SlideShare a Scribd company logo

Machine Learning Project

Download as PPTX, PDF
A
Abhishek Singh

This document summarizes a machine learning project for Homesite to predict customer quote conversions. The team members are Jack, Harry, and Abhishek. Homesite wants to predict the likelihood of customers purchasing insurance contracts based on their quote. The training data has 261k rows and 298 predictors, while the test data has 200k rows and the same 298 columns. Some key steps included data cleaning, using gradient boosting and random forests, and calculating the AUC (area under the ROC curve) metric to evaluate model performance. The team's model achieved an AUC of 0.95, indicating it does not overfit and has little bias.

1 of 11
Downloaded 245 times
1
2
3
4
Most read
5
6
7
Most read
8
9
10
11
Most read
Machine Learning project Team members: Jack, Harry & Abhishek
Homesite problem: Predicting Quote conversion - Homesite sells Home-insurance to home buyers in United States - Insurance quotes are offered to customers based on several factors What Homesite knows - Customer’s geographical, personal, Financial Information & HomeOwnership details - Quote for every customer Enter What Homesite Doesn’t know: Customer’s likelihood of buying that Insurance contract
Data shared: Training & Test Task: Binary classification Training set: 261k rows, 298 predictors, 1 Binary response Test set: 200k rows, 298 columns Predictors: Customer Activity, Geography, Personal, property & Coverage Response: Customer Conversion What’s good about Homesite data: - 296 variables don’t have NA’s or bad data entry points - Not many Levels in Nominal variables - Plenty of binary variables - Plenty of ordinal variables - No unbalanced variables - No missing values - No Textual columns
Data cleaning steps Removing Constants Removing Identifier rows Synthesizing Date column 1 2 3 Treating NA variables 4 Treating bad levels (-1) 5 Treating false categorical 6 Categorical to dummy 7
Gradient Boosting (Iterative corrections) Learning from past mistakes Could get nearly 0 training error Weighted scoring of multiple trees Hard to tune, as there are too many parameters to adjust Often overfit and hard to decide the stopping point
Random Forests (Majority wins) Handles missing data Handles redundancy easily Reduces variations in results Produces Out of Bag error rate Produces De-correlated trees Random subspace & split Bias sometimes Increases as Trees are shallower
Gradient Boosting + Random Forest Handles missing data Handles redundancy easily Reduces variations in results Produces Out of Bag error rate Produces De-correlated trees Random subspace & split Quite slow & Computational expensive, optimizing these constraints could be an excellent area for research Our Score AUC = .95 Does not overfit Little bias, due to correction Easy to tune
Calculating AUC ID True class Predicted probability 1 1 .8612 2 0 .2134 3 0 .1791 4 0 .1134 5 1 .7898 6 0 .0612 AUC - Randomly decide a threshold - Calculate True Positive Rate (y) & False Positive Rate (x) - Based on (x,y) plot the point - Repeat steps for each value of threshold [0,1] - We now have a curve and we call it ROC - Area under this curve becomes AUC
War for the highest AUC
What we have already employed - Categorical to Continuous conversion - Continuous to Ordinal conversion - Variable bucketing - SVM / Logistic Regression - Random Forest/ Trees - Lasso / Ridge / Elastic Net - Gradient Boosting - Multicollinearity elimination - Outlier treatment - K-Fold Cross validation - Imputation for NA’s - Model tuning - Variable transformation - Most importantly, Your Suggestions What we look forward to use
THANK YOU

More Related Content

PPTX
Machine Learning project presentation
Ramandeep Kaur Bagri
 
PPTX
Support Vector Machines
CloudxLab
 
PPTX
Multiclass classification of imbalanced data
SaurabhWani6
 
PDF
Naive Bayes Classifier Tutorial | Naive Bayes Classifier Example | Naive Baye...
Edureka!
 
PDF
Understanding random forests
Marc Garcia
 
PPTX
Machine Learning Final presentation
AyanaRukasar
 
PPTX
Machine Learning lecture6(regularization)
cairo university
 
PPTX
Machine Learning Unit 4 Semester 3 MSc IT Part 2 Mumbai University
Madhav Mishra
 
Machine Learning project presentation
Ramandeep Kaur Bagri
 
Support Vector Machines
CloudxLab
 
Multiclass classification of imbalanced data
SaurabhWani6
 
Naive Bayes Classifier Tutorial | Naive Bayes Classifier Example | Naive Baye...
Edureka!
 
