Learn Gradient Boosting Algorithm for better predictions (with codes in R)

tavish 26 Jun, 2020
5 min read

Introduction

The accuracy of a predictive model can be boosted in two ways: Either by embracing feature engineering or by applying boosting algorithms straight away. Having participated in lots of data science competition, I’ve noticed that people prefer to work with boosting algorithms as it takes less time and produces similar results.

There are multiple boosting algorithms like Gradient Boosting, XGBoost, AdaBoost, Gentle Boost etc. Every algorithm has its own underlying mathematics and a slight variation is observed while applying them. If you are new to this, Great! You shall be learning all these concepts in a week’s time from now.

In this article, I’ve explained the underlying concepts and complexities of Gradient Boosting Algorithm. In addition, I’ve also shared an example to learn its implementation in R.

Note: This guide is meant for beginners. Hence, if you’ve already mastered this concept, you may skip this article here.

gradient boosting, predictive modeling, machinelearning

 

Quick Explanation

While working with boosting algorithms, you’ll soon come across two frequently occurring buzzwords: Bagging and Boosting. So, how are they different? Here’s a one line explanation:

Bagging: It is an approach where you take random samples of data, build learning algorithms and take simple means to find bagging probabilities.

Boosting: Boosting is similar, however the selection of sample is made more intelligently. We subsequently give more and more weight to hard to classify observations.

Okay! I understand you’ve questions sprouting up like ‘what do you mean by hard? How do I know how much additional weight am I supposed to give to a mis-classified observation.’ I shall answer all your questions in subsequent sections. Keep Calm and proceed.

 

Let’s begin with an easy example

Assume, you are given a previous model M to improve on. Currently you observe that the model has an accuracy of 80% (any metric). How do you go further about it?

One simple way is to build an entirely different model using new set of input variables and trying better ensemble learners. On the contrary, I have a much simpler way to suggest. It goes like this:

Y = M(x) + error

What if I am able to see that error is not a white noise but have same correlation with outcome(Y) value. What if we can develop a model on this error term? Like,

error = G(x) + error2

Probably, you’ll see error rate will improve to a higher number, say 84%. Let’s take another step and regress against error2.

error2 = H(x) + error3

Now we combine all these together :

Y = M(x) + G(x) + H(x) + error3

This probably will have a accuracy of even more than 84%. What if I can find an optimal weights for each of the three learners,

Y = alpha * M(x) + beta * G(x) + gamma * H(x) + error4

If we found good weights, we probably have made even a better model. This is the underlying principle of a boosting learner. When I read the theory for the first time, I had two quick questions:

  1. Do we really see non white noise error in regression/classification equations? If not, how can we even use this algorithm?
  2. Wow, if this is possible, why not get near 100% accuracy?

I’ll answer these questions in this article, however, in a crisp manner. Boosting is generally done on weak learners, which do not have a capacity to leave behind white noise.  Secondly, boosting can lead to overfitting, so we need to stop at the right point.

 

Let’s try to visualize one Classification Problem

Look at the below diagram :

boosting

 

We start with the first box. We see one vertical line which becomes our first week learner. Now in total we have 3/10 mis-classified observations. We now start giving higher weights to 3 plus mis-classified observations. Now, it becomes very important to classify them right. Hence, the vertical line towards right edge. We repeat this process and then combine each of the learner in appropriate weights.

Explaining underlying mathematics

How do we assign weight to observations?

We always start with a uniform distribution assumption. Lets call it as D1 which is 1/n for all n observations.

Step 1 . We assume an alpha(t)

Step 2: Get a weak classifier h(t)

Step 3: Update the population distribution for the next step

formula1where,

formula2

 

Step 4 : Use the new population distribution to again find the next learner

Scared of Step 3 mathematics? Let me break it down for you. Simply look at the argument in exponent. Alpha is kind of learning rate, y is the actual response ( + 1 or -1) and h(x) will be the class predicted by learner. Essentially, if learner is going wrong, the exponent becomes 1*alpha and else -1*alpha. Essentially, the weight will probably increase if the prediction went wrong the last time. So, what’s next?

Step 5 : Iterate step 1 – step 4 until no hypothesis is found which can improve further.

Step 6 : Take a weighted average of the frontier using all the learners used till now. But what are the weights? Weights are simply the alpha values. Alpha is calculated as follows:

formula3

 

Time to Practice – Example

I recently participated in an online hackathon organized by Analytics Vidhya. For making the variable transformation easier, I combined both test and train data in the file complete_data. I started with basic import function and splitted the population in Devlopment, ITV and Scoring.

library(caret)
rm(list=ls())
setwd("C:\\Users\\ts93856\\Desktop\\AV")
library(Metrics)
complete <- read.csv("complete_data.csv", stringsAsFactors = TRUE)
train <- complete[complete$Train == 1,]
score <- complete[complete$Train != 1,]
set.seed(999)
ind <- sample(2, nrow(train), replace=T, prob=c(0.60,0.40))
trainData<-train[ind==1,]
testData <- train[ind==2,]
set.seed(999)
ind1 <- sample(2, nrow(testData), replace=T, prob=c(0.50,0.50))
trainData_ens1<-testData[ind1==1,]
testData_ens1 <- testData[ind1==2,]
table(testData_ens1$Disbursed)[2]/nrow(testData_ens1)
#Response Rate of 9.052%

Here is all you need to do, to build a GBM model.

