# LightGBM (Light Gradient Boosting Machine)

**LightGBM** is a gradient boosting framework based on decision trees to increases the efficiency of the model and reduces memory usage.

It uses two novel techniques: **Gradient-based One Side Sampling** and **Exclusive Feature Bundling (EFB)** which fulfills the limitations of histogram-based algorithm that is primarily used in all GBDT (Gradient Boosting Decision Tree) frameworks. The two techniques of GOSS and EFB described below form the characteristics of LightGBM Algorithm. They comprise together to make the model work efficiently and provide it a cutting edge over other GBDT frameworks **Gradient-based One Side Sampling Technique for LightGBM: **

Different data instances have varied roles in the computation of information gain. The instances with larger gradients(i.e., under-trained instances) will contribute more to the information gain. GOSS keeps those instances with large gradients (e.g., larger than a predefined threshold, or among the top percentiles), and only randomly drop those instances with small gradients to retain the accuracy of information gain estimation. This treatment can lead to a more accurate gain estimation than uniformly random sampling, with the same target sampling rate, especially when the value of information gain has a large range. **Algorithm for GOSS: **

Input: I: training data, d: iterationsInput: a: sampling ratio of large gradient dataInput: b: sampling ratio of small gradient dataInput: loss: loss function, L: weak learner models ? {}, fact ? (1-a)/b topN ? a × len(I), randN ? b × len(I)fori = 1toddopreds ? models.predict(I) g ? loss(I, preds), w ? {1, 1, ...} sorted ? GetSortedIndices(abs(g)) topSet ? sorted[1:topN] randSet ? RandomPick(sorted[topN:len(I)], randN) usedSet ? topSet + randSet w[randSet] × = fact . Assign weight f act to the small gradient data. newModel ? L(I[usedSet], g[usedSet], w[usedSet]) models.append(newModel)

**Mathematical Analysis for GOSS Technique (Calculation of Variance Gain at splitting feature j)**

For a training set with n instances {x_{1}, · · ·, x_{n}}, where each x_{i} is a vector with dimension *s* in space X^{s}. In each iteration of gradient boosting, the negative gradients of the loss function with respect to the output of the model are denoted as {g_{1}, · · ·, g_{n}}. In this GOSS method, the training instances are ranked according to their absolute values of their gradients in the descending order. Then, the top-a × 100% instances with the larger gradients are kept and we get an instance subset A. Then, for the remaining set A^{c }consisting (1- a) × 100% instances with smaller gradients., we further randomly sample a subset B with size b × |A^{c}|. Finally, we split the instances according to the estimated variance gain at vector *V*_{j }(d) over the subset A ? B.

(1)

where *A _{l} = {x_{i }? A : x_{ij} ? d}, A_{r} = {x_{i }? A : x_{ij} > d}, B_{l} = {x_{i} ? B : x_{ij} ? d}, B_{r }= {x_{i }? B : x_{ij }> d},* and the coefficient

*(1-a)/b*is used to normalize the sum of the gradients over B back to the size of A

^{c}.

**Exclusive Feature Bundling Technique for LightGBM:**

High-dimensional data are usually very sparse which provides us a possibility of designing a nearly lossless approach to reduce the number of features. Specifically, in a sparse feature space, many features are mutually exclusive, i.e., they never take nonzero values simultaneously. The exclusive features can be safely bundled into a single feature (called an Exclusive Feature Bundle). Hence, the complexity of histogram building changes from

*O(#data × #feature)*to

*O(#data × #bundle)*, while

*#bundle<<#feature*. Hence, the speed for training framework is improved without hurting accuracy.

