18-10-2016, 02:39 PM
1459678988-DocumentationChurnHub.docx (Size: 342.59 KB / Downloads: 4)
1. Abstract
As markets become increasingly saturated and competitive, user base churn prediction and retention management has become of great concern to many industries. A company wishing to retain its customers needs to be able to predict those who are likely to churn in the near future and will make those, the focus of customer retention efforts. Today, the data available is extremely large and in particular, the data collected in association with activities of what users do can be of large samples, high dimensions and more noise-prone . There are three possible strategies to generate more revenue: acquire more customers, upsell existing customers, or increase customer retention. All the efforts made as part of one of the strategies have a cost, and what we’re ultimately interested in is the return on investment: the ratio between the extra revenue that results from these efforts and their cost. But the cost of retaining an existing customer is less than acquiring a new customer. Understanding what keeps customers engaged, therefore, is incredibly valuable. Consequently, there's growing interest among companies to develop better churn-detection techniques, leading many to look to data mining and machine learning for new and creative approaches. Given all the research that has been carried out in this area, we could use the past churn results to find patterns and dependencies based on feature extraction from the available data, making use of that to build a predictive model, which could be used to say who will churn or not in the near future. This when done as a third party app, where in it is not restricted to one single product, but like Software as a service product, make it highly usable and profitable.
Introduction
With the rapid development of the Internet, the global enterprises establish a global trading network through the Internet to provide consumers with greater choice, e-commerce has brought great changes to business process and consumer behaviour. Compared with the traditional methods, the biggest disadvantage of e-commerce is the churn rate of their users is very high, therefore identifying those customers likely to churn, to take appropriate measures to reduce customer churn, maximizing the enterprise’s profits, e-commerce has become a hot field of research. Aiming to the loss forecast for ecommerce customers, many scholars have conducted in depth and extensive research, and have achieved good results.
E-commerce customers currently are classified into two kinds of prediction methods: statistical analysis and artificial intelligence. Statistical analysis includes linear regression, time series, cluster analysis, decision trees and Bayesian networks, etc. These methods are all supposed that e-commerce customer loss is presenting a linear variation, but because of the minds of consumers, buying behaviour, economics, culture and other factors, the loss of e-commerce presents nonlinear and high-dimensional characteristics, statistical analysis model does not fully reflect the characteristics of e-commerce customer churning. Artificial intelligence technology has self-learning ability and nonlinear processing capabilities, with respect to statistical analysis, the correct rate of predict is improved. However, the actual customer loss data has the characteristics with noise, the sample being extremely uneven, the high dimensionality and high nonlinearity.
The predicted difficulty is increased, using a single prediction model is hard to achieve accurate churn prediction in e-commerce. If multiple predictive models are combined together, each of predictive models play its prediction advantage, the accuracy of customer churn prediction in E-commerce is expected to improve .Thus, most of the churn prediction modelling methods rely on quantifying risk based on static data and metrics, i.e., information about the customer as he or she exists right now. The most common churn prediction models are based on older statistical and data-mining methods, such as logistic regression and other binary modelling techniques. These approaches offer some value and can identify a certain percentage of at-risk customers, but they are relatively inaccurate and end up leaving money on the table. The users are where the money comes from. Although a few remain with the site forever, it does make sense to keep their dwell time high, and their attrition rate as low as possible. Naturally, there are exceptions to this. Subscriber churning (also known as Customer attrition) in many industry refers to the movement of subscribers from one provider to another. Many subscribers frequently churn from one provider to another in search of better Rates/services or for any other reason.
It is estimated that the average churn rate for the mobile telecommunications is 2.2% per month (Berson et al., 2000).That is, about 27% of a given carrier's subscribers are lost each year, making it essential to develop an effective churn-reduction method.The cost of acquisition of a new mobile service subscriber is estimated to be from $300 to $600 in sales support, marketing, advertising, and commissions (Berson et al., 2000; SPSS, 1999).
Now, let's look at what churn rate is and why the present type of modelling churn is difficult,
Churn is at its most basic level the percentage of subscribers who leave within a given period. This makes a useful reciprocal for retention rate. Although subscribers are not necessarily the only users, we are more-or-less stuck with them when measuring churn. This is because we have no way to assess the moment when casual customers drift away (although perhaps somewhere, sometime, someone will come up with the metric).
"Churn Rate" is a business term describing the rate at which customers leave or cease paying for a product or service. It's a critical figure in many businesses, as it's often the case that acquiring new customers is a lot more costly than retaining existing ones (in some cases, 5 to 20 timeS more expensive).
2.1. Benefits
Having sticky customers is great for future cash flow. It helps us budget - and set stock levels and reorder points if we are selling goods. We can also use predicted turnover to streamline our human resources, so all-in-all low churn is good for business, but that’s not all.
