230 likes | 234 Vues
Bayesian Network Classifiers for Identifying the Slope of the customer Lifecycle of Long-Life Customers. Authored by: Bart Baesens, Geert Vertraeten, Dirk Poel, Michael Petersen, Patrick Kenhove, Jan Vanthienen Presentation by: Oksana Myachina, Jeff Janies. INTRODUCTION.
E N D
Bayesian Network Classifiers for Identifying the Slope of the customer Lifecycle of Long-Life Customers Authored by: Bart Baesens, Geert Vertraeten, Dirk Poel, Michael Petersen, Patrick Kenhove, Jan Vanthienen Presentation by: Oksana Myachina, Jeff Janies
INTRODUCTION • Acquiring a new customer is more costly, than selling additional products to existing ones. • Traditional brand strategies should be replaced by customer strategies. • It’s very important to make informed decisions on customers level.
CRM is successful only if customers remain at least to a certain extent ,loyal to the company in case. • Research shows large heterogeneity in long-term customers spending. • Responding to this fact , the study explained in the paper,was performed.
The relevance of estimation of a customer’s spending evaluation • Traditional relationship marketing claims: - loyal customers raise their spending - generate new customers - ensure diminishing serving costs - have reduced consumer price sensitivity • RM main idea : the longer customer stays loyal to company, the more Profit it has
Reinartz and Kumar state that LLC are not necessary: - cheaper to serve - less price sensitive - more effective in bringing new business to the company • Mail Company example
What is the aim of the study? To elaborate an accurate indication of customer’s future spending evaluation To account for heterogeneity within the group of long-life customer To estimate whether newly acquired customers will increase or decrease their future spending
Aim and Methodology • Binary classification problem: 'Will newly acquired customers increase or decrease their spending after their first purchase experiences?‘ • Previous experience: • traditional statistical methods • nonparametric statistical models • neural networks • Innovation -adaptation of Bayesian network classifiers
Naïve Bayes classifiers • Often work well in practice • Learns the class-conditional probabilities P( Xi = xi | C = cl) • New test cases are classified by using Bayes’ rule to compute the posterior probability of each class cl given the vector of observed variable values (see handout)
TANs • Tree Augmented Naïve Bayes Classifiers (TANs) • Extension of the Naïve Bayes Classifiers • Relax the independence assumption by allowing arcs between the variables • The class variable has no parents and each variable has as parents the class variable and at most one other variable • The variables are only allowed to form a tree structure
GBN: Learning Algorithm • Assumes an a priori ordering of the variables • D-separation plays a pivotal role in the structure learning algorithm • A four phase algorithm • Create a draft • Add and remove arcs based on the concept of d-separation and conditional independence • Establish parameters
Multinet Bayesian Network Classifiers • GBN and TANs assume relations between the variables are the same for all classes • Multinet Bayesian networks allows for more flexibility and is composed of a separate, local network for each class and prior probability distribution of the class node • (see handout for formulas)
Other Methods used, but not discussed • CL multinet • C4.5 and C4.5rules • White-box classifiers for classification decisions • Linear Discriminant Analysis (LDA) • Well-known benchmark statistical classifiers • Quadratic Discriminant Analysis (QDA) • Well-known benchmark statistical classifiers
Training • Naïve Bayes and TAN used Matlab toolbox of Kevin Murphy • GBN and GBN multinet classifiers used PowerPredictor software
Data Set • Variables of the Study • Time Frame • Attributes, Values, and Encodings
Performance Classification • Measured by area under the Receiver operating characteristic curve (AUROC) • Uses a 2D graph of the sensitivity on the Y-axis (true alarms) versus the false alarms on the X -axis
Performance Classification • Percentage of correctly classified (PCC) • This is the most commonly used measure of performance of a classifier • Contingency table analysis to detect statistically significant performance differences between classifiers.
The results • Naïve Bayes and TAN did not remove any attributes • TAN added 14 arcs to the Naïve Bayes classifier with minimal performance improvement • GBN multinet looks simpler, but bad performance • GBN classifier was able to prune 12 attributes
Practical implementation • Marketing investment decision • Monitor of customer-acquisition policies • To design an a-priori segmentation scheme for a company's customer base