Customer Churn Prediction, Prevention and Analysis in The Telecommunication Industry

The telecommunication industry is one of the fastest proliferating sectors around the world. Indicators clearly reveal increased competition induces a customer to opt for low cost options. This, in turn with a lack of personal touch from large corporations, can cause disloyalty and can persuade a customer to switch to an alternative supplier, known as ‘customer churning’. Therefore, it has become imperative for the mobile operators to shift their focus from rapid acquisition strategies to aiding customer maintenance and enhancing margins from the existing customer base.

To face this operational challenge, it is essential to utilise the customer behaviour data available to identify unique, actionable factors that influence customer loyalty. This project gained insights into the relationships between churn and different attributes of customer behaviour such as: tenure, contract, payment methods, monthly charges etc. After extracting these valuable insights, a model is built using Logistic Regression predict if a customer will churn. The model returned an AIC (Akaike Information Criterion) of 4158.2, with a VIF (Variance Inflation Factor) of below 2. Since various statistical tests were performed to evaluate the predictions, and avoid over fitting, the model proved accurate at predicting the churn rate of customers.

Identifying the churn risk score for each customer and identifying which customer behaviours anticipate churning, are the essential foundations for targeted proactive retention for customers switching to another supplier. Allowing tailor made marketing action for each and every customer.

Uncovering the associations between a retailers products

Transactional data can be used to develop models that predict which products a user will buy again, try for the first time, or add to their cart next during a session.

Market Basket Analysis is a modelling technique based upon the theory that if you buy a certain group of items, you are more (or less) likely to buy another group of items.

Market basket Analysis can be used as Recommendation System, by using data from MBA company can suggest the next best product which a customer is likely to buy.

This can also be used to give discounts to user on those products which user is not likely going to purchase.

This Problem is to predict which previously purchased products will be in user’s next order by using anonymized data of customer’s orders over time. Our model can predict if a customer will repurchase a certain product again with an accuracy of 87%.

Market Basket Analysis has proved essential for the maintenance of inventory, creating promotion strategies such as cross-selling. Market Basket Analysis can also be used to make decisions regarding product placement in store and online.

Recommendation systems have become ubiquitous in our everyday lives, with uses including but not limited to e-commerce, entertainment, research and academia. A recommendation system is an information filtering system, which predicts the preferences of a user, and makes suggestions based on these preferences. Popular examples include YouTube, Netflix and Amazon, all offering tailored recommendations to users.

These systems can collect information a user’s behaviour, using this information to improve their suggestions in the future. Content-based filtering, is used by recommending items based on interests contained in a user profile, and the contents of the items. Alternatively, collaborative filtering, groups similar users together and uses information about the group to make recommendations to the user.

This project involved building a movie rating collaborative filtering recommendation system to model the data and predict the ratings for the movies which have not yet been given by the users. In the beginning, we extracted the data from the client database and have made a dictionary for every user and the movies they have watched along with their respective ratings. An artificial neural network was used, and using data to test the accuracy of the model, an RMSE value of was found to be 0.96, proving there is no over-fitting in the model.

Parking Solutions

A Parking Solution system is a service in which the number of empty and occupied lots is made available to the users in a parking area on a real-time basis. In this project we used machine-learning to develop an image-based classification system to predict if a parking space is occupied or empty, with the objective of providing this information to the end user in a web-application.

An Inception V3 model is trained and developed using images of car parks. Using an existing camera system, we can get real-time images of a car park, which Inception V3 model can classify the lot as occupied or vacant, and update this information to the database frequently. This information can provide many details regarding the parking lot to end users via web applications.

Parking solution systems are vital for hotels, leisure and retail areas as they gain important insights into customer capacity and peak times, which can aid many other business decisions. By adopting this data led solution, a car park can increase its efficiency and optimise car park usage. Businesses can use our service to give their customers valuable information regarding how busy the area is and how convenient it would be for them to visit.

Email Classification

Spam and scam emails are unsolicited messages sent to consumers, often for the purposes of marketing or performing fraudulent activity. These messages are time consuming, annoying, and most importantly can be a threat to their recipient. Despite the “Junk” folder being used to filter these messages out, recently more advanced spamming techniques have been used. Subsequently our junk folders are often overlooking these spam messages and placing recipients again at risk.

Our team used supervised classification to build a spam filter, to automatically categorise an email based on its content into ‘Spam’ of ‘Ham’. Using natural language processing and a Naïve Bayes classification, a fast and highly scalable method, this machine learning

Not having an accurate spam filter can put you or your company at risk of viruses and harmful malware. A good spam filter will also not let important emails slip through.

It is important for every employee of a company to be working efficiently. By allowing spam emails to flood into inbox’s there will be a considerable amount of time spent looking for important emails and deleting all spam emails. This may seem benign on a day-to-day basis however over the space of months and years hours and days can be wasted.

