data analytics for marketing decision support
DESCRIPTION
Very Useful documents of IBM for Marketing AnalyticsTRANSCRIPT
IBM Research
© 2007 IBM Corporation
Data Analytics for MarketingDecision Support
Saharon Rosset, Naoki Abe*
IBM T.J. Watson Research Center
*Acknowledgements to: Andrew Arnold, Chid Apte, John Langford, Rick Lawrence, Srujana Merugu,
Edwin Pednault, Claudia Perlich, Rikiya Takahashi and Bianca Zadrozny
IBM Research
© 2006 IBM Corporation2
Tutorial outline
Challenges of marketing analytics (SR)
– Integrating the “marketing” and “data mining” approaches
– Customizing data mining approaches to the challenges of marketing decision support
Survey of some useful ML methodologies (NA)
– Bayesian network modeling
– Utility-based classification (Cost-sensitive learning)
– Reinforcement learning and Markov decision processes
Detailed analysis and case studies:
– Customer lifetime value modeling (NA)
– Customer wallet estimation (SR)
IBM Research
© 2006 IBM Corporation4
The grand challenges of marketing
Maximize profits (duh)
Initiate, maintain and improve relationships with customers:
– Acquire customers
– Create loyalty, prevent churn
– Improve profitability (lifetime value)
Optimize use of resources:
– Sales channels
– Advertising
– Customer targeting
IBM Research
© 2006 IBM Corporation5
Some of the concrete modeling problems Channel optimization
Cross/up-sell (customer targeting)
New customer acquisition
Churn analysis
Product life-cycle analysis
Customer lifetime value modeling
– Effect of marketing actions on LTV?
Advertising allocation
RFM (Recency, Frequency, Monetary) analysis
...
IBM Research
© 2006 IBM Corporation6
Data analytics for decision support: grand challenge
Beyond “modeling” the current situation, we need to offer insight about the effect or potential of possible actions and decisions:
– How would different channels / incentives affect LTV of our customers?
– How much more money could this customer be spending with us (customer wallet)
– Can we predict the effects of new actions that have never been tried in historical data? What if they have been tried on non-representative set?
– Can we be confident our results are actionable? Can we differentiate causality from correlation in our models?
IBM Research
© 2006 IBM Corporation7
Tutorial outline
Challenges of marketing analytics
– Integrating the “marketing” and “data mining” approaches
– Customizing data mining approaches to the challenges of marketing decision support
Survey of some useful ML methodologies
– Bayesian network modeling
– Utility-based classification (Cost-sensitive learning)
– Reinforcement learning and Markov decision processes
Detailed analysis and case studies:
– Customer lifetime value modeling
– Customer wallet estimation
IBM Research
© 2006 IBM Corporation8
CRM analytics:
– Relies on primary research (=surveys) to understand needs and wants
– Relies on (more or less) detailed models of customer behaviorUsually parametric statistical models
– Often estimates customer-level parameters
Data mining:
– Typically relies on data in Data Warehouse /Mart
– Uses minimum of parametric assumptions
– Often attempts to fit problem into “standard” modeling framework: classification, regression, clustering...
Typical marketing analytics vs. data mining
IBM Research
© 2006 IBM Corporation9
Comparison of approaches
Criterion Marketing DM
Parametric models formalize knowledge of domain and problems
+ -
Robust against incorrect assumptions about domain and problems
- +
Actively collect the data to estimate model quantities (active learning)
+ -
Rely on existing, abundant data in Corporate Data Warehouses
- +
Integrate expert input from managers and customers (“wants and needs”)
+ -
Use data to learn new, surprising patterns about customer behavior
- +
IBM Research
© 2006 IBM Corporation10
Rust, Lemon and Zeithaml (2004), “Return on Marketing: Using Customer Equity to Focus Marketing Strategy”, J. of Marketing
Modeling customer equity / lifetime value
– Combine several previous approaches
– Model the brand “switching matrix” as a function of customer preference, history and product properties
– Want to identify drivers of satisfaction (levers)
– Calculate effect (ROI) of marketing actions – pulling levers
Mostly relies on primary research collected specifically for this study
– Interviews with managers
– Survey of consumer preferences
Example 1: modeling and improving LTV
IBM Research
© 2006 IBM Corporation11
Simplified version of paper’s business model
Marketing investment
Costs
Pullinglevers
Increasedequity
Return on marketing investment
Main goals:
Identify relevant levers
Quantify their effect
IBM Research
© 2006 IBM Corporation12
Analytic setup (main components only)
logit(pijk) = 0k LASTijk + xik k
– pijk is probability that customer i buys item k given they bought item j previously
– LAST is a dummy variable for “inertia”
– Xik is a feature vector for customer i, product k
This is used to compute the brand switching matrix {pijk} and customer lifetime value is calculated as:CLVij = t PROFij Bijt
– PROF is a profit measure considering discounting, price & cost (assumed known)
– Bijt is probability customer i buys product j in time t, calculated
from the stochastic matrix {pijk}
IBM Research
© 2006 IBM Corporation13
Data definitions
Potential drivers (marketing activities) are reflected in the components of xi
– Price
– Quality of service
etc.
The data to estimate the logit model is based on:
– Expert (manager) input
– Questionnaires of customers
– Corporate data warehouse (not implemented in their case study...)
IBM Research
© 2006 IBM Corporation14
Results: important drivers for airline industry?
...
Etc. (all factors deemed important)
Driver Coefficient Std error Z score (coeff/std)
Inertia .849 .075 11.34
Quality .441 .041 10.87
Price .199 .020 9.86
Convenience .609 .093 6.56
......
......
IBM Research
© 2006 IBM Corporation15
What would a data miner do?
Count more (or only) on historical data in data warehouse
– Variables would have different meaning
– Identify correlations, not necessarily drivers
Could use same analytic formulation, but also try alternative approaches
– Relate LTV directly to variables observed?
