CHAID Segmentation

A self-contained Python package for population segmentation using CHAID decision trees. You give it a KPI/target and a set of predictors; it auto-bins the continuous predictors, grows a CHAID tree, and hands back interpretable segments with an event rate (or mean), population, population share and lift — plus a static tree chart where each node is a population and each branch is a choice.

It is built for the workflow: "input a KPI, generate a tree, read off the high/low segments of the customer base." For example, for a 90+DPD target:

age < 25 AND region = Phnom Penh AND bank = ABA → 60% 90+DPD rate, 10% of population, 2.5x lift

The CHAID tree engine is bundled in this repository — there is no external CHAID dependency to install. It is based on the Rambatino/CHAID project (Apache 2.0); see Credits & License.

Features

• Automatic binning of continuous predictors — supervised (target-based), equal-width, equal-frequency (quantile) or manual cut points, chosen per variable.
• Binary and continuous KPIs — node rate is the event rate P(target = positive) for a binary target, or the mean for a continuous target.
• Interpretable segments — every terminal node becomes a readable rule with rate, population, population share and lift.
• Static visualisation (matplotlib + seaborn) — node = population, branch = choice, colour = rate.
• Predict / score new data by re-applying the fitted bins and rules.
• Load straight from CSV or Parquet with one call.

Installation

Requires Python 3.9+. Install from this repository:

pip install .                       # core (numpy, pandas, scipy, matplotlib, seaborn, ...)
pip install '.[segmenter-target]'   # + optbinning, for method="target" (supervised binning)
pip install '.[parquet]'            # + pyarrow, for ChaidSegmenter.from_parquet

matplotlib and seaborn are core dependencies (binning + plotting). optbinning and pyarrow are only needed for target-based binning and Parquet loading respectively, and are imported lazily with a clear error if missing.

How to use

import pandas as pd
from chaid_segmenter import ChaidSegmenter

df = pd.read_csv("loan_book.csv")     # columns: age, income, tenure, score, region, bank, dpd90

seg = ChaidSegmenter(
    target="dpd90",
    positive_class=1,                                    # binary event-rate target
    predictors={
        "age":    {"method": "target", "max_bins": 4},      # supervised (optbinning)
        "income": {"method": "equal_width", "bins": 4},     # fixed-interval bins
        "tenure": {"method": "equal_frequency", "bins": 4}, # quantile bins
        "score":  {"method": "manual", "edges": [550, 650, 750]},
        "region": {"method": "nominal"},                    # categorical, used as-is
        "bank":   {"method": "nominal"},
    },
    max_depth=3,
    min_child_node_size=0.02,         # int count, or a fraction of the dataset
    alpha_merge=0.05,
)
seg.fit(df)

seg.summary()                         # tidy DataFrame, highest rate first
seg.segments()                        # list[Segment]
seg.predict(df_new)                   # assign rows to terminal segments
seg.plot("tree.png")                  # static matplotlib/seaborn chart
seg.tree_frame()                      # raw tree, one row per node -- DataFrame
seg.tree_dict()                       # raw tree, nested dict -- for custom plots/graphs

Load and fit in a single call:

seg = ChaidSegmenter.from_csv("loan_book.csv", "dpd90", predictors, positive_class=1)
seg = ChaidSegmenter.from_parquet("loan_book.parquet", "dpd90", predictors, positive_class=1)

A runnable, self-contained demo lives at examples/dpd_segmentation.py.

You don't have to spell out every predictor

predictors accepts three forms — pick whichever is least effort:

# 1. Full control: a spec (or method string) per column
predictors={"age": {"method": "target", "max_bins": 4}, "region": "nominal"}

# 2. Just the column names — the method is inferred from each column's dtype
#    (numeric -> default_numeric_method, non-numeric -> nominal)
predictors=["age", "income", "region", "bank"]

# 3. Omit it entirely — auto-select every column except the target/weight
ChaidSegmenter(target="dpd90", positive_class=1).fit(df)

In full-auto mode, constant columns and high-cardinality text columns (IDs, names, free text — anything with more than max_nominal_cardinality distinct values) are skipped automatically. You can always mix inference with overrides — e.g. {"age": "auto", "score": {"method": "manual", "edges": [550, 650]}} — and inspect what was chosen via seg.resolved_predictors after fit. Inferred numeric columns use default_numeric_method (default "target", falling back gracefully if you prefer "equal_frequency"/"equal_width").