Understanding random forests
Marc Garcia
 
Machine Learning Final presentation
AyanaRukasar
 
Machine Learning lecture6(regularization)
cairo university
 
Machine Learning Unit 4 Semester 3 MSc IT Part 2 Mumbai University
Madhav Mishra
 

What's hot (20)

PPTX
Machine Learning Using Python
SavitaHanchinal
 
PDF
Support Vector Machines ( SVM )
Mohammad Junaid Khan
 
PPTX
Machine Learning and Real-World Applications
MachinePulse
 
PPTX
Credit card fraud detection using machine learning Algorithms
ankit panigrahy
 
PDF
Feature Engineering in Machine Learning
Knoldus Inc.
 
PPTX
Housing price prediction
Abhimanyu Dwivedi
 
PPTX
Deep neural networks
Si Haem
 
PPTX
Support Vector Machine - How Support Vector Machine works | SVM in Machine Le...
Simplilearn
 
PPTX
Credit card fraud detection using python machine learning
Sandeep Garg
 
PPTX
Machine Learning in Cyber Security
Rishi Kant
 
PPTX
House price prediction
AdityaKumar1505
 
PDF
Credit card fraud detection through machine learning
dataalcott
 
PPT
Back propagation
Nagarajan
 
PDF
Bias and variance trade off
VARUN KUMAR
 
PPTX
Spam email detection using machine learning PPT.pptx
Kunal Kalamkar
 
PPTX
Introduction to Machine Learning
Lior Rokach
 
PPTX
YOLO
geothomas18
 
PPTX
Sms spam-detection
Tanvirul Islam
 
PDF
Linear regression
MartinHogg9
 
PPTX
supervised learning
Amar Tripathi
 
Machine Learning Using Python
SavitaHanchinal
 
Support Vector Machines ( SVM )
Mohammad Junaid Khan
 
Machine Learning and Real-World Applications
MachinePulse
 
Credit card fraud detection using machine learning Algorithms
ankit panigrahy
 
Feature Engineering in Machine Learning
Knoldus Inc.
 
Housing price prediction
Abhimanyu Dwivedi
 
Deep neural networks
Si Haem
 
Support Vector Machine - How Support Vector Machine works | SVM in Machine Le...
Simplilearn
 
Credit card fraud detection using python machine learning
Sandeep Garg
 
Machine Learning in Cyber Security
Rishi Kant
 
House price prediction
AdityaKumar1505
 
Credit card fraud detection through machine learning
dataalcott
 
Back propagation
Nagarajan
 
Bias and variance trade off
VARUN KUMAR
 
Spam email detection using machine learning PPT.pptx
Kunal Kalamkar
 
Introduction to Machine Learning
Lior Rokach
 
YOLO
geothomas18
 
Sms spam-detection
Tanvirul Islam
 
Linear regression
MartinHogg9
 
supervised learning
Amar Tripathi
 
Ad

Similar to Machine Learning Project (20)

PPTX
Presentation - Predicting Online Purchases Using Conversion Prediction Modeli...
Christopher Sneed, MSDS, PMP, CSPO
 
PPTX
Module 3_ Classification.pptx
nikshaikh786
 
PPTX
Learning machine learning with Yellowbrick
Rebecca Bilbro
 
PPTX
House Sale Price Prediction
sriram30691
 
PDF
Cheatsheet machine-learning-tips-and-tricks
Steve Nouri
 
PPTX
Data mining model
Siddharth Negi
 
PDF
The Beginnings of a Search Engine
VirenKhandal
 
PDF
The Beginnings Of A Search Engine
VirenKhandal
 
PPTX
Machine Learning Project - 1994 U.S. Census
Tim Enalls
 
PDF
Evaluating Classifiers' Performance KDD2002
Anna Olecka
 
PDF
Predictive modeling
Prashant Mudgal
 
PPTX
Alex Korbonits, "AUC at what costs?" Seattle DAML June 2016
Seattle DAML meetup
 
PPTX
Churn Modeling For Mobile Telecommunications
Salford Systems
 
PDF
Machine learning project
BabatundeSogunro
 
PPTX
AUC: at what cost(s)?
Alexander Korbonits
 
PPTX
Building and deploying analytics
Collin Bennett
 
PPTX
Competition16
Saurabh Vashist
 
PPTX
NN Classififcation Neural Network NN.pptx
cmpt cmpt
 
PDF
Assessing Model Performance - Beginner's Guide
Megan Verbakel
 
PDF
housepriceprediction-180915174356.pdf
VinayShekarReddy
 
Presentation - Predicting Online Purchases Using Conversion Prediction Modeli...
Christopher Sneed, MSDS, PMP, CSPO
 