– Model transaction sizes in addition to switching?
– Use non-parametric modeling tools?
Etc.
IBM Research
© 2006 IBM Corporation16
Common practice in marketing:
Define static, fixed customer segments
– Supposed to capture “true essence” of customers’ behaviors, needs and wants
– Often given catchy names: “Upwardly mobile businessmen” representing the “average” profile
Make marketing decisions at segment level, based on understanding of needs and wants
Example 2: the segmentation approach
IBM Research
© 2006 IBM Corporation17
A market segmentation methodology
Based on Kotler (2000). Marketing Management. Prentice-Hall
1. Survey stage: primary research to capture motivations, attitudes, behaviors
2. Analysis stage: factor analysis, then clustering of survey data Identify segments
3. Profiling stage: analyze segments and give them names
Additional stage often taken is to assign all customers to the defined segments:
4. Assignment stage: build classification model to assign all customers to learned segments
IBM Research
© 2006 IBM Corporation18
What would a data-miner do?
Option 1: clustering
– Replace primary research by warehouse data
– Cluster all customers
– Lose the “needs and wants” aspect
Option 2: supervised learning
– Treat each decision problem as separate modeling taskE.g., find “positive” and “negative” examples for each binary decision, learn model
– Advantage: customized
– Disadvantages:
• May not have right data to model decisions we want to make
• Past correlations may not be indicative of future outcomes
IBM Research
© 2006 IBM Corporation19
Comparison of approaches
Criterion Marketing DM
Parametric models formalize knowledge of domain and problems
+ -
Robust against incorrect assumptions about domain and problems
- +
Actively collect the data to estimate model quantities (active learning)
+ -
Rely on existing, abundant data in Corporate Data Warehouses
- +
Integrate expert input from managers and customers (“wants and needs”)
+ -
Use data to learn new, surprising patterns about customer behavior
- +
IBM Research
© 2006 IBM Corporation20
Count on historical data as much as possible
Avoid complex parametric models
– Let the data guide us
– Still want to integrate domain knowledge
Analyze and understand the special aspects of marketing modeling problems
– Importance of long-term relationship (lifetime value, loyalty)
– Effects of competition (customer wallet vs. customer spending)
Modify existing, or develop new, data analytics approaches to address problems properly
An integrated approach
IBM Research
© 2006 IBM Corporation21
Tutorial outline
Challenges of marketing analytics
– Integrating the “marketing” and “data mining” approaches
– Customizing data mining approaches to the challenges of marketing decision support
Survey of some useful ML methodologies
– Bayesian network modeling
– Utility-based classification (Cost-sensitive learning)
– Reinforcement learning and Markov decision processes
Detailed analysis and case studies:
– Customer lifetime value modeling
– Customer wallet estimation
IBM Research
© 2006 IBM Corporation22
Moving beyond revenue modelingTo really understand the profitability and potential of our
customers, we need to move beyond modeling their short-term revenue contribution
Revenue over time: Lifetime Value modeling
– How much can we expect to gain from customer over time?
– Incorporates loyalty/churn, prediction of future customer revenue
– LTV = t S(t) v(t) D(t) dt (S(t) is customer survival function, v(t) customer value over time, D(t) discounting factor)
Potential revenue: Customer Wallet Estimation
– How much revenue could we be generating from this customer?
– Incorporates competition, brand switching etc.
IBM Research
© 2006 IBM Corporation23
LTV and Wallet: beyond standard modelingTime
Revenue
Now
Future
Next year
Sales / revenue modeling
Sales forecasting
LTV
mod
elin
g
Potential salesActual sales
Wallet estimation
?
IBM Research
© 2006 IBM Corporation24
Types of decision support
Passive decision support
– Understand more about problems and causes
– Identify areas of need, under-performance etc.
– Help in making better decisions
Active decision support
– Model the effect of actions
– Actively help in deciding between alternative actions
Active decision support is typically more challenging in terms of data needed to learn models
IBM Research
© 2006 IBM Corporation25
Depth and actionability of insightsDepth
Actionability
Basic concepts
Real insight
Revenue modeling
ActivePassive
Correlation Causality
Revenue forecast
Lever identification
LTV modeling
Wallet estimation
Understand effect of potential actions on LTV and Wallet
attainment
IBM Research
© 2006 IBM Corporation26
The causality challenge Predictive models discover correlation
– Example: linear regressionSignificant t-statistic for coefficients imply they have a significant effect, not that they are actually causing the response
For active decision support we need to identify levers to pull to affect outcome
– Only works with causality
Causality is difficult to find or prove from observation data
– If we have knowledge about causality, we can formalize it as (say) Bayesian network and use in our models
– We can get closer to causality by case-control experiments
IBM Research
© 2006 IBM Corporation27
Assume we observe for some companies:X = company’s marketing budget,Y = company’s salesand want to understand how to affect Y by controlling X
Assume we find that X is very “predictive” about Y
Possible scenarios:
Illustration: predictive power is not causality
Z
Y X
x y
x y
Causality successfully identified “lever”
Fixed percent of revenue to marketing?
Z=Company size independently determining both quantities?
IBM Research
© 2006 IBM Corporation28
Some other challenges
Modeling effects of new/unobserved actions
– Critical for active support, often difficult or impossible
– Even for established actions, they may have been applied in different context than our planned campaign
Integrating expert knowledge into process
– Can be done formally via graphical models
Handling data issues: matching, leaks, cleaning
– Always critical
Delivering solutions and results
IBM Research
© 2006 IBM Corporation29
Example: Telecom Churn Management
Cell phone company has set of customers, some leave (churn) every month
The goals of a Churn Management system:
Analyze the process of churn
– Causes
– Dynamics
– Effects on company
Design policies and actions to improve the situation
– Marketing campaigns
– Incentive allocation (offer new features or presents)
– Change in plans to contend with competition
IBM Research
© 2006 IBM Corporation30
First step: understand current situation Who is likely to churn (predictive patterns)?
