Business Introduction

Abstract

Cross-selling is a sales strategy employed to encourage customers to increase their spending by acquiring a product that complements their current purchase. This approach finds utility in the insurance industry, where it serves as a means for companies to introduce new products to their existing customer base. Leveraging machine learning models can streamline the labor-intensive process of sifting through customer records, resulting in significant time and cost savings. However, the implementation of machine learning models poses its own set of challenges, some of which we aim to address in this project. Recognizing the pivotal role of data in business, we prioritize customer privacy, adhering to the principles of the Ethical Machine Learning framework. Our team conducts data analysis and model development following the CRISP-DM methodology.

Background

Triks Insurance Inc., a prominent Life Insurance Agent in the town of XYZ, boasts an extensive client base of 381,109 households within the region. Presently, they are in negotiations with a prominent Auto Insurance provider, GOOL Auto Insurance, with plans to introduce their insurance product. GOOL Auto Insurance has requested Triks Insurance to furnish a report indicating the proportion of their client base likely to express interest in Auto insurance.

Rather than embarking on the arduous task of manually sifting through their customer database, Triks Insurance Inc. has opted to engage the SK team to create a machine learning system. This system’s purpose is not only to predict potential Auto Insurance candidates once but continuously over time. By comprehending the customer base and leveraging demographic information about a prospect, this machine learning system will assess whether a prospect falls into the category of a “good” or “bad” candidate.

By implementing this solution, Triks Insurance’s sales team will be better prepared for their cross-selling endeavors. They can hone their focus on the most promising leads, thus maximizing the efficiency of their marketing budget while concentrating on the most lucrative prospects.

The proposed solution serves several valuable purposes:

*It provides GOOL Auto with insights into the potential business that can be generated from Triks Insurance’s existing client base.

*It equips the staff at Triks Insurance with a tool for client prioritization, facilitating focused and effective marketing campaigns.

*It enables Triks Insurance to project future revenues, enhancing their financial planning.

*The system offers real-time notifications for potential cross-selling opportunities.

The client was initially unaware that the data they possess holds the answers to their questions. Some of their concerns and inquiries include:

-What information or resources do you require from us to deliver this solution?

-Can the system accurately classify customers?

-How will the system handle cases where a customer who should have qualified for GOOL Insurance is not selected?

Objective

The aim of this project is to create one or more models to generate an initial report for GOOL Insurance on behalf of Triks Insurance Ltd. Additionally, the project seeks to develop a predictive tool that can determine whether a customer, based on specific characteristics, qualifies as a potential candidate for their upcoming cross-selling campaign.

Business Understanding:

To understand and address business problems:

-Determine the key variables to predict (good/bad candidates). -Define relevant metrics related to these variables. -Grasp project objectives and requirements. -Create specific questions, such as what distinguishes a good from a bad prospect. -Assess current sales practices for identifying prospects. -Establish a success metric for the project. -Identify data sources with answers to these questions. -Select data that accurately measures the target and relevant features. -Consider if the existing system requires additional data for the project. -Evaluate the need for external data sources or system updates.

Then, translate this knowledge into a data mining problem and create a preliminary plan to achieve the project goals.

Data Understanding:

Having understood the business statement, it’s clear that Triks Insurance faced a choice. They could have promoted Auto Insurance to all their customers, but that wouldn’t have been the most efficient use of their marketing budget. It’s more cost-effective to focus on customers likely to respond to the GOOL Auto campaign. This targeted approach saves money and doesn’t bother customers uninterested in the new product.

To achieve this, we can leverage historical data from past campaigns to build a model for predicting which customers are promising prospects. We start by collecting relevant data, addressing data quality issues, gaining initial insights from the data, and identifying interesting trends or patterns within it.

The data for this project was downloaded from Kaggle, weblink: https://www.kaggle.com/anmolkumar/health-insurance-cross-sell-prediction. For privacy and security, the customer names have been masked with an id.The data is in csv format.

Data Description :

Variable Name Variable Description
id Unique ID for the customer
Gender Gender of the customer
Age Age of the customer
Driving_License 0 : Customer does not have DL
1 : Customer already has DL
Region_Code Unique code for the region of the customer
Previously_Insured 1 : Customer already has Vehicle Insurance.
0 : Customer doesn’t have Vehicle Insurance.
Vehicle_Age Age of the Vehicle
Vehicle_Damage 1 : Customer’s vehicle damaged in the past.
0 : Customer”s vehicle no damaged in the past.
Annual_Premium The premium insurec paid in the year
Policy_Sales_Channel Anonymized Code for outreach channels for sales.
Vintage Number of Days, Customer has been associated with.
the company
Response 1 : Customer is interested
0 : Customer is not interested

Loading Data

test <- read.csv("c://DATA//test.csv")
train <- read.csv('c://DATA//train.csv')
submission <- read.csv('c://DATA//sample_submission.csv')

###Determine the dimension of our dataset:

## [1] 381109     12

View the contents and structure of our dataset:

View the main structure of the raw data

## 'data.frame':    381109 obs. of  12 variables:
##  $ id                  : int  1 2 3 4 5 6 7 8 9 10 ...
##  $ Gender              : chr  "Male" "Male" "Male" "Male" ...
##  $ Age                 : int  44 76 47 21 29 24 23 56 24 32 ...
##  $ Driving_License     : int  1 1 1 1 1 1 1 1 1 1 ...
##  $ Region_Code         : num  28 3 28 11 41 33 11 28 3 6 ...
##  $ Previously_Insured  : int  0 0 0 1 1 0 0 0 1 1 ...
##  $ Vehicle_Age         : chr  "> 2 Years" "1-2 Year" "> 2 Years" "< 1 Year" ...
##  $ Vehicle_Damage      : chr  "Yes" "No" "Yes" "No" ...
##  $ Annual_Premium      : num  40454 33536 38294 28619 27496 ...
##  $ Policy_Sales_Channel: num  26 26 26 152 152 160 152 26 152 152 ...
##  $ Vintage             : int  217 183 27 203 39 176 249 72 28 80 ...
##  $ Response            : int  1 0 1 0 0 0 0 1 0 0 ...

