BUS2004 – Predictive Analytics Assignment 2

Introduction

This report presents the process and outcomes of developing a predictive classification model to determine whether a customer will churn using the dataset. The dataset comprises several customer-related attributes, as described in the accompanying Metadata worksheet.

The analysis was performed using the Orange Data Mining platform, which supports visual, modular workflows for machine learning and data analysis. The workflow involved data exploration, cleaning, preprocessing, model building, evaluation, and performance improvement.

Descriptive Statistics

I performed the descriptive statistics of the data using the following orange data workflow:

Data information:

There are total 13 columns with 3150 data rows. Data contains 3 categorical variables and 9 numerical variables. The target variable has two classes. There are 123 missing values in the features that is 0.3% of the data.

Feature Statistics

The above screenshot of statistic summary provides different features of the dataset. The dataset has unique records of 3150 customers.

Call Failure

There are 7.63 calls on average that are failed. The median value is 0 and mode is 6. Minimum value is 0 and maximum value is 36. This shows that most customers experience fewer call failures, but some customers report high failures (up to 36). The distribution is right-skewed, meaning many have low failures, but a few have very high failures. Customers with higher call failures may be more likely to churn.

Subscription Length

The Subscription Length has a mean of 32.54 months, with a median of 35 and values ranging from 3 to 47 months. The distribution suggests that many customers have relatively long subscription histories, which may indicate loyalty. However, there are also customers with much shorter subscription lengths, potentially representing new users who may be at higher risk of churn if they do not find value in the service. Longer subscription lengths typically correlate with customer stability and lower churn probability.

Charge Amount

The Charge Amount feature shows a mean of 0.94, a median of 0, and a mode of 0, with a maximum of 10. The majority of customers have little to no additional charges, as reflected by the skewed distribution toward zero. A smaller group of customers experiences higher charges, which could lead to dissatisfaction if perceived as unfair or excessive. This feature may therefore help identify whether additional charges are linked to an increased likelihood of churn.

Seconds of Use

The Seconds of Use variable has a mean of 4472.46 seconds, with a median of 2990 and a wide range from 0 to 17,090. This indicates that customer usage varies greatly, with some users being heavy callers while others use the service minimally. The right-skewed distribution suggests that most customers fall within lower ranges, but heavy users form a significant minority. Higher usage levels may indicate dependency on the service, which could lower churn risk, unless service quality issues prompt dissatisfaction.

Frequency of Use

The Frequency of Use feature has a mean of 68.88, a median of 53, and values ranging from 0 to 255. About 1% of data in this feature is missing. The distribution indicates that most customers make a relatively low to moderate number of calls, while a smaller group has very high call frequencies. Regular and frequent use of the service could signal loyalty, but such users may also be highly sensitive to network quality, making them more vulnerable to churn if service performance declines.

Frequency of SMS

The Frequency of SMS variable has a mean of 73.17, a median of 21, and a maximum of 522. The data is highly skewed, with most customers sending very few SMS messages, while a small group sends SMS at very high frequencies. This variation may reflect different customer profiles, such as light users and heavy communicators. The extent to which SMS usage influences churn will depend on whether heavy SMS users are satisfied with the service or find cheaper alternatives elsewhere.

Distinct Called Numbers

The Distinct Called Numbers feature shows a mean of 23.51, a median of 21, and values ranging from 0 to 97. This suggests that, on average, customers call around 20 unique contacts. A wider range of distinct contacts may indicate higher engagement with the service and possibly lower churn risk. Conversely, customers with fewer unique contacts might be less attached to the service, making them more prone to switching providers.