– Phones features / plans
– Usage patterns
– Demographics
Tools: segmentation, classification, etc.
Which of these patterns are causal?Tools: expert knowledge, Bayesian networks, etc.
Which causal effects not in data? Competition, economy etc.
Which of these customers are profitable?
– Short term: customer value
– Long term: lifetime value
– Growth potential: customer wallet
IBM Research
© 2006 IBM Corporation31
Second step: design actions Can we affect causal churn patterns?
– For example, by improving customer service
Given possible incentives and marketing actions, what effect will they have?
– Loyalty and relationship
– Current customer value and wallet attainment
– Customer lifetime value
– Cost to company
How can we optimize use of our marketing resources?
– Identify segments we want to retain
– Identify effective marketing actions
IBM Research
© 2006 IBM Corporation32
Tutorial outline
Challenges of marketing analytics
– Integrating the “marketing” and “data mining” approaches
– Customizing data mining approaches to the challenges of marketing decision support
Survey of some useful ML methodologies
– Bayesian network modeling
– Utility-based classification (Cost-sensitive and active learning)
– Reinforcement learning and Markov decision processes
Detailed analysis and case studies:
– Customer lifetime value modeling
– Customer wallet estimation
IBM Research
© 2006 IBM Corporation33
Survey of Useful Methodologies
Bayesian Networks
– Motivation: need to address causality vs. correlation issue; need to formalize domain knowledge about relationships in data
– Example domain: Customer wallet estimation
Utility-based classification* (Cost-sensitive Learning)
– Motivation: need to handle utility of decision and cost of data acquisition in marketing decision problems
– Example domains: Targeted marketing, Brand switch modeling
Markov Decision Processes (MDP) and Reinforcement Learning
– Motivation: need to consider long term profit maximization
– Example domain: Customer lifetime value modeling
*c.f. Utility-Based Data Mining Workshop at KDD’05 and KDD’06
IBM Research
© 2006 IBM Corporation34
Bayesian Network a.k.a Graphical Model
P(E) ¬P(E)
0.3 0.7
E P(C) ¬P(C)
F 0.4 0.6
T 0.7 0.3
M C P(R) ¬P(R)
F F 0.3 0.7
T F 0.9 0.1
F T 0.2 0.8
T T 0.6 0.4
Bayesian Network is a directed acyclic graphical model and defines a probability model
Here is a simple example…
Economy
Marketing Competition
Revenue
P(M,E,C,R) = P(E) P(M|E) P(C|E) P(R|M,C)
E P(M) ¬P(M)
F 0.3 0.7
T 0.9 0.1
IBM Research
© 2006 IBM Corporation35
Bayesian Network as a General Unifying Framework
Bayesian Network provides a general framework that subsumes numerous known classes of probabilistic models, e.g.
– Naïve Bayes Classification
– Clustering (Mixture models)
– Auto regressive models
– Hidden Markov models, etc, etc
Bayesian Network provides a framework for discussing modeling, inference, causality, hidden variables, etc
Naïve Bayes classification
Class
Variable 1 Variable N…. Variable 1 Variable N….
Clustering/Mixture
Unobserved
Class
Hidden Markov Model
Symbol Symbol
State State
Unobserved
IBM Research
© 2006 IBM Corporation36
Estimation and Inference Problems for Bayesian Networks
Parameter estimation from data given structure
– Given a graphical structure as input, and a model class, estimate the parameters of the models
Inference given model
– Given a full Bayesian network (i.e. graph and model parameters) and partial information on the realized values, infer the unknown values
– Useful for business scenario analyses
Latent variable estimation given structure
– Given a full Bayesian network and data for the observed variables, infer the values for the unobserved (latent) variables
Bayesian network structure learning from data
– Given data only, infer the best Bayesian network, including both the graphical structure and the model parameters
Inferring causal structure from data
– Given data only, infer not only the underlying Bayesian Network but the causality between variables
IBM Research
© 2006 IBM Corporation37
Parameter Estimation (for Linear Gaussian Models)
Parameter estimation given graph structure reduces to standard estimation problem (e.g. maximum likelihood estimation) for the underlying model class
For example, for linear Gaussian models, it is solvable by linear regression
– P(M,E,C,R) = P(E) P(M|E) P(C|E) P(R|M,C)
– M = 1E + 1, 1 ~ N(0, 1)
– C = 2E + 2, 2 ~ N(0, 2)
– R = 3M + 4C + 3, 3 ~ N(0, 3)
There is active research for many other underlying model classes
Economy
Marketing Competition
Revenue
IBM Research
© 2006 IBM Corporation38
Inference in a Given Model
Given an estimated model and realized values for a subset of the variables, it may be possible to compute the most likely values for unknown variables.
Inference in unrestricted Bayesian Networks is intractable (#P-hard)
For restricted classes, it is possible to efficiently perform inference: e.g. dependency trees
– P(M,E,C,R) = P(E) P(M|E) P(C|E) P(R|M)
– P(M|E,C,R) = P(M|E,C) P(R|M)
– = P(M|E) P(R|M)
– Simplified due to the conditional independence (d-separation) between M and C, implied by the graph structure
But there are considerable challenges for graph structures including undirected cycles (e.g. original graph for P(M,E,C,R))
Economy
Marketing Competition
Revenue
IBM Research
© 2006 IBM Corporation39
Structure Learning Given Data
For unrestricted classes, structure learning is known to be intractable.
