Customer churn is the loss of customers. As a group, we had been tasked with analyzing the customer churn rate for the eCommerce clothing website we had created. Our goal is to identify the factors that contribute to customer churn and suggest measures to prevent it. To achieve this, we have analyzed the historical data generated using Mockaroo [10] in a typical e-commerce scenario, which includes customer behavior, demographics, purchase history, and feedback. The code and the detailed project here is only about calculating and predicting customer churn.

Churn Prediction using Machine Learning
The churning of customers is the phenomenon in which customers leave the business. Customers are individuals who have already undertaken the process of buying products or engaging in business contracts. To predict customers’ next move, machine learning was utilized because it is important to take action to avoid further withdrawal of customers. The huge amounts of data collected in a real situation cannot be handled by humans efficiently thus, it made sense to employ these techniques to make the data suitable for providing necessary value to our business. Machine learning is the process by which machines learn from data rather than humans having to explicitly write lines of code [8]. The data used and the algorithms employed are described further to discuss the results of our analysis and the changes we would make to our business to improve customer satisfaction and reduce customer churning.

Dataset
It would be helpful to know customers’ opinions on the store to predict if they would leave our store. As a reason, it was set up so that each customer had an opportunity to provide feedback on their experience of using the store and answer useful questions that would enable us to improve the store and their experience so that we minimize customer churning or customer deleting their accounts in our website. A google form was made as a touchpoint to collect data from the customers after each of their purchases.

It has questions related to the customer’s rating 1-5 for the quality of the products, delivery experience, how appropriate they think the prices of the products in the store are, the website’s design and look, customer service provided by our company, and lastly, how likely they are to recommend our company to others. These ratings will help the company understand the customers’ opinions on the offerings of a business (their products, website, and services) and these in turn could explain the drop or increase in the number of customers.

All the attributes correspond to the questions in the form and have been given self-explanatory names. For example, last_purchase_before is the number of days since the last purchase, time_spent_last is the duration of the customer’s last session on the website, and account_creation_before means the number of days from the time of creating the account. The churn value could be 0 or 1 where 0 means the customer stays with the company and 1 is for when they leave it.

Since we could not collect enough data for using the model, we have generated 1000 records that will be used for this purpose. The data has been generated using Mockaroo (See samples in Figure. 3). All the records are generated randomly while the churn value was generated using a customized random distribution as it is supposed to be at least slightly dependent on or influenced by the other factors in the dataset.

The attributes- likeliness_to_recommend, product_quality, delivery_experience, price_aptness, website_design, and customer_service have been used for this and it has been set up such that churn is inclined towards (more probability is given during generation based on the conditions setup) 1 for 1 and 2 ratings and ten towards 0 for 4 and 5 ratings. For example, for likeliness to recommend, if the rating is 5, the churn value is 3 times more likely to be 0. For a rating of 3 associated with an attribute, both 0 and 1 are equally likely. The churn rate for this data is given by-> (number of customers at the beginning - number of customers at the end) / number of customers at the beginning - > number of churned customers/number of customers at the beginning -> (1000-522)/1000 = 0.478 or 47.8%.

Python programming language has been selected to implement the model so the code has been written in a Jupyter notebook for better readability. The data has been cleaned to not have any duplicate records, missing values in any of the attributes, or any null values to ensure that it is suitable to be used in Python. Although the data was properly generated with all valid values, this step was still taken to ensure the data is ready to be used. It has also been explored to see the distribution of values for each column and visualize their impact on the value of churn.

Algorithms
For the task of analyzing the most influential factors behind customers leaving the store and predicting the next move of the customer to either continue shopping or to completely engage with the online store, a machine learning model has been implemented. The model experimented with in this project is called a Decision Tree. Other popular alternatives to decision trees include Random Forests which are a collection of decision trees [4]. These seem to be better alternatives according to [3][5] however, they have not been explored due to it being out of the scope of this project. We have also experimented with Regression apart from classification. These topics will be explained further below.