Helper Function

Helper functions are used in conjunction with the main verbs to make specific tasks and calculations a bit easier.

Exploratory Data Analysis (EDA)

We can create a profile of the ideal customer for our insurance company:

The ideal customer doesn’t currently have active insurance. They have a history of vehicle accidents or damage. Their vehicle is relatively new, between 1 to 2 years old. Their age range of 30 to 55. Their annual premium between 30,000 and 40,000. They are more inclined to use channels like channel 25 or channel 125, rather than channel 155. ### Customer Response

What affect customer response?

Response by previous insurance

Response by past vehicle damage

Response by gender

Response by vehicle age

Response by age

Response by Annual Premium

What is affecting vehicle damage?

Age by Vehicle Damage

Vehicle damage by vehicle age

vehicle damage by gender

Annual Premium

annual premium by vehicle damage

## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.

## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.

Age distribution by Vehicle Age

Policy sales channel

Correlation

Examine the relationships between independent numerical variables by creating a correlation matrix to assess the correlation coefficients between these variables. Before constructing the correlation matrix, exclude the ‘id’ and ‘Response’ columns(correlation can be perform on numeric only).

Loading head of training data

## [1] "data.frame"

##Create The XGBoost model XBBoost is a boosting technique in machine learning, known for its ability to generate highly accurate predictive models.

####How deal with am imbalanced dataset? There are numerous approaches, and while many of them are valid, this time we decided to take a different path.

How we proceed:

Our approach is as follows:

We’ll fine-tune certain XGBoost parameters using grid search. We’ll apply the XGBoost Classifier to the imbalanced dataset (the original dataset), while setting the scale_pos_weight : sum(negative)/sum(positive). We’ll assess the XGBoost Classifier’s performance on a balanced dataset that matches the length of the training dataset. Finally, we’ll submit the results.

Since all customers in the submission dataset are uninterested, we will focus solely on evaluating the model’s metrics.

We will tune some XGB parameters with gridsearch.

Oversampling for gridsearch

XGB model

####scale_pos_weight= 334399/46710 = 7.15

Response #
0 334399
1 46710

Launch XGBoost

Test XGB on a balanced dataset

As a further test for our model, we will rebalance the original dataset and evaluate the model performance on a balanced dataset

Rebalance dataset

## [1] "Length of train_3 dataset(Balanced dataset): 381109"
Response %
0 0.4997101
1 0.5002899

Launch Xgboost

## [1] "XGB is starting..."
## [1] "XGB Train:"
## [1]  train-auc:0.798917 
## [201]    train-auc:0.883364 
## [401]    train-auc:0.899102 
## [601]    train-auc:0.910498 
## [801]    train-auc:0.921392 
## [1000]   train-auc:0.930757 
## [1] "XGB Cross Validation:"
## [1]  train-auc:0.841687+0.000625 test-auc:0.841301+0.001094 
## [201]    train-auc:0.908711+0.000917 test-auc:0.885836+0.001570 
## [401]    train-auc:0.937619+0.001331 test-auc:0.904483+0.001312 
## [601]    train-auc:0.956170+0.000755 test-auc:0.917808+0.001107 
## [801]    train-auc:0.967871+0.000623 test-auc:0.927173+0.001033 
## [1000]   train-auc:0.976454+0.000681 test-auc:0.934779+0.000742
## Warning: The `<scale>` argument of `guides()` cannot be `FALSE`. Use "none" instead as
## of ggplot2 3.3.4.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.
## Setting direction: controls < cases

## $`Features Importance:`

## 
## $`Confusion Matrix`
## TableGrob (2 x 2) "arrange": 3 grobs
##   z     cells    name                grob
## 1 1 (2-2,1-1) arrange      gtable[layout]
## 2 2 (2-2,2-2) arrange  gtable[rowhead-fg]
## 3 3 (1-1,1-2) arrange text[GRID.text.963]
## 
## $`Roc Curve`
## 
## Call:
## roc.default(response = label_train, predictor = cv$pred, levels = c(0,     1))
## 
## Data: cv$pred in 190444 controls (label_train 0) < 190665 cases (label_train 1).
## Area under the curve: 0.9348

Submission (Checking Results)

Submit XGB predictions

Model Deployment

Once we’ve selected a model, we proceed to deploy it with a data pipeline into a production or production-like environment for final user acceptance. This prepares the model for seamless integration with the client’s existing applications. We successfully deployed our model and presented it to Triks Insurance Ltd.

From our initial understanding of the business, we identified several ways in which the client aims to benefit from this project:

-Providing GOOL Auto with insights into the potential business Triks Insurance can generate from their existing client base. -Equipping Triks Insurance staff with a tool to prioritize clients for targeted marketing campaigns. -Enabling Triks Insurance to project future revenues. -Offering real-time notifications for potential cross-selling opportunities.

Pipelines can be developed for each of these requirements, although they fall outside the scope of this project. Predictions can be generated in real-time or on a batch basis. We have deployed the model to predict cross-sells based on user-provided customer input, including gender, age, region code, policy sales channel, vehicle age, and vehicle damage. Additionally, another deployment facilitates data file uploads, allowing predictions for entire datasets.

Conclusion

To achieve better accuracy we performed a hyper parameter tuning of the other parameters such (learning rate), max_depth (maximum depth of tree), sub sample values, etc by using a hyper grid and then achieved a better tuned mode.