– Even for the class of poly-trees, robust estimation (i.e. when the true distribution may not be in the target class) is NP-hard
For some restricted classes, structure learning is efficient
– Dependency Trees can be efficiently robustly learned
– Poly-trees can be efficiently learned, given the assumption that the true model is in the target class
There is active research on proving conditions under which variable selection methods in regression (e.g. Lasso) can provably learn structure of general graphs
Economy
Marketing Competition
Revenue
Economy
Marketing Competition
Revenue
A dependency tree A poly-tree
IBM Research
© 2006 IBM Corporation40
Inferring Causality from Data
Economy
Marketing Revenue
Marketing Competition
Revenue
P(M,E,R) = P(E) P(M|E) P(R|E)
P(M,E,C) = P(M) P(C) P(R|M,C)
M ┴ C | R
M ┴ R | E
Economy
Marketing Revenue
Economy
Marketing Revenue
The causal structure cannot be determined from data !
P(M,E,R) = P(M) P(E|M) P(R|E) P(M,E,R) = P(R) P(E|R) P(M|E)
The causal structure can be determined from data !
It can be inferred that Marketing can be
a “lever” for controlling Revenue !
C.f. [P. Spirtes, C. Glymour, and R. Scheines (2000)]
IBM Research
© 2006 IBM Corporation41
Summary: Estimation and Inference with Bayesian Networks
Parameter estimation from data given structure
– It is efficiently solvable for many model classes
Inference given model
– Exact inference is known to be NP-complete for sub-class including undirected cycles
– It is efficiently solvable for tree structures and many models used in practice
Latent variable estimation, given structure
– Local optimum estimation is often possible via EM-algorithms
Bayesian network structure learning from data
– It is known to be “intractable” for general classes
– It is even NP-complete to estimate “polytrees” robustly
Inferring causal structure from data
– Sometimes possible but in general not
Given these facts, determining network structure using domain knowledge and using it to do parameter estimation and inference is common practice
example
IBM Research
© 2006 IBM Corporation42
Tutorial outline
Challenges of marketing analytics
– Integrating the “marketing” and “data mining” approaches
– Customizing data mining approaches to the challenges of marketing decision support
Survey of some useful ML methodologies
– Bayesian network modeling
– Utility-based classification (Cost-sensitive learning)
– Reinforcement learning and Markov decision processes
Detailed analysis and case studies:
– Customer lifetime value modeling
– Customer wallet estimation
IBM Research
© 2006 IBM Corporation43
Cost-sensitive Learning for Marketing Decision Support
Use of Basic Machine Learning (e.g. Classification and Regression) in Marketing Decision Support is well accepted
– Example applications include: targeted marketing, credit rating, and others
– But are they the best we have to offer ?
Regression is an inherently harder problem than is required
– One does not necessarily need to predict business outcome, customer behavior, etc, but is merely required to make business decisions
– Regression may fail to detect significant patterns, especially when data is noisy
Classification is an over simplification
– By mapping to classification, one loses information on the degree of goodness/badness of a business decision in the past data
Cost-sensitive classification provides the desired middle ground
– It simplifies the problem almost to classification and thus allows discovery of significant patterns;
– Yet retains and exploits the information on the degree of goodness of business decisions, in a way that is motivated by Utility theory
IBM Research
© 2006 IBM Corporation44
Cost-sensitive Learning a.k.a. Utility-based Classification
In regression: given (x,r) ε X x R, generated from a sampling distribution, find F: – F(x) ≈ r – E.g. r = profit obtained by targeting customer x
In classification: given (x,y) ε X x {0,1} , generated from a sampling distribution, find F: – F(x) ≈ y – E.g. y = 1 if customer x is “good”, 0 otherwise
In utility-based classification: given (stochastic) utility function U and (x,y) ε X x {0,1} generated from a sampling distribution, find F: – E[U(x,y,F(x))] is maximized (or equivalently E[C(x,y,F(x))] is minimized)
– E.g. U(x,1,1) = Profit(x) = Profit obtained by targeting customer x, when x is indeed a “good” customer.
IBM Research
© 2006 IBM Corporation45
Example Cost and Utility Functions
Simple formulation (cost/benefit matrix)
More realistic formulation (utility/cost dependent on individuals)
Predicted
True
0 1
0 1 0
1 0 1
Classification utility matrix
“Credit rating” utility
Predicted
True
bad good
bad 0 - Default Amt
good 0 Interest
“Targeted marketing” utility
Predicted
True
bad good
bad 0 - C
good 0 Profit – C
Predicted
True
0 1
0 0 1
1 1 0
Misclassification cost matrix
IBM Research
© 2006 IBM Corporation46
Bayesian Approach with Regression
For each example x, choose the class that minimizes the expected cost:
Problem: Requires conditional density estimation and regression to solve a classification problem.
– Price is high computational and sample complexity Merit: more flexibility and general applicability
– Business constraints
– Variability in fixed costs
– But, is it necessary ?
ji
jixCxjPxi ),,()|(minarg)(*
need be estimated!
IBM Research
© 2006 IBM Corporation47
A Classification Approach: Reducing cost-sensitive learning to weighted classification via “Costing” [ZLA’03]
• If Y is {0,1}, then minimizing cost is equivalent to minimizing
where
• Even though we have a 2 x 2 cost matrix, its minimization can be done using one weight per labeled example
• Given a distributional assumption on w(x,y), minimizing the above weighted error in the training data will generalize to unseen test instances !