Expected output

seg.summary()

A pandas.DataFrame, one row per terminal segment, highest rate first:

 node_id                                                           description population population_pct  rate  lift
       2                                         age < 24.9859 AND score < 550        530           6.6% 48.3% 1.99x
       8                         bank = ABA AND age < 24.9859 AND score >= 550        188           2.4% 41.5% 1.71x
       4                         bank = ABA AND age >= 24.9859 AND score < 550      1,043          13.0% 37.9% 1.56x
       9             bank in {ACLEDA, Wing} AND age < 24.9859 AND score >= 550        410           5.1% 30.0% 1.24x
       5             bank in {ACLEDA, Wing} AND age >= 24.9859 AND score < 550      2,107          26.3% 25.1% 1.03x
      11             bank = ABA AND age in [24.9859, 66.5531) AND score >= 550      1,167          14.6% 22.4% 0.92x
      13                                       age >= 66.5531 AND score >= 550        282           3.5% 20.9% 0.86x
      12 bank in {ACLEDA, Wing} AND age in [24.9859, 66.5531) AND score >= 550      2,273          28.4% 10.7% 0.44x

seg.segments()

Each Segment is a small object you can read off directly:

top = seg.segments()[0]
top.description     # 'age < 24.9859 AND score < 550'
top.rate            # 0.4830...   (48.3% 90+DPD rate)
top.population      # 530.0
top.population_pct  # 0.06625     (6.6% of the book)
top.lift            # 1.99        (vs the 24.3% overall rate)
top.node_id         # 2
top.rules           # [{'variable': 'age', 'label': 'age < 24.9859', 'data': [...]},
                    #  {'variable': 'score', 'label': 'score < 550', 'data': [...]}]

seg.plot("tree.png")

Each node shows its population (count + % of total) and rate; each branch is labelled with the choice that leads into it; node colour encodes the rate:

seg.tree_frame() / seg.tree_dict()

Everything plot() draws, as data instead of a picture — one row/dict per tree node (internal and terminal, not just the final segments), with the same split variable, incoming-branch condition, population, rate, lift and layout position (x, y) used to render the chart above. Use these to build your own plot (Plotly, Graphviz, D3, NetworkX, ...):

>>> seg.tree_frame().head()
   node_id  parent_id  depth  is_terminal split_variable branch_variable    branch_label  population  population_pct   rate  lift   x    y  children
0        0       None      0        False            age           None            None      8000.0          1.0000 0.2430  1.00 3.5  0.0  [1, 2]
1        1          0      1         True           None            age       age < 25      1200.0          0.1500 0.4830  1.99 1.0 -1.0        []
2        2          0      1        False         score            age      age >= 25      6800.0          0.8500 0.2010  0.83 4.5 -1.0  [3, 4]

>>> seg.tree_dict()
{'node_id': 0, 'parent_id': None, 'split_variable': 'age', 'branch_label': None,
 'rate': 0.243, ..., 'children': [{'node_id': 1, 'branch_label': 'age < 25', ...}, ...]}

tree_dict() is plain nested dicts/lists (json.dumps-able), so it drops straight into JSON-based/hierarchical viz libraries.

seg.predict(df, with_rate=True)

Assigns every row to its terminal segment (and, optionally, that segment's rate):

>>> seg.predict(df_new, with_rate=True).head()
   node_id      rate
0       11  0.223650
1        4  0.378715
2       12  0.106907
3        5  0.250593
4        4  0.378715

Rows that match no segment (e.g. an unseen category at predict time) come back as <NA>.

Binning methods

Each predictor's method selects how it is turned into branches:

method	Spec keys	Description
target	max_bins	Supervised optimal binning via optbinning — monotonic event rate. Works on numeric columns (rate-ordered ranges) and on high-cardinality categoricals (groups categories into rate tiers). Needs the segmenter-target extra.
equal_width	bins	Fixed-width intervals across the value range.
equal_frequency	bins	Quantile bins of roughly equal population.
manual	edges	User-supplied interior cut points.
nominal	—	Categorical predictor, used as-is (no binning).

Continuous predictors are converted to ordinal bin codes, so only contiguous bins ever merge and every branch renders as a clean range (age < 25, in [25, 40), >= 40). Missing values become their own missing branch.

A spec may be written as a bare method string when it takes no options, e.g. "region": "nominal".