A decision tree is a supervised (trained with labels) machine-learning model which learns from the decision rules of the data and its labels. Its structure resembles that of a flowchart, and it consists of nodes where ones without any further branches (or children) are called leaves. The target variable is understood by analyzing the attributes associated with it. This has been selected because it is a popular choice for this purpose, it is easy to comprehend and can be visualized using Python which enhances the experience further. However, they can have overfitting of data which means it underperforms unseen data.

The decision tree classification model (to classify instances as churn or no churn) has been instantiated from [7]. The values it was initialized with are criterion- ‘gini’, ‘min_samples_leaf’ is 3, ‘max_depth’ is 5, and splitter is 'random'. Criterion is the splitting criterion (a measure of node impurity) which has been selected to be entropy, minimum samples leaf is the number of leaves to be there as the leaves, maximum depth is the depth of the tree or the number of levels down the tree, and splitter is the condition at which the node should be split (best random split is chosen) [7]. Gini is the measure of impurity or of how many instances were misclassified at a split and values close to 0.5 mean the assignment was completely random (0 is pure) [2]. These have been selected after performing Grid Search over the values.

Grid Search is an exhaustive search over all possible values of the hyperparameters to select the best ones for a given dataset and model and it has been implemented using [6]. The decision tree generated is given below. It can be inferred that the most influential factor given on the root node (first node) is the customers’ likeliness to recommend. The data has been split into training and testing datasets in a 3:1 ratio randomly.

‘likeliness_to_recommend’, ‘time_spent_last’, ‘product_quality’, ‘delivery_experience’, ‘price_aptness’, ‘website_design’, and ‘customer_service’ are used for constructing the decision tree. Training the model involves learning patterns from the data [9].

The trained model’s performance is assessed on the testing dataset as it is unseen data for the model, but it comes from the same distribution as the training set. The chosen metrics for model evaluation are accuracy, precision, recall, and F1 score. Accuracy is the number of correctly identified instances when compared with the actual value. Precision is the number of correct positive predictions made, recall is the number of positives predicted from the entire data, and the F1 score is an average combination of the recall and precision values [1].

The data is almost balanced (contains an almost equal number of instances that have a churn value of 0 and 1) as 53-47 is considered balanced. Due to this reason, accuracy around 50% could mean the model is giving only 0/1 thus matching almost half the records that have that value already. So, it is now a better training set enabling the model to learn both values equally thus improving the model’s ability to produce both 0s and 1s depending on the other values. The final values (unweighted averages) obtained after training and after testing are given in Table 1. Since the values are satisfactory and did not improve after trying different initial decision tree parameters, the search was stopped here.

Process	Accuracy	Precision	Recall	F1 Score
Training	70.6%	71%	70%	70%
Testing	72%	73%	72%	72%

Analysis and Solution
For coming up with a plan to improve customers’ experience with the store, all the attributes suggested by the implemented decision tree and regression model were considered. Since the results obtained from both are similar, they hold greater weight and validate the decisions and changes considered. The likeliness to recommend is the most influential and it does imply that customers who would recommend the website to others would also prefer to continue buying from the same. An idea is that customers can be sent referral coupons after their purchase which can be shared with their acquaintances.

These would give small discounts to the sender and the receiver. Sharing the website or products can also be rewarded with occasional coupons and this would be announced to all. The quality of the products seems to impact the churning too as it is necessary for any business to function for long so it would be taken care of to not be compromised. Additionally, the sustainable nature and sources of material can be advertised for better reach. Time spent last on the website can be improved by adding more products, categories, filters to find products, an AI-based product finder so that they keep coming back to search for any new design, and daily rewards to collect to improve their interaction with the website.

All changes need to be communicated to the customers so that they are aware of these changes and take an interest in exploring them. Customers unsatisfied with these aspects are more likely to leave and vice versa. Good reviews from other customers regarding all these factors are also crucial for others to continue buying products. Regular feedback forms with related questions can be helpful to get their opinions time-to-time.