• “Costing” Algorithm
• “Costing” repeatedly performs weighted rejection sampling with w(x,y) to obtain an ensemble of hypotheses
• Similar approaches have been applied to class probability estimation (“probing” [LZ’05) and quantile regression (“quanting” [LOZ’06])
)],())(([ yxwyxhIE YX
),()1,(),( yxCyxCyxw
example
IBM Research
© 2006 IBM Corporation48
Exceptional
Performance
Empirical Evaluation with Targeted Marketing data sets
Method Costing
(200)
Transparent Box
Resampling (100k)
No weight
NB $13163 $12367 $12026 $0.24
Boosted NB $14714 $14489 $13135 -$1.36
C4.5 $15016 -$118 $2259 $0
SVMLight $13152 $13683 $12808 $0
KDD-98:Charity donation
DMEF-2:Targeted marketing
Method Costing
(200)
Transparent Box
Resampling (100k)
No weight
NB $37629 $32608 $12026 $16462
Boosted NB $37891 $36381 $13135 $121
C4.5 $37500 $478 $2259 $0
SVMLight $35290 $36443 $12808 $0
Rejection Sampling Feeding weights Classification
*Costing is state-of-the-art, but is restricted to 2-class problems
Resampling
Test Set Profits
IBM Research
© 2006 IBM Corporation49
A Closer Look: “Costing” (cost based bagging) [ZLA’03]
Costing (Learner A, Data S, count T)
(1) For all set
(2) For t=1 to T do
(1) Let ht = A(S, w)
(3) Output H:
S y)(x,
),()1,(, yxCyxCw yx
yxh
y
t
xH)(
1maxarg)(
Same weight in every iteration
It only makes sense for 2-class
IBM Research
© 2006 IBM Corporation50
A Multi-class Extension: Cost-sensitive boosting algorithm [AZL 2004]
T
tii yxh
1
),(
),()],([ 0,y x, yxCyxCEw HyS
||/1),(0 YyxH
}'),(|))0(),,{((' yx, SyxwIyxT
S' y)(x,
S' y)(x,
),()],([ 1,y x, yxCyxCEw tHyS
GBSE (Learner A, Expanded data S’, count T)
(1) For all initialize
(2) For all initialize weight
(3) For t=1 to T do
(a) For all (x,y) in S’ update weight
(b) Let
(c) Let ht = A(T,|w|)
(d) ft = Stochastic(hi )
(e) Ft = (1- α)Ft-1+αft
(4) Output h(x) = arg max( )
Weight updated in each iteration
the difference between average cost
by the current ensemble and cost of y
Y}y' S,y)(x,y |)y'{(x,S' y)(x, Define the “expanded sample” S’ as:
IBM Research
© 2006 IBM Corporation51
Gradient Boosting with Stochastic Ensembles: Illustration
C(x,y)
At learning iteration t At learning iteration t+1
+
Cost C(x,y)
Predicted
Label, y+ - - + -
Training Labels
• The difference between the current average cost and the cost associated with a particular label is the boosting weight
• The sign of the weight, E[C(x,y)]– C(x,y), is the training label
Ave Cost
E[C(x,y)]
y
IBM Research
© 2006 IBM Corporation52
Cost-sensitive boosting outperforms existing methods of cost-sensitive learning as well as classification and regression
Existing methods
Data Set Bagging AvgCost MetaCost GBSE
Annealing 1059±174 127±12 207±42 34±4
Solar 5403±397 237±38 5317±390 48±10
KDD-99 319±42 42±8 49±9 2±1
Letter 151±3 92±1 130±2 85±2
Splice 64±5 61±4 50±3 58±4
Satellite 190±10 108±6 104±6 93±6
Ave Test Set Cost (±SE)
IBM Research
© 2006 IBM Corporation53
Tutorial outline
Challenges of marketing analytics
– Integrating the “marketing” and “data mining” approaches
– Customizing data mining approaches to the challenges of marketing decision support
Survey of some useful ML methodologies
– Utility-based classification (Cost-sensitive learning)
– Reinforcement learning and Markov decision processes
– Bayesian network modeling
Detailed analysis and case studies:
– Customer lifetime value modeling
– Customer wallet estimation
IBM Research
© 2006 IBM Corporation54
Sequential Cost-sensitive Decision Making by Reinforcement Learning
Cost-sensitive classification provides an adequate framework for single marketing decision making
– Real world marketing decision making is rarely made in isolation, but is made sequentially
– Need to address the sequential dependency in decision making
Cost-sensitive classification
– Maximizes E[U(x,h(x)]
We now wish to
– Maximize Σt E[U(xt,h(xt)], where x may depend on earlier decisions …
This is nothing but Reinforcement Learning, if we view x as the “state”
– Maximize Σt E[U(st,π(st))], where st is determined stochastically according to a transition probability determined by st-1 and π(st-1).
IBM Research
© 2006 IBM Corporation55
Review: Markov Decision Process (MDP)
At any given time t, the agent is in some state s.
It takes an action a, and makes a transition to the next state s’, dictated by transition probability T(s,a)
It then receives a “reward”, or utility U(s,a), which also depends on state s and action a.
The goal of a reinforcement learner in MDP is to learn a policy, namely π: S → A, mapping states to actions, so as to maximize the cumulative discounted reward:
),( R0t
ttt asU
IBM Research
© 2006 IBM Corporation56
Modeling CRM process using Markov Decision Process (MDP) Customer is in some "state" (his/her attributes) at any point in time
Retailer's action will move customer into another state
Retailer's goal is to take sequence of actions to guide customer's path to maximize customer's lifetime value
Reinforcement Learning produces optimized targeting rules of the form If customer is in state "s", then take marketing action "a"
Customer state “s” represented by current customer attribute vector
estimates LTV(s,a) -- best policy is to choose a to maximize LTV(s,a)
Typical CRM Process
BargainHunter
Repeater
LoyalCustomer
ValuableCustomer
One Timer
Repeater
Defector Defector
Repeater
LoyalCustomer
PotentiallyValuable
Campaign A
Campaign B
Campaign C
Campaign E
Campaign D
MDP and Reinforcement Learning provide an advanced framework for modeling customer lifetime value
p 64
IBM Research
© 2006 IBM Corporation57
Observed lifetime value reflects only customer’s lifetime value attained by current marketing policy, and therefore fails to capture their potential lifetime value
MDP based lifetime value modeling allows modeling of lifetime value based on optimized marketing policy (= the output of system !)