Title: Web Traffic Analysis and Forecasting

Web traffic is the interaction between users and a website; how many users visit a website, which pages a user clicks and the time spent on a page, to name a few. By analysing this traffic, we can find trends, or the most and least popular pages on a website.

Time series analysis and forecasting can be performed on the web traffic time data, and in this project, we forecasted the counts of people who will visit the web pages within next 61 days. The SMAPE (Symmetric Mean Absolute Percent Error) was chosen as the metric to evaluate our model’s performance, and was found to be 0.13.

Web traffic monitoring is vital for the success of any company. It gives an understanding of which products and services are popular within the page. Monitoring web traffic is only effective, when teamed with analytics and forecasting, for making informed decisions to optimising a website.

Understanding along with business insights allows for panels on the web page containing products and services can be moved to give customers more ease of use when finding the product/service they desire. For the webpage this will increase customer usage as well as sales performances.

Big data is a term used to define very large amount of all varying information logged by a company. Data may be considered ‘big’ depending on. Customary data processing is often inadequate and costly for big data as it requires high computational power, constant maintenance and regular scaling.

This project was designed to create a pipeline to ingest an organisations data using big data technologies. After transferring the data from the local system to an Amazon Web Services (AWS) EC2 instance, there are two ways to proceed with taming the big data.

The first way involves importing the data into SQLite using python, then using the SQLite browser, checking the schema and exporting into SQL format. The SQL file can then be opened in python to run a command which inserts the data into AWS RDS. The second way follows a similar structure, but uses Hadoop HDFS and Sqoop to insert the data into AWS RDS.

Using HDFS and Sqoop is the faster method for operation, however if the data contains unusual information, Sqoop is likely to output errors. Using SQLite and Python will minimise these errors, but with a slightly slower computational time.

The advantages of completing this project for the client include:

• The flexibility of being able to access the data from anywhere
• The system will not become slow when processing the data
• Since the data is not in a local system, if this crashes the data is safe
• Flexible scaling of the system

Title: Bike Sharing Demand

In many cities across the UK, ‘docks’ of bicycles have been made available for public use, under a “Bike-Sharing Scheme”. This scheme allows people to borrow a bike from the dock on a short time basis, either for free or for a price. Bike-sharing aims to reduce the effects of traffic congestion and pollution, by providing an economical and environmental alternative transport system, used by both locals and tourists.

In this project, our team performed extensive and thorough analysis on the data collected, to try and investigate the factors that influenced a person to borrow bike, to ultimately predict how many users a bike sharing scheme may see on a day. We explored attributes like the effect of weather, temperature, year, season and hour on bike users and used a Random Forest machine learning algorithm to predict how many users a bike sharing company had per day. The model predicted with an accuracy of 92.71%, with mean squared residuals of 0.1598 on our test data, and our model showed neither under-fitting or over-fitting.

Computer vision is a rapidly growing and improving field of artificial intelligence, and has practicalities in a diverse range of fields. Humans have no issues recognising the difference between a cat and a dog, however this problem has proved troublesome for image recognition systems.

A set of labelled images of cats and dogs was used to build a machine learning that classifies new images of cats or dogs. Our team built a Deep Convolutional Neural Network, a computational model which works in a similar way to the human brain. Each neuron takes an input, performs an operation, then passes its conclusion onto the next neuron. The model first needed to learn the different features of the given image. To achieve this, we first applied conventional layers, which comprise various techniques like feature detection, max pooling and flattening. After this, we applied the fully connected layers of neural networks to classify the objects. In the model evaluation phase, parameters are tuned to improve the model. The accuracy of the model is above 95% within 10 epochs.

There are many industries where an image recognition system can be used including:

• Medical industry- Gives doctors the ability to diagnose medical conditions,
• Parking industry- Can calculate the time a car spent in the parking lot using recognition of the registration plates to be aware of what that customer must pay. This in turn gives the company knowledge of average time spent in the lot.
• Hotel and restaurants- System can give information of how many spaces they have free in their lot to organise customers, the system would indicate how many free spaces are available.

Customer retention has a significant impact on bank's profits. This kind of impact has exceeded that caused by scale, market share, unit cost and other relevant factors of competitive advantage. Customers churn does not only bring the effect level of sales decreasing, and it would also decrease the number of new customers using that bank.

This project has various applications, which can be applied to modern-day industries such as identifying the behaviour of the customers and employees.

We build a Deep Artificial Neural Network in Azure, Deep Learning Virtual Machine. The Deep Learning Virtual Machine (DLVM) is a specially configured variant of the Data Science Virtual Machine (DSVM) to make it easier to use GPU-based VM instances for training deep learning models. The accuracy of the model is 84% within 100 epochs. In model evaluation Parameters Tuning and K Fold is used to improve the model.

A small improvement in a customer retention rates would yield a considerable increase in profits. Therefore, a different marketing strategy could be carried out for different a client base. This would attract a new and different group of clients to the business and a client base that brings in profits from areas not attained before. The aim of this project is to find the unusual trends happening inside, which is causing unusual churning of customers.