Users grow in loyalty as they stay longer with us, and become more open to online surveys and one-on-one discussions (although some prefer we leave them be). The social ones are great for trialing new ideas. Which one of us does not appreciate a supplier asking us for our opinion?
Finally, stable customers are lower maintenance. Once we have them up and running and they know our systems, we can largely leave them on their own. This is a huge saving on expensive call-center operators, provided, that is, we remember to tell them before something changes.
2.2. Reasons for Churn
• Spoiler alert - the biggest reason customers leave is not, in fact, because they found a better price.
• Value is another top reason for customer churn. However, it's important not to equate "value" with "price." In this case, value is a more encompassing term that describes the entire customer experience. This is especially apparent in that 55% of customers would pay extra to guarantee better service, instead of just opting for the lowest price.
• Another top reason for customer churn is due to amateurish and unengaging communication tactics that end up turning off the customer. With email subscription rates, for example, 35% of customers unsubscribe from emails because they are sent too frequently.
• When shoppers leave, it is usually not because they've sworn off your product forever. Generally, shoppers run from you and straight into the welcoming arms of your competitors (again, most often due to service-related complaints, per Bain & Company).
Clients churn away for a variety of reasons. They may choose to do so voluntarily because they find a supplier they prefer, or because we managed to annoy them with poor service or something else. There can also be involuntary reasons. They could die (we all do sometime), suffer a disablement, or switch to another lifestyle where they either cannot afford, or no longer need our offering.
Clearly, we do not want to lose good, paying customers unnecessarily. The main reason for retaining users is that the cost of acquiring new customers is higher than the cost of keeping existing customers. Let's take an example, Quettra Mobile Analytics company's tabular data showing the retention rate of Android apps,
This proves that even a normal app, which is used on the world’s biggest mobile platform has a churn rate of greater than 70%, then we are doing something wrong.
The existing system do the following,
• Clustering : Classifies the customers according to different levels of churn risk.
• Map: Allows to see the geographical nature of churn, verifying it through relative frequency.
• Profile :Identifies which variables describe these customers and which do not, in order to know in depth how they are.
• Venn Diagram : Creates a quantile in the three groups of benefits, to see the value that churns have for the company.
• Predictive modelling : creating models that predict whether user leaves the app based on patterns found in historical data.
• Retention Strategy : Motivating users to avoid churn and introducing offers to regain users
While in this documentation we will go through a dynamic system, which will incorporate the features periodically. The system is developed in Ruby on Rails and Python scripts.
. Literature Survey
With the growth of the online industries, in particular applications which offer software as a service, there is a huge increase in the graph of online users, making the users a most important product to be taken care of. In an attempt to leverage and lure the competition to one’s side, there is a huge battle going on out there to make a mark on the customer. In an attempt to go beyond, researches have analysed a lot of user bases and statistically analysed with the balance sheets of marketing departments and operations, to find that user churn reduction and retention is much more viable than acquiring a new user.Ankit Jain of Quettra, in his paper with Andrew, quotes the same. To understand totally about churn the book, “Evergage Guide to Churn Analysis”, contains the mathematical model behind the research done in this area.
Now, knowing this, the next question which was arrised was “How?”, because user base churn modelling was one of the toughest, as the team at Kissmetrics quoted during their project of modelling churn, “Why modelling churn is difficult? “, A major challenge with modelling is churn rates change each month, one data used may not be useful for the other.
There has been a lot of research going into this, modelling the churn pattern by running algorithms through millions of user records and datasets. Tang Chiu and Ping Wei, tried to model telecom churn, “ Turning telecommunications call details to churn prediction: a data mining approach” (2002) [2] , where in he took a data mining approach. Similarly, Gideon Dror of Yahoo, tried the community based CQA churn rate prediction, “ Churn Prediction in New Users of Yahoo! Answers ” [3], which uses a set of classification algorithms to obtain the same.
Given all these, large multinational corporations such as Limeroad, Kissmetrics and Netflix have employed these algorithms in their systems. A few links can be found at their engineering blogs.
Much of the earlier work emphasised dominant relationships between the patterns, such as in Junxiang Lu’s “Predicting Customer Churn in the Telecommunications Industry –– An Application of Survival Analysis Modeling Using SAS “ but still there is something missing which still makes the graph face down.Which says it is much more viable if we build a prediction model that could align and learn by itself, predicting when a customer might leave in the near future, and finding the data points or the class which he falls into to build an effective retention strategy for the same user. The same is reiterated in white papers of FPSC.
Clearly, there is scope here for a great deal more research that, makes the present model much more adaptive, dynamic and third party integratable.
4. Problem Statement
Customer churn is a painful reality that all businesses have to deal with. Even the largest and most successful companies suffer from customer churn, and understanding what causes formerly loyal customers to abandon ship is crucial to lasting, sustainable business growth.