BargainHunter
Repeater
LoyalCustomer
ValuableCustomer
One Timer
Repeater
Defector Defector
Repeater
LoyalCustomer
PotentiallyValuable
Campaign A
Campaign B
Campaign C
Campaign E
Campaign D
Current marketing policy
Optimized marketing policy
• Estimated (potential) lifetime value will be based on the optimal path
• Output policy will lead the customer through the same path
MDP enables genuine lifetime value modeling, in contrast to existing approaches that use observed lifetime value
Customer A’s path under…
IBM Research
© 2006 IBM Corporation58
And here is how this is possible…
The MDP enables the use of data for many customers in various stages (states) to determine potential lifetime value of a particular customer in a particular state
Reinforcement Learning can estimate the lifetime value (function) without explicitly estimating the MDP itself
– The key lies in the value iteration procedure based on “Bellman’s equation”
Repeater
LoyalCustomer
ValuableCustomer
Repeater Repeater
LoyalCustomer
PotentiallyValuable
Each rule is, in effect, trained with data corresponding to all subsequent states
LTV of a state = reward now + LTV of best next state
Rule a Rule b
Rule c
Rule d
)a',Q(s'maxa)]E[U(s, a)Q(s, 'a
IBM Research
© 2006 IBM Corporation59
Reinforcement Learning Methods with Function Approximation
Value Iteration (based on Bellman Equation)
– Provides the basis for classic reinforcement learning methods like Q-learning
Batch Q-Learning (with Function Approximation)
– Solves value iteration as iterative regression problems
a)(s,Qmax arg (s)
)a',(s'Qmaxa)]E[U(s, a)(s,Q
]a)E[U(s, a)(s,Q
a
k'1k
0
a
))a',(s'Qmax),((a)(s,)Q-(1 a)(s,Q
a)U(s,a)(s,Q
k'k1k
0
aasU
Estimate using function approximation (regression)
IBM Research
© 2006 IBM Corporation60
The graph below plots profits per campaign obtained in monthly campaigns over 2 years (in an empirical evaluation using benchmark data, i.e. KDD cup 98 data)
Lifetime value modeling based on reinforcement learning can achieve greater long term profits than the traditional approach
… to yield greater long term profits
Output policy of MDP approach (CCOM) “invests” in initial campaigns…
Output policy of MDP approach (CCOM) “invests” in initial campaigns…
IBM Research
© 2006 IBM Corporation61
Tutorial outline
Challenges of marketing analytics
– Integrating the “marketing” and “data mining” approaches
– Customizing data mining approaches to the challenges of marketing decision support
Survey of some useful ML methodologies
– Utility-based classification (Cost-sensitive and active learning)
– Reinforcement learning and Markov decision processes
– Bayesian network modeling
Detailed analysis and case studies:
– Customer lifetime value modeling
– Customer wallet estimation
IBM Research
© 2006 IBM Corporation62
Lifetime Value Modeling and Cross-Channel Optimized Marketing (CCOM)
Direct Mail
Kiosk
Web
Store
Call Center
$
$ $ $ $
Optimizes targeted marketing across multiple channels for lifetime value maximization.
Combines scalable data mining and reinforcement learning methods to realize unique capability.
IBM Research
© 2006 IBM Corporation63
CCOM Pilot Project with Saks Fifth Avenue
Business Problem addressed: Optimizing direct mailing to maximize lifetime revenue at the store (and other channels)
Provided solution for the “Cross-Channel Challenge”: No explicit linking between marketing actions in one channel and revenue in another
CCOM mailing policy shown to achieve 7-8% increase in expected revenue in the store (in laboratory experiments) !
Direct Mail
Store
$ $ $ $
$
CCOM-pilot business problem
p56
IBM Research
© 2006 IBM Corporation64
Some Example Features
Demographic Features action reward
FULL_LINE_STORE_OF_RES.: If a full-line store exists in the area 0.018 0.004
NON_FL_STORE_OF_RES.: If a non full-line store exists in area 0.012 -0.004
Transaction Features (concerning divisions relevant to current campaign)
CUR_DIV_PURCHASE_AMT_1M: Pur amt in last month in curr div 0.065 0.090
CUR_DIV_PURCHASE_AMT_2_3M: Pur amt in 2-3 month in curr div 0.099 0.080
CUR_DIV_PURCHASE_AMT_4_6M: Pur amt in 4-6 month in curr div 0.133 0.091
CUR_DIV_PURCHASE_AMT_1Y: Pur amt in last year in curr div 0.162 0.128
CUR_DIV_PURCHASE_AMT_TOT: Total Pur amt in current division 0.153 0.147
Promotion History Features (on divisions relevant to current campaign)
CUR_DIV_N_CATS_1M: Num cat sent last month in curr div 0.294 0.028
CUR_DIV_N_CATS_2_3M: Num cat sent 2-3 months ago in curr div 0.260 0.025
CUR_DIV_N_CATS_4_6M: Num cat sent 4-6 months ago in curr div 0.158 0.062
CUR_DIV_N_CATS_TOT: Total num cat sent in curr div to date 0.254 0.062
Control Variable
ACTION: To mail or not to mail 1.000 0.008
Target (Response) Variable
REWARD: Expected cumulative profits 0.008 1.000
IBM Research
© 2006 IBM Corporation65
The Cross-Channel Challenge and Solution
The Challenge: No explicit linking between actions in one channel (mailing) and rewards in another (revenue)
• Very low correlation observed between actions and responses• Other factors determining “life time value” may dominate over the control variable
(marketing action) in estimation of “expected value” • Obtained models can be independent of the action and give rise to useless rules !
The Cross-Channel Solution: Learn the relative advantage of competing actions!
