialTreeSynthesizer(
    continuous_columns=["age", "bmi"],
    discrete_columns=["sex_code", "income_bin", "risk_code"],
)

If both lists are omitted, SeqTrees can infer all-float columns as continuous and all-integer columns as discrete. With preprocessor="ifcfill", SeqTrees uses IFCTransformer.column_types_ so original categorical columns do not generate unseen labels while original integer columns remain numeric.

Practical Difference¶

Topic	Synthpop	SeqTrees
Language	R	Python
Input data	R data frames with type-aware behavior	pandas DataFrames
Categorical handling	Can use factor/categorical-aware methods	Integer codes, or opt-in IFCFill label encoding
Continuous handling	Parametric and tree-based numeric methods	Empirical or interpolated leaf sampling
Variable roles	Largely driven by data type and method settings	Explicitcontinuous_columns and discrete_columns
One-hot input	Not the core design point	Not intended; use one integer-coded categorical column
API style	R functions such assyn()	Estimator-stylefit() / sample()
Backends	R modelling ecosystem	Native, scikit-learn, LightGBM

SeqTrees is therefore closer to a Python estimator with a focused DataFrame preparation layer, while Synthpop is a broader R toolkit with more built-in data-type aware synthetic data machinery.