fitControl <- trainControl(method = "repeatedcv", number = 4, repeats = 4)
trainData$outcome1 <- ifelse(trainData$Disbursed == 1, "Yes","No")
set.seed(33)
gbmFit1 <- train(as.factor(outcome1) ~ ., data = trainData[,-26], method = "gbm", trControl = fitControl,verbose = FALSE)
gbm_dev <- predict(gbmFit1, trainData,type= "prob")[,2] 
gbm_ITV1 <- predict(gbmFit1, trainData_ens1,type= "prob")[,2] 
gbm_ITV2 <- predict(gbmFit1, testData_ens1,type= "prob")[,2]
auc(trainData$Disbursed,gbm_dev)
auc(trainData_ens1$Disbursed,gbm_ITV1)
auc(testData_ens1$Disbursed,gbm_ITV2)

As you will see after running this code, all AUC will come extremely close to 0.84 . I will leave the feature engineering upto you, as the competition is still on. You are welcome to use this code to compete though. GBM is the most widely used algorithm. XGBoost is another faster version of boosting learner which I will cover in any future articles.

 

End Notes

I have seen boosting learners extremely quick and highly efficient. They have never disappointed me to get high initial scores on Kaggle and other platforms. However, it all boils down to how well can you do feature engineering.

Have you used Gradient Boosting before? How did the model perform? Have you used boosting learners in any other capacity. If yes, I would love to hear your experiences in the comments section below.

If you like what you just read & want to continue your analytics learning, subscribe to our emailsfollow us on twitter or like our facebook page

 

tavish 26 Jun, 2020

Frequently Asked Questions

Lorem ipsum dolor sit amet, consectetur adipiscing elit,

Responses From Readers

Clear

Vishwajyoti
Vishwajyoti 11 Sep, 2015

Nicely explained . I have used ada boosting and the performance of random forest and ada boosting were almost same in my case. However, I was wondering if different types of boosting have any advantage over one another.

Krishna Priya
Krishna Priya 11 Sep, 2015

Thank you Tavish for the wonderful post. I have used Ada Boost Algorithm. I was just comparing the performance of Ada Boost , Random Forest , SVM , and Decision Tree Induction with various data sets and various metrics. Ada just dominated than all others and showed an accuracy of approximately 98-99% with all data sets .

venugopal
venugopal 11 Sep, 2015

Yes this is the Essenble methods ERA...

Krishna Priya
Krishna Priya 11 Sep, 2015

Thank you for the wonderful post. I worked with Ada Boost Algorithm quite a while back. I was just comparing the performance of Ada Boost, SVM, Random Forest and Decision Tree with various data sets and metrics . Ada Boost performed far better than others and showed an accuracy of 98-99 % in all cases .

vishwa
vishwa 11 Sep, 2015

Hi, while running the below code , i am getting the error "Error: cannot allocate vector of size 56.4 Gb" code: gbmFit1 <- train(as.factor(outcome1) ~ ., data = trainData[,-26], method = "gbm", trControl = fitControl,verbose = FALSE) I tried to fix this issue with "memory.limit(size=)" and gc() , but of no use.I am running on 64bit machine.

Ram
Ram 14 Sep, 2015

Vishwa, Were you able to solve the problem? If the data size is beyond your machine's capacity, you can try fitting the model on a random subset of the data. numrows=nrow(trainData) ind=sample(1:numrows, numrows/10) ## select only 10% of the data trainData_s=trainData[ind,] Then fit a model on trainData_s and see if the machine can handle this data.

Kuber Jain
Kuber Jain 22 Oct, 2015

Hi, How can I include weight variable as a part of above gradient boosting R code and get the parameter estimates of predictor variables? Please help.

Pankaj
Pankaj 22 Apr, 2016

hi, wonderful article; can you please share the data you used along the exercise above. Many Thanks,

Sharon
Sharon 13 Sep, 2016

Thank you Tavish. This is a great article. Quick question - epsilon refers to the error rate, right? (in step 6). That would indicate higher weights to trees with lower error rate. Just want to confirm. Also, to work with larger amounts of data, you can adjust your train fraction and bag fraction to choose a smaller set of rows for each tree, but set a higher number of trees so it converges well. Also, # variables makes a huge difference in memory requirements. If you have a few hundred variables to work with, you may want to split into multiple groups, build separate GBM models and then combine the best variables from each iteration. One thing the examples may be missing is taking care of overfitting. Adjusting final node size is important.

Yasho Joshi
Yasho Joshi 24 Oct, 2016

Can you please attach a link to the dataset that used to explain this method?

Rich
Rich 08 Nov, 2016

Echoing previous requests: can you please post the data set???

Krishna
Krishna 09 May, 2017

Is this really Gradient Boosting? I don't see any loss function in this post, It just seems to be AdaBoost. Could you please clarify?

SF
SF 18 Jul, 2017

????????????????????? Let’s begin with an easy example : Gradient boosting Explaining underlying mathematics : Adaboost Time to practice : Gradient boosting

Mike
Mike 30 Aug, 2017

An introduction to gradient boosting, with no mentioning of gradient in the actual explanation... something is missing...

salman bootwala
salman bootwala 12 Oct, 2017

Can you please provide link to the dataset

Nadjib
Nadjib 28 Dec, 2017

Hi, thank you fr this amazing explanation, i used Gradient Boosting and AdaBoost and gradient boosting gave me amazing results, i didn't understand the step 6 of the algorithme, and how the formula of calculating alpha works ?.

Siva
Siva 30 Jan, 2018

I want Ada algorithm code and can u tell me which algorithm is best?