**Algorithm for Exclusive Feature Bundling Technique:**

Input: numData: number of dataInput: F: One bundle of exclusive features binRanges ? {0}, totalBin ? 0forfinFdototalBin += f.numBin binRanges.append(totalBin) newBin ? new Bin(numData)fori = 1tonumDatadonewBin[i] ? 0forj = 1tolen(F)doifF[j].bin[i] != 0thennewBin[i] ? F[j].bin[i] + binRanges[j]Output: newBin, binRanges

**Architecture : **

LightGBM splits the tree leaf-wise as opposed to other boosting algorithms that grow tree level-wise. It chooses the leaf with maximum delta loss to grow. Since the leaf is fixed, the leaf-wise algorithm has lower loss compared to the level-wise algorithm. Leaf-wise tree growth might increase the complexity of the model and may lead to overfitting in small datasets.

Below is a diagrammatic representation of Leaf-Wise Tree Growth:

**Code: Python Implementation of LightGBM Model: **

The data set used for this example is Breast Cancer Prediction. Click on this to get data set : Link to Data set.

## python

`# installing LightGBM (Required in Jupyter Notebook and` `# few other compilers once)` `pip install lightgbm` `# Importing Required Library` `import` `pandas as pd` `import` `lightgbm as lgb` `# Similarly LGBMRegressor can also be imported for a regression model.` `from` `lightgbm ` `import` `LGBMClassifier` `# Reading the train and test dataset` `data ` `=` `pd.read_csv("cancer_prediction.csv)` `# Removing Columns not Required` `data ` `=` `data.drop(columns ` `=` `[` `'Unnamed: 32'` `], axis ` `=` `1` `)` `data ` `=` `data.drop(columns ` `=` `[` `'id'` `], axis ` `=` `1` `)` `# Skipping Data Exploration` `# Dummification of Diagnosis Column (1-Benign, 0-Malignant Cancer)` `data[` `'diagnosis'` `]` `=` `pd.get_dummies(data[` `'diagnosis'` `])` `# Splitting Dataset in two parts` `train ` `=` `data[` `0` `:` `400` `]` `test ` `=` `data[` `400` `:` `568` `]` `# Separating the independent and target variable on both data set` `x_train ` `=` `train.drop(columns ` `=` `[` `'diagnosis'` `], axis ` `=` `1` `)` `y_train ` `=` `train_data[` `'diagnosis'` `]` `x_test ` `=` `test_data.drop(columns ` `=` `[` `'diagnosis'` `], axis ` `=` `1` `)` `y_test ` `=` `test_data[` `'diagnosis'` `]` `# Creating an object for model and fitting it on training data set` `model ` `=` `LGBMClassifier()` `model.fit(x_train, y_train)` `# Predicting the Target variable` `pred ` `=` `model.predict(x_test)` `print` `(pred)` `accuracy ` `=` `model.score(x_test, y_test)` `print` `(accuracy)` |

OutputPrediction array : [0 1 1 1 1 1 1 1 0 1 1 1 1 0 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 0 1 0 0 1 0 1 1 1 1 1 0 1 1 0 1 0 1 1 0 1 0 1 1 1 1 1 1 1 1 0 0 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 0 1 0 1 0 0 1 1 1 1 1 0 0 1 0 1 0 1 1 1 1 1 0 1 1 0 1 0 1 0 0 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 0 1 0 0 1 1 1 1 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0] Accuracy Score : 0.9702380952380952

**Parameter Tuning**

Few important parameters and their usage is listed below :

**max_depth :**It sets a limit on the depth of tree. The default value is 20. It is effective in controlling over fitting.**categorical_feature :**It specifies the categorical feature used for training model.**bagging_fraction :**It specifies the fraction of data to be considered for each iteration.**num_iterations :**It specifies the number of iterations to be performed. The default value is 100.**num_leaves :**It specifies the number of leaves in a tree. It should be smaller than the square of*max_depth*.**max_bin :**It specifies the maximum number of bins to bucket the feature values.**min_data_in_bin :**It specifies minimum amount of data in one bin.**task :**It specifies the task we wish to perform which is either train or prediction. The default entry is*train*. Another possible value for this parameter is*prediction.***feature_fraction**: It specifies the fraction of features to be considered in each iteration. The default value is one.