Customer churn or subscriber churn is also similar to attrition, which is the process of customers switching from one service provider to another anonymously. From a machine learning perspective, churn prediction is a supervised (i.e. labeled) problem defined as follows: Given a predefined forecast horizon, the goal is to predict the future churners over that horizon, given the data associated with each customer in that horizon.
Our aim is to keep the existing customers and predict potential churn early and respond fast. Identify the signs of potential churn, understand customer wants and needs and automate campaigns designed to revive and renew loyalty for a solid strategy that minimizes acquisition costs.
5. Problem Solution
The churn prediction problem represented here involves
• Feature selection: Selecting the most relevant attributes
• Feature extraction: Combining attributes into a new reduced set of features
• Logistic regression: Logistic regression measures the relationship between the categorical dependent variable and one or more independent variables by estimating probabilities using a logistic function, which is the cumulative logistic distribution. In our case, we perform logistic regression to measure the relationship between the churn attribute and other attributes.
• Training Phase: This is used to build up our prediction model. This phase tries to tune itself to the quirks of the training data sets. The input includes dataset on the past history of the churned customers in the firm.
• Cross-Validation: This is used to compare the performances of the prediction algorithms that were created based on the training set. We choose the algorithm that has the best performance.
• Predictive Modelling: Now the model is trained with highest accuracy, the model must be able to predict the list of churners from the real dataset which does not include any churn label.
• Deployment: The model is deployed on a rails application. The application provides options for a company to upload the present customer’s dataset and get the churn probability of each customer. This helps to identify the possible churners in advance before they leave the company.
• Retention Strategy: The company then prevent customers who are likely to churn in future by taking the required retention policies like a push mail giving offers to attract the likely churners and retain them.
6. Novelty
The method developed is a two stage process, combining both logistic regression and predictive modelling, in this way we get a more accurate result, based on the patterns developed dynamically. Say there is a stock market crash, that would certainly be reflected in churn rate of e-commerce sites, but if we have a set model where in we are not updating, then this leads to wrong analytics. Thus this model ensures market adaptability.
ChurnHub, the online dashboard, which offers the same churn analysis as a service enable third party integration like service. This makes it possible for even small players without much computational power to take up churn analysis into their account. This is of ease to access method,
. Feature Extraction
Each piece of information we use to represent customers is called a “feature” and the activity of finding useful features is called “feature engineering.” For churn, we would have 4 types of features:
1. Customer Features : basic information about the customer
• Gender
• SeniorCitizen
• Partner
• CustomerID
• Dependent
2. Support Features : characterizations of the customer’s interactions with customer support (e.g., number of interactions, topics of questions asked, satisfaction ratings)
• Contract
• Paperless
• Payment Method
• PhoneService
3. Usage Features : characterizations of the customer’s usage of the service
• Multiple Lines
• Internet Service
• Online Security
• Online Backup
• Device Protection
• Monthly Charges
• Total Charges
• Tenure
4. Contextual Features : any other contextual information we have about the customer
• Tech Support
• Streaming TV
• Streaming Movies
7.3. Logistic regression
Logistic Regression is a statistical technique capable of predicting a binary outcome. It's a well-known strategy, widely used in disciplines ranging from credit and finance to medicine to criminology and other social sciences. Logistic regression is fairly intuitive and very effective.
Logistic Regression helps to find the dependency of each feature on the churn label. It helps us to get a great overview of the coefficients of the model, how well those coefficients fit, the overall fit quality, and several other statistical measures.
The result object also lets you to isolate and inspect parts of the model output. The confidence interval gives you an idea for how robust the coefficients of the model are. Taking the exponential of each of the coefficients, generate the odds ratios. This tells you how a 1 unit increase or decrease in a variable affects the odds of being churned.
7.4. Predictive Modelling
It is based on the fact that the behavior patterns of individual customers frequently change over time.
Dynamic and updated continually based on changes in the data.
The history of each customer is an extremely important factor determining when and why the customer may churn.
We develop a predictive model using the following strategy
• Past - Using historical data to provide personalization or increase user retention.
• Present - Use current session data to provide personalized content for increase user affinity to site/brands or increase retention.
• Future - Use past data to predict when users are going to churn.
Retention Strategy
This helps in designing a retention strategy for the churn risk customers, i.e.
1. Motivating users to “avoid” characteristics common to churned users, i.e. poor profile is an important characteristic, so you give some “perks” for user to connect Facebook or Twitter.
2. Retargeting users — since you know the user id’s and probably emails of the users of each cluster, then you may reach them in social networks or other apps & motivate to use the app again.
3. Notifying users by Push/Email — you may send out push notifications and/or personalised emails.