Module 3_ Classification.pptx
nikshaikh786
 
Learning machine learning with Yellowbrick
Rebecca Bilbro
 
House Sale Price Prediction
sriram30691
 
Cheatsheet machine-learning-tips-and-tricks
Steve Nouri
 
Data mining model
Siddharth Negi
 
The Beginnings of a Search Engine
VirenKhandal
 
The Beginnings Of A Search Engine
VirenKhandal
 
Machine Learning Project - 1994 U.S. Census
Tim Enalls
 
Evaluating Classifiers' Performance KDD2002
Anna Olecka
 
Predictive modeling
Prashant Mudgal
 
Alex Korbonits, "AUC at what costs?" Seattle DAML June 2016
Seattle DAML meetup
 
Churn Modeling For Mobile Telecommunications
Salford Systems
 
Machine learning project
BabatundeSogunro
 
AUC: at what cost(s)?
Alexander Korbonits
 
Building and deploying analytics
Collin Bennett
 
Competition16
Saurabh Vashist
 
NN Classififcation Neural Network NN.pptx
cmpt cmpt
 
Assessing Model Performance - Beginner's Guide
Megan Verbakel
 
housepriceprediction-180915174356.pdf
VinayShekarReddy
 
Ad

Machine Learning Project

  • 1. Machine Learning project Team members: Jack, Harry & Abhishek
  • 2. Homesite problem: Predicting Quote conversion - Homesite sells Home-insurance to home buyers in United States - Insurance quotes are offered to customers based on several factors What Homesite knows - Customer’s geographical, personal, Financial Information & HomeOwnership details - Quote for every customer Enter What Homesite Doesn’t know: Customer’s likelihood of buying that Insurance contract
  • 3. Data shared: Training & Test Task: Binary classification Training set: 261k rows, 298 predictors, 1 Binary response Test set: 200k rows, 298 columns Predictors: Customer Activity, Geography, Personal, property & Coverage Response: Customer Conversion What’s good about Homesite data: - 296 variables don’t have NA’s or bad data entry points - Not many Levels in Nominal variables - Plenty of binary variables - Plenty of ordinal variables - No unbalanced variables - No missing values - No Textual columns
  • 4. Data cleaning steps Removing Constants Removing Identifier rows Synthesizing Date column 1 2 3 Treating NA variables 4 Treating bad levels (-1) 5 Treating false categorical 6 Categorical to dummy 7
  • 5. Gradient Boosting (Iterative corrections) Learning from past mistakes Could get nearly 0 training error Weighted scoring of multiple trees Hard to tune, as there are too many parameters to adjust Often overfit and hard to decide the stopping point
  • 6. Random Forests (Majority wins) Handles missing data Handles redundancy easily Reduces variations in results Produces Out of Bag error rate Produces De-correlated trees Random subspace & split Bias sometimes Increases as Trees are shallower
  • 7. Gradient Boosting + Random Forest Handles missing data Handles redundancy easily Reduces variations in results Produces Out of Bag error rate Produces De-correlated trees Random subspace & split Quite slow & Computational expensive, optimizing these constraints could be an excellent area for research Our Score AUC = .95 Does not overfit Little bias, due to correction Easy to tune
  • 8. Calculating AUC ID True class Predicted probability 1 1 .8612 2 0 .2134 3 0 .1791 4 0 .1134 5 1 .7898 6 0 .0612 AUC - Randomly decide a threshold - Calculate True Positive Rate (y) & False Positive Rate (x) - Based on (x,y) plot the point - Repeat steps for each value of threshold [0,1] - We now have a curve and we call it ROC - Area under this curve becomes AUC
  • 9. War for the highest AUC
  • 10. What we have already employed - Categorical to Continuous conversion - Continuous to Ordinal conversion - Variable bucketing - SVM / Logistic Regression - Random Forest/ Trees - Lasso / Ridge / Elastic Net - Gradient Boosting - Multicollinearity elimination - Outlier treatment - K-Fold Cross validation - Imputation for NA’s - Model tuning - Variable transformation - Most importantly, Your Suggestions What we look forward to use