Data Mining: The Brains Behind eCRM

Introduction

While customer relationship management (CRM) gets a lot of attention, one important component is often overlooked - how a company effectively uses its existing customer data to improve customer loyalty while also identifying new potential buyers of products and services. Data mining has emerged from obscure beginnings in artificial intelligence to become a viable and increasingly popular tool for putting data to work.

What is eCRM?

Customer Relationship Management is an information industry term for methodologies and software that help companies manage customer relationships in a structured way. For example, an enterprise might build a database about its customers that describe relationships in sufficient detail. Ideally, the information in the database can then be accessed by management, salespeople, helpdesk personnel, and others to match customer needs with product plans and offerings, remind customers of service requirements, and know what other products a customer has purchased. CRM of course existed before there was an information industry; even the great salesmen of literature such as Willie Loman and "Professor" Harold Hill owed their successes to knowing their customers and what those customers wanted.

CRM has been a vital information industry segment for many years, but the special needs of Internet businesses for real-time processing of huge volumes of data led to the creation of a flock of startups and the coining of the term "eCRM" to describe their niche. Some of the special characteristics of eCRM were that it primarily served (and was offered by) dot-com businesses, that it tended to be weak on the traditional kinds of CRM processing, and that it made effective uses of Internet technology, such as connecting an on-line customer with a phone representative and automatically displaying the customer's data on the representative's screen. However, as bricks-and-mortar firms expand to the Internet and old-line CRM and analytics firms do likewise, the line between CRM and eCRM will become too hazy to matter.

Some objectives of eCRM are to:

• Enable marketing departments to identify and target their best customers, manage marketing campaigns with clear goals and objectives, and generate quality leads for the sales team
  
  
• Assist organizations in improving telesales, account, and sales management by optimizing information shared by multiple employees, and streamlining existing processes
  
  
• Tailor relationships with customers to improve customer satisfaction and maximize profits
  
  
• Identify the most profitable customers and provide them the highest level of service
  
  
• Provide employees with the information and processes necessary to understand their customers' needs and build strong relationships between the company, its customer base, and distribution partners

What is Data Mining?

Data mining is an automated approach for discovering or inferring hidden patterns or knowledge buried in data. By "hidden" we mean patterns that are not made apparent through casual observation. As a component of eCRM, data mining can aid marketing personnel by helping to answer the following types of questions:

• Are there additional people whom I should market to, or whom I am currently missing?
  
  
• Which product traits do my customers perceive most favorably?
  
  
• Can I create additional products which contain the best of these traits?

Though various techniques for data mining exist, most of these fall into one of the following two categories, supervised and unsupervised learning. Users apply data mining software employing one of these two learning algorithms to construct a data mining model.

Supervised and Unsupervised Learning

In supervised learning, the user constructs the model by supplying example sets of input data and corresponding sets of output data. Training pets provides a good analogy for supervised learning. When training a dog to roll over, the owner will say "roll over" (the input), then grasp Fido on either side and physically rotate him laterally over the ground to simulate the desired response: a roll (the output). Eventually the dog will respond automatically to the "roll over" command with the appropriate action. The analogy fails, however, when extending it to data mining. In data mining, the trained model needs to respond to more than just known inputs, but must give meaningful results when receiving novel inputs. (Our dog would be confused by "roll out" or "roll up".)

A data mining example might involve a marketing manager who wants to spur the sales of a new CD. The manager would comb the database for records of customers who had already purchased the CD, and present some of these to the learning algorithm, instructing it to recognize from the other purchases that these are likely purchasers of the target CD. The algorithm would be tested with the remaining customer records. Records from customers who didn't purchase that CD would also be used in the training and testing. If the algorithm shows a high rate of predicting purchasers from non-purchasers it would be used as a recommendation engine; when customers visit the site those that score highly would be offered the target CD as a recommendation.

Data mining models created by supervised learning are often applied in cases where users wish to understand the outcome or consequences of a given set of customer profile data never before encountered. By supplying both the inputs and expected outputs, supervised learning allows users to create models for prediction and identification.

In contrast to supervised learning, unsupervised learning constructs data mining models independent of any supplied performance criteria. The algorithm that creates these models searches through a large collection of input data and finds relationships on its own without requiring the user to supply an expected output. In our dog training analogy, this would be equivalent to shouting a series of commands at Fido and letting him decide how to respond. Similar commands (measured within a specified tolerance) would presumably cause him to perform the same trick, but the trick might be completely new, i.e., one that we might not have anticipated (like fetching leftovers from the refrigerator).

Unsupervised learning models are ideal for identifying novel patterns in data that might otherwise go undiscovered. For example, if a book is reviewed favorably on nationwide radio show its sales might suddenly shoot up. A website would want to recognize this and recommend the book to likely buyers. Since the review was not known about in advance and the phenomenon only lasts for about a day, the website needs to be able to recognize the pattern as it emerges and respond to it without marketing intervention. Indeed, on any given day there might be hundreds of such short-lived patterns that the website could profit from.

Summary

BOTTOM LINE

User Recommendations

Data mining capabilities can play an important role in successful customer relationship management. In evaluating a data mining package, users should ask in which category, supervised or unsupervised, it falls and whether it is suitable for their needs. For instance, insurance companies seeking to predict insurance fraud should consider packages that use unsupervised learning, as there are typically no baselines to serve as predictors. Consumer products retailers can use unsupervised techniques to discover customer buying patterns in cash register receipt data in order to group items optimally in their stores. Supervised learning should be used where one wishes to measure the deviation from a known result. For example, a marketing department may know what product combinations (e.g., cell phones and accessories) are purchased by certain age groups. Supervised learning can create a model that records these known "truths" about customer buying habits which can then be used to qualify buying data that does not conform exactly to these truths. This can help the marketers adjust their strategy accordingly.

For many companies the benefits of data mining are first realized in the area of marketing. The process of collecting and segmenting customer data for effective marketing campaigns has traditionally been more of an art than a science. However, automated approaches such as data mining provide businesses with better ways to plan and manage marketing campaigns. Companies with large customer databases should consider a data mining solution for campaign management and customer analysis that can leverage their data volumes to improve customer loyalty, qualify and understand their customers, and generate repeat business.

Editor's Note:
This article has been modified from its original form since the original publication date.