– Standard Method
– Proposed Method
Actions
Value in state s1 Value in state s2
Actionsa1 a2 a1 a2
Value in state s1 Value in state s2
Actionsa1 a2 a1 a2
Value in state s1 Value in state s2
Actions
a1 a2 a1 a2
Approximation
IBM Research
© 2006 IBM Corporation66
The Learning Method
Definition of Advantage
– A(s,a):= 1/Δt(Q(s,a) – maxa’ Q(s,a’))
Advantage Updating Procedure [Baird ’94]
– Modifications: 1. Initialization with empirical life time value
– 2. Batch Learning with optional function approximation
Repeat 1. Learn 1.1. A(s,a):=(1-α)A(s,a)
+α (Amax(s)+(R(s,a)+γΔtV(s’)-V(s))/Δt) 1.2. Use Regression to estimate A(s,a) 1.3. V(s):=(1-β)V(s)
+β(V(s)+(Amax-new(s)-Amax-old(s))/α) 2. Normalize
A(s,a):=(1- ω)A(s,a)+ω(A(s,a)-Amax(s))
IBM Research
© 2006 IBM Corporation67
Evaluation Results
Significant policy advantage observed with small number of iterations
Obtained policy with 7- 8% policy advantage, i.e. 7- 8% increase in expected revenue (for 1.6 million customers considered)
Mailing policy was constrained to mail same number of catalogues in each campaign as last year
CCOM to evaluate sequence of models and output best model
Policy Advantage
-4
-2
0
2
4
6
8
10
1 2 3 4 5
Learning iterations
Ad
va
nta
ge
(p
erc
en
tag
e)
Policy Advantage
-4
-2
0
2
4
6
8
1 2 3 4 5
Learning iterations
Ad
va
nta
ge
(p
erc
en
t)
Typical run (version 1)
Typical run (version 2)
IBM Research
© 2006 IBM Corporation68
Evaluation Method
Challenge in Evaluation: Need to evaluate new policy using data collected by existing (sampling) policy
Solution: Use bias-corrected estimation of “policy advantage” using data collected by sampling policy
Definition of policy advantage:
– (Discrete Time) Advantage
– Policy Advantage
Estimating policy advantage with bias corrected sampling
Aπ(s,a):= Qπ (s,a) – maxa’ Qπ (s,a’)
As~π(π’):= Eπ [Ea~π’ [Aπ(s,a)]]
As~π(π’):= Eπ [(π’(a|s)/ π(a|s)) [Aπ(s,a)]]
IBM Research
© 2006 IBM Corporation69
Combination of reinforcement learning (MDP) with predictive data mining enables automatic generation of trigger-based marketing targeting rules
Optimized with respect to the customer’s potential lifetime value
Stated in simple “if then” style, which supports flexibility and compatibility
Refined to make reference to detailed customer attributes and hence, well-suited to event and trigger-based marketing
This is made possible by
– Representing the states in MDP by customer’s attribute vectors
– Combining reinforcement learning with predictive data mining to estimate lifetime value as function of customer attributes and marketing actions
An example marketing targeting ruleoutput by CCOM system
IBM Research
© 2006 IBM Corporation70
Some examples of rules output by CCOM
Interpretation: If a customer has spent in the current division but enough catalogues have been sent, then don’t mail
• Avoid saturation effects
• Differentiate between customers who may be near saturation and those who are not … Interpretation: If a customer has spent in the current division and has received moderately many relevant catalogues, then mail
• Invest in a customer until it knows it is not worth it
Interpretation: If a customer has spent significantly in the past and yet has not spent much in the current division (product group) then don’t mail
IBM Research
© 2006 IBM Corporation71
Marketing Event
Event IdentifierChannel IdentifierEvent DateEvent Category DescriptionFixed Cost
Customer
Customer IdentifierFirst NameLast NameAgeGender
Transaction
Customer IdentifierTransaction DateProduct Category IdentifierEvent IdentifierChannel IdentifierTransaction RevenueTransaction Profit
Customer Marketing Action
Event IdentifierCustomer IdentifierMarketing Action DateMarketing Action
Period
Period IdentifierPeriod Duration
Customer Profile History
Customer IdentifierProfile History DatePeriod IdentifierProduct Category IdentifierChannel IdentifierAggregated Count of EventAggregated RevenueAggegated Profit
Channel
Channel IdentifierChannel Description
Product Category
Product Category IdentifierProduct Category Description
Customer Loyalty Level History
Customer IdentifierLoyalty Level Start DateLoyalty Level End DateLoyalty Level
EventProduct Category
Event IdentifierProduct Category IdentifierWeight
CCOM Output Models
Marketing Policy Model
Model IdentifierModel TypeModel
Lifetime Value Model
Model IdentifierModel TypeModel
CCOM - Logical Data Model
Optional Entity
CCOM is generically applicable by mapping physical data to this model
*Developed with CBO
IBM Research
© 2006 IBM Corporation72
Tutorial outline
Challenges of marketing analytics
– Integrating the “marketing” and “data mining” approaches
– Customizing data mining approaches to the challenges of marketing decision support
Survey of some useful ML methodologies
– Bayesian network modeling
– Utility-based classification (Cost-sensitive learning)
– Reinforcement learning and Markov decision processes
Detailed analysis and case studies:
– Customer lifetime value modeling
– Customer wallet estimation
IBM Research
© 2006 IBM Corporation73
Wallet Estimation Case Study Outline Introduction
– Business motivation and different wallet definitions
Modeling approaches for conditional quantile estimation
– Local and global models
– Empirical evaluation
A graphical model approach to wallet estimation
– Generic algorithm for class of latent variable modeling problems
MAP (Market Alignment Program)
– Description of application and goals
– The interview process and the feedback loop
– Evaluation of Wallet models performance in MAP
IBM Research
© 2006 IBM Corporation74
What is Wallet (AKA Opportunity)?
Total amount of money a company can spend on a certain category of products.
Company Revenue
IT Wallet
IBM Sales
IBM sales IT wallet Company revenue
IBM Research
© 2006 IBM Corporation75
Why Are We Interested in Wallet?
Customer targeting
– Focus on acquiring customers with high wallet
– Evaluate customers’ growth potential by combining wallet estimates and sales history
– For existing customers, focus on high wallet, low share-of-wallet customers
Sales force management
– Make resource assignment decisions
• Concentrate resources on untapped
– Evaluate success of sales personnel and sales channel by share-of-wallet they attain
OnT
argetM
AP
IBM Research
© 2006 IBM Corporation76
Wallet Modeling Problem
Given:
– customer firmographics x (from D&B): industry, emloyee number, company type etc.
