We organized the KDD cup 2009 around a marketing problem with the goal of identifying data mining techniques capable of rapidly building predictive models and scoring new entries on a large database. Customer Relationship Management (CRM) is a key element of modern marketing strategies. The KDD Cup 2009 o ered the opportunity to work on large marketing databases from the French Telecom company Orange to predict the propensity of customers to switch provider (churn), buy new products or services (appetency), or buy upgrades or add-ons proposed to them to make the sale more pro table (up-selling). The challenge started on March 10, 2009 and ended on May 11, 2009. This challenge attracted over 450 participants from 46 countries. We attribute the popularity of the challenge to several factors: (1) A generic problem relevant to the Industry (a classi cation problem), but presenting a number of scienti c and technical challenges of practical interest including: a large number of training examples (50,000) with a large number of missing values (about 60%) and a large number of features (15,000), unbalanced class proportions (fewer than 10% of the examples of the positive class), noisy data, presence of categorical variables with many di erent values. (2) Prizes (Orange o ered 10,000 Euros in prizes). (3) A well designed protocol and web site (we bene tted from past experience). (4) An e ective advertising campaign using mailings and a teleconference to answer potential participants questions. The results of the challenge were discussed at the KDD conference (June 28,2009). The principal conclusions are that ensemble methods are very e ective and that ensemble of decision trees o er o -the-shelf solutions to problems with large numbers of samples and attributes, mixed types of variables, and lots of missing values. The data and the platform of the challenge remain available for research and educational purposes at http://www.kddcup-orange.com/.
Customer Relationship Management (CRM) is a key element of modern marketing strategies. The KDD Cup 2009 o ered the opportunity to work on large marketing databases from the French Telecom company Orange to predict the propensity of customers to switch provider (churn), buy new products or services (appetency), or buy upgrades or add-ons proposed to them to make the sale more pro table (up-selling). The most practical way to build knowledge on customers in a CRM system is to produce scores. A score (the output of a model) is an evaluation for all target variables to explain (i.e., churn, appetency or up-selling). Tools producing scores provide quanti able information on a given population. The score is computed using customer records represented by a number of variables or features. Scores are then used by the information system (IS), for example, to personalize the customer relationship.
The rapid and robust detection of the most predictive variables can be a key factor in a marketing application. An industrial customer analysis platform developed at Orange Labs, capable of building predictive models for datasets having a very large number of input variables (thousands) and instances (hundreds of thousands), is currently in use by Orange marketing. A key requirement is the complete automation of the whole process. The system extracts a large number of features from a relational database, selects a subset of informative variables and instances, and efficiently builds in a few hours an accurate classiffier. When the models are deployed, the platform exploits sophisticated indexing structures and parallelization in order to compute the scores of millions of customers, using the best representation.
The challenge was to beat the in-house system developed by Orange Labs. It was an opportunity for participants to prove that they could handle a very large database, including heterogeneous noisy data (numerical and categorical variables), and unbalanced class distributions. Time eciency is often a crucial point. Therefore part of the competition was time-constrained to test the ability of the participants to deliver solutions quickly. The fast track of the challenge lasted ve days only. To encourage participation, the slow track of the challenge allowed participants to continue working on the problem for an additional month. A smaller database was also provided to allow participants with limited computer resources to enter the challenge.
This challenge uses important marketing problems to benchmark classi cation methods in a setting, which is typical of large-scale industrial applications. A large database was made available by the French Telecom company, Orange with tens of thousands of examples and variables. This dataset is unusual in that it has a large number of variables making the problem particularly challenging to many state-of-the-art machine learning algorithms. The challenge participants were provided with masked customer records and their goal was to predict whether a customer will switch provider (churn), buy the main service (appetency) and/or buy additional extras (up-selling), hence solving three binary classification problems. Churn is the propensity of customers to switch between service providers, appetency is the
propensity of customers to buy a service, and up-selling is the success in selling additional good or services to make a sale more pro table. Although the technical difficulty of scaling up existing algorithms is the main emphasis of the challenge, the dataset proposed offers a variety of other difficulties: heterogeneous data (numerical and categorical variables), noisy data, unbalanced distributions of predictive variables, sparse target values (only 1 to 7 percent of the examples examples belong to the positive class) and many missing values.