Overview

LightGBM is Microsoft's gradient boosting framework. 2–10x faster than XGBoost on large datasets thanks to leaf-wise tree growth, histogram-based splitting, and native categorical feature support (no encoding required). Industry standard for high-performance tabular ML.

Binary Classification

uses lightgbm;

{ load from CSV — label in column 0 }
var train := DatasetFromCSV(ReadFile('train.csv'), 0);
var valid := DatasetFromCSV(ReadFile('valid.csv'), 0);

{ quick start }
var model := TrainDefault(train, lgbmBinary);

{ predict probabilities }
var preds := Predict(model, valid);
var i: Integer;
for i := 0 to Length(preds) - 1 do
  WriteLn('P(+1) = ', preds[i]:0:4);

FreeBooster(model);
FreeDataset(train);
FreeDataset(valid);

Custom Parameters

uses lightgbm;

var params := DefaultParams(lgbmBinary);
params.NumRounds       := 300;
params.NumLeaves       := 63;     { more leaves = more complex model }
params.LearningRate    := 0.05;
params.FeatureFraction := 0.8;
params.BaggingFraction := 0.8;
params.BaggingFreq     := 5;
params.EarlyStopRounds := 20;     { stop if no improvement for 20 rounds }
params.EvalMetric      := 'auc';
params.Verbose         := -1;     { silent }

var result: TLGBMTrainResult;
var model := TrainWithHistory(train, valid, params, result);
WriteLn('Best iteration: ', result.BestIteration);
WriteLn('AUC: ', result.EvalMetric:0:4);

Categorical Features

uses lightgbm;

{ LightGBM handles categoricals natively — no one-hot encoding needed }
var train := DatasetFromCSV(ReadFile('customers.csv'), 0);

{ mark columns 1 (country), 3 (plan), 5 (device) as categorical }
SetCategoricalFeatures(train, '1,3,5');
SetFeatureNames(train, 'churned,country,age,plan,income,device,tenure');

var model := TrainDefault(train, lgbmBinary);
WriteLn('Trees: ', GetNumTrees(model));

Regression

uses lightgbm;

var params := DefaultParams(lgbmRegression);
params.NumRounds    := 200;
params.NumLeaves    := 31;
params.LearningRate := 0.1;
params.EvalMetric   := 'rmse';

var train := DatasetFromFile('housing.csv', '');
var model := Train(train, params, 0);

{ predict a single sample }
var features: array of Double;
SetLength(features, 10);
features[0] := 3.5;   { rooms }
features[1] := 120.0; { sqm }
{ ... }

var price := PredictOne(model, features);
WriteLn('Estimated price: $', price:0:0);

Feature Importance

uses lightgbm;

var model := Train(train, params, 0);

{ "split" = number of times feature used in splits }
{ "gain"  = total gain from splits using this feature }
var imp := GetFeatureImportance(model, 'gain');
WriteLn(imp);
{ tenure     0.312
  income     0.241
  age        0.187
  ... }

WriteLn('Features: ', GetNumFeatures(model));
WriteLn('Trees:    ', GetNumTrees(model));

Save, Load & Cross-Validation

uses lightgbm;

{ cross-validation }
var cv := CrossValidate(train, params, 5);
WriteLn('5-fold CV: ', cv:0:4);

{ save model }
SaveModel(model, '/models/churn.txt');

{ load and predict later }
var loaded := LoadModel('/models/churn.txt');
var preds  := Predict(loaded, testDataset);
FreeBooster(loaded);

Partial Prediction (Early Trees)

uses lightgbm;

{ predict using only first 50 trees — useful for speed/accuracy tradeoff }
var fast := PredictNumTrees(model, testDataset, 50);

{ raw scores before sigmoid/softmax }
var raw := PredictRaw(model, testDataset);