Decision tree classifier
Paste a labelled table, pick the class column, and get a readable decision tree: numbered if/then rules, feature importance, accuracy, and predictions for new rows.
About this tool
Fit a single decision tree to a pasted, labelled table and read back exactly how it decides. The tool parses CSV, TSV, semicolon, pipe, or whitespace-delimited rows, treats one column as the class label, and greedily grows a tree by Gini impurity decrease (CART), Shannon information gain (ID3), or the gain ratio (C4.5). Feature columns can be numeric — split on a midpoint threshold — or categorical, split either one-vs-rest or with one branch per value. Everything runs locally in WebAssembly; the table never leaves the browser.
The report includes numbered IF … THEN rules, a text tree, normalised feature importance,
training accuracy with a confusion matrix, an optional hold-out check, and predictions for rows you
paste in. Output can also be JSON, flat CSV, or a Graphviz DOT digraph you can render as a diagram.
Worked example:
color,size,ripe
red,small,yes
red,large,yes
green,small,no
green,large,no
With the class column ripe and the default Gini criterion, the tree splits once and prints:
Tree:
├─ color = green → no [n=2, 100.0%]
└─ color != green → yes [n=2, 100.0%]
Rules:
1. IF color = green THEN ripe = no [n=2, 100.0%]
2. IF color != green THEN ripe = yes [n=2, 100.0%]
size never gets used, so its importance is 0 and color takes the full 1.0 — the tree tells you
which column actually carries the signal.
Limits and edge cases: tables are capped at 20,000 rows and 100 columns, tree depth at 20, and
pasted prediction rows at 1,000. A feature or class column with more than 200 distinct values is
rejected as an id-like column. Rows with a blank, NA, null, or ? value in the class column or
any selected feature are dropped and counted in the report. Accuracy measured on the same rows the
tree was fitted to is optimistic — set a hold-out test split, or shrink max_depth, to see how much
of the fit is real.
FAQ
Which criterion should I pick?
gini is the CART default and is fast and stable — start here. entropy is classic ID3
information gain and usually produces a very similar tree. gain_ratio is the C4.5 correction:
it divides the information gain by how finely a feature fragments the data, which stops a
many-valued column (like a date or a product code) from winning every split just because it splits
the rows into lots of tiny pure groups.
How do I stop the tree from overfitting?
Use the pre-pruning knobs. max_depth caps how long a rule can get, min_samples_split stops
small nodes from splitting at all, min_samples_leaf refuses splits that would strand a handful of
rows in a branch, and min_gain throws away splits whose score is below a threshold. Combine any of
them with a test_split hold-out to check the effect on unseen rows. Cost-complexity pruning
(ccp_alpha) and C4.5 confidence-factor post-pruning are not implemented.
Can I mix text and numeric columns?
Yes. Each feature column is classified automatically: if every value in it parses as a number it is
treated as numeric and split with a <= threshold at the midpoint between two observed values;
otherwise it is categorical and split on its values. Categorical splits are one-vs-rest by default
(color = green versus color != green); switch Categorical splits to multiway for one branch
per distinct value, the way ID3 and C4.5 present them.
How do I classify new rows?
Paste them into Rows to classify, one per line. Three layouts are accepted: the full table
layout (the class column is ignored), just the feature columns in order, or a header row naming the
columns followed by the values. Each prediction reports the class, the leaf's purity as a
confidence, and which numbered rule fired. An unseen category takes the != branch; a missing value
follows whichever branch held the most training rows.
What do the importance numbers mean?
They are impurity-decrease shares, the same idea as a standard feature_importances_ vector: every
split adds the impurity it removed, weighted by the share of rows reaching that node, and the totals
are normalised to sum to 1. A column that is never split on scores 0. They describe this tree on
this data — they are not causal claims, and a correlated column can absorb another one's credit.
Is the result reproducible?
Yes. Tree fitting is completely deterministic — no random feature sampling, no tie-breaking by
chance — so the same table and options always produce the same rules. The only randomness is the
shuffle behind the hold-out test split, and that is driven by seed.
Developer & Automation Access
Run it from the terminal
Same engine as this page, headless — via the gizza CLI:
gizza tool decision-tree-classifier "color,size,ripe
red,small,yes
red,large,yes
green,small,no
green,large,no"New to the CLI? Get gizza →
Open it by URL
Pre-fill and auto-run this tool with query parameters — the names match the API/CLI:
https://gizza.ai/tools/decision-tree-classifier/?data=color%2Csize%2Cripe%0Ared%2Csmall%2Cyes%0Ared%2Clarge%2Cyes%0Agreen%2Csmall%2Cno%0Agreen%2Clarge%2Cno&target=last&features=color%2Csize&criterion=gini&splits=binary&max_depth=5&min_samples_split=2&min_samples_leaf=1&min_gain=0&class_weight=none&test_split=0&seed=42&predict=color%2Csize%0Agreen%2Csmall&header=auto&decimals=4&format=textMachine-readable descriptor: tool.json — title + parameters JSON Schema for agents.