High-cardinality IDs (member / institution codes)

A categorical with hundreds of distinct values — a member id, institution code, merchant id — can't be used as plain nominal (CHAID would try to merge hundreds of categories: slow and unreadable). Use target to group it by the event rate into a few risk tiers:

predictors={"MEMBER_ID": "target", "PRODUCT_TYPE": "nominal", ...}

Segments then read like MEMBER_ID in {003, 014, 019, 024, 045, 048, +11} → 40.7% — a concrete high-risk member group. This also happens automatically when you simply list such a column (its cardinality exceeds max_nominal_cardinality); in full-auto mode (predictors omitted) high-cardinality columns are dropped instead, since the tool can't tell a meaningful id from a row identifier like ACCOUNT_ID.

Numeric-looking ids. Whether a column is binned as a number or grouped as a category is decided from its dtype. An id like 001..200 read from a CSV becomes int64, so it would be binned into meaningless ranges (MEMBER_ID < 100). Either cast it to text — df["MEMBER_ID"] = df["MEMBER_ID"].astype(str) (or read with pd.read_csv(..., dtype={"MEMBER_ID": str})) — or force grouping with {"method": "target", "categorical": True}.

Targets

• Binary — pass positive_class (the event value, e.g. 1). Node rate is P(target == positive_class) and lift is rate / overall_rate.
• Continuous — omit positive_class. Node rate is the mean of the target and lift is mean / overall_mean.

API reference

ChaidSegmenter(...)

Parameter	Default	Description
target	—	Name of the KPI/target column.
predictors	None	A {column: spec} dict, a list of column names (methods inferred from dtype), or None for full auto-select. See Binning methods and above.
positive_class	None	Event value for a binary target; None ⇒ continuous target.
default_numeric_method	"target"	Binning method for auto-inferred numeric predictors (target / equal_frequency / equal_width).
default_bins	5	Bin count for auto-inferred numeric predictors.
max_nominal_cardinality	20	In full-auto mode, non-numeric columns with more distinct values are skipped.
max_depth	3	Maximum tree depth.
min_child_node_size	30	Minimum observations per child. Values in (0, 1) are treated as fractions of the dataset.
min_parent_node_size	None	Minimum observations to split a node; defaults to min_child_node_size. Fractions supported.
alpha_merge	0.05	Significance threshold for merging predictor categories.
split_threshold	0	Surrogate-split threshold (passed through to the tree engine).
max_splits	None	Maximum number of children per split.
weight	None	Optional weight column; populations and rates use weighted sums.

Methods

• fit(df) — fit the binners and grow the tree from a pandas.DataFrame.
• segments(sort_by_rate=True) — list of Segment objects.
• summary() — segments as a tidy DataFrame.
• segment_rates — {node_id: rate} for the terminal nodes.
• predict(df, with_rate=False) — terminal node_id per row (optionally with rate).
• plot(path=None, **kwargs) — render the tree; returns (fig, layout) and writes to path if given. Accepts figsize, cmap, dpi, font-size overrides.
• tree_frame() — every tree node (not just terminals) as a tidy DataFrame, for building a custom plot/graph of the tree.
• tree_dict() — the tree as a nested dict (root → children → ...), for hierarchical/JSON-based plotting libraries.
• ChaidSegmenter.from_csv(path, target, predictors, *, read_csv_kwargs=None, **kwargs) and from_parquet(...) — load a file, construct and fit in one call.

Segment

node_id, description, rate, population, population_pct, lift, and a structured rules list of {variable, label, data}.

How it works

ChaidSegmenter fits a Binner per continuous predictor and feeds the resulting integer bin codes (as ordinal columns) plus the nominal predictors into a bundled CHAID tree engine, which splits on the predictor most strongly associated with the target (chi-squared for categorical targets, Bartlett's/Levene's test for continuous targets). Because the bins enter as contiguous ordinal codes, merged groups always describe a single, readable range. Terminal nodes are then translated back into rate/population/lift segments.

Low-level tree engine

The underlying CHAID tree implementation is bundled and importable as CHAID (from CHAID import Tree) for advanced use — building a tree by hand, exporting classification_rules(), treelib conversion, etc. The segmentation API above is the recommended entry point for the KPI-segmentation workflow.

Testing

pip install -e '.[segmenter-target,parquet,test]'
pytest tests/

Credits & License

This project is built on top of CHAID by Mark Ramotowski, Richard Fitzgerald and contributors. The CHAID/ package in this repository is that upstream implementation, bundled unmodified as the underlying tree engine. The chaid_segmenter/ package — automatic binning, KPI segmentation and the matplotlib/seaborn visualisation — is an original addition.

Distributed under the Apache License 2.0. See LICENSE.txt for the full license text and NOTICE for attribution.