– customer revenue r
– IBM relationship variables z: historical sales by product
– IBM sales s
Goal: model customer wallet w, then use it to “predict” present/future wallets
No direct training data on w or information about its distribution!
IBM Research
© 2006 IBM Corporation77
Historical Approaches within IBM Top down: this is the approach used by IBM
Market Intelligence in North America (called ITEM)
– Use econometric models to assign total “opportunity” to segment (e.g., industry geography)
– Assign to companies in segment proportional to their size (e.g., D&B employee counts)
Bottom up: learn a model for individual companies
– Get “true” wallet values through surveys or appropriate data repositories (exist e.g. for credit cards)
Many issues with both approaches (won’t go into detail)
– We would like a predictive approach from raw data
IBM Research
© 2006 IBM Corporation78
Relevant Work in the Literature
While wallet (or share of wallet) is widely recognized as important, not much work on estimating it:
Du, Kamakura and Mela (2006) developed “list augmentation” approach, using survey data to model spending with competitors
Epsilon Data Management in white paper in 2001, proposed survey-based methodology
Zadrozny, Costa and Kamakura (2005) compared bottom-up and top-down approaches on IBM data. Evaluation is based on a survey.
IBM Research
© 2006 IBM Corporation79
Traditional Approaches to Model Evaluation
Evaluate models based on surveys
– Cost and reliability issues
Evaluate models based on high-level performance indicators:
– Do the wallet numbers sum up to numbers that “make sense” at segment level (e.g., compared to macro-economic models)?
– Does the distribution of differences between predicted Wallet and actual IBM Sales and/or Company Revenue make sense? In particular, are the % we expect bigger/smaller?
– Problem: no observation-level evaluation
IBM Research
© 2006 IBM Corporation80
Proposed Hierarchical IT Wallet Definitions TOTAL: Total customer available IT budget
– Probably not quantity we want (IBM cannot sell it all)
SERVED: Total customer spending on IT products covered by IBM
– Share of wallet is portion of this number spent with IBM?
REALISTIC: IBM sales to “best similar customers”
– This can be concretely defined as a high percentile of:P(IBM revenue | customer attributes)
– Fits typical definition of opportunity?
REALISTIC SERVED TOTAL
TOTAL
SERVED
REALISTIC
IBM Research
© 2006 IBM Corporation81
An Approach to Estimating SERVED Wallets
Wallet is unobserved, all other variables are
Two families of variables --- firmographics and IBM relationship are conditionally independent given wallet
We develop inference procedures and demonstrate them
Theoretically attractive, practically questionable
(We will come back to this later)
Company
firmographics
IT spendwith IBM
Historical relationship
with IBM
SERVEDWallet
IBM Research
© 2006 IBM Corporation82
Distribution of IBM sales to the customer given customer attributes: s|r,x,z ~ f,r,x,z
E.g., the standard linear regression assumption:
What we are looking for is the pth percentile of this distribution
REALISTIC Wallet: Percentile of Conditional
),(~,,| 2 zrxNzrxs
E(s|r,x,z) REALISTIC
IBM Research
© 2006 IBM Corporation83
Estimating Conditional Distributions and Quantiles Assume for now we know which percentile p we
are looking for
First observe that modeling well the complete conditional distribution P(s|r,x,z) is sufficientIf have good parametric model and distribution
assumptions can also use it to estimate quantiles
– E.g.: linear regression under linear model and homoskedastic iid gaussian errors assumptions
Practically, however, may not be good idea to count on such assumptions
– Especially not a gaussian model, because of statistical robustness considerations
IBM Research
© 2006 IBM Corporation84
Modeling REALISTIC Wallet Directly
REALISTIC defines wallet as pth percentile of conditional of spending given customer attributes
– Implies some (1-p)% of the customers are spending full wallet with IBM
Two obvious ways to get at the pth percentile:
– Estimate the conditional by integrating over a neighborhood of similar customers Take pth percentile of spending in neighborhood
– Create a global model for pth percentile Build global regression models
IBM Research
© 2006 IBM Corporation85
Local Models: K-Nearest Neighbors Design distance metric, e.g.:
– Same industry
– Similar employees/revenue
– Similar IBM relationship
Neighborhood sizes (k):
– Neighborhood size has significant effect on prediction quality
Prediction:
– Quantile of firms in the neighborhood
Indu
stry
Employees IBM
spe
nd
Universe of IBM customers with D&B information
Neighborhood of target company
Target company i
Fre
qu
en
cy
IBM Sales
Wallet Estimate
IBM Research
© 2006 IBM Corporation86
Global Estimation: the Quantile Loss Function Our REALISTIC wallet definition calls for estimating the
pth quantile of P(s|x,z).
Can we devise a loss function which correctly estimates the quantile on average?Answer: yes, the quantile loss function for quantile p.
This loss function is optimized in expectation when we correctly predict REALISTIC:
yyyyp
yyyypyyLp ˆ if )ˆ()1(
ˆ if )ˆ()ˆ,(
)|( of quantile p)|)ˆ,((minarg thˆ xyPxyyLE py
IBM Research
© 2006 IBM Corporation87
-3 -2 -1 0 1 2 3
01
23
4
Some Quantile Loss Functions
p=0.8
p=0.5 (absolute loss)
Residual (observed-predicted)
IBM Research
© 2006 IBM Corporation88
Quantile Regression Squared loss regression:
– Estimation of conditional expected value by minimizing sum of squares
Quantile regression:
– Minimize Quantile loss:
Implementation:
– assume linear function in some representation y = t f(x,z), solution using linear programming
– Linear quantile regression package in R (Koenker, 2001)
n
iiiip xzfsL
1
)),,(,(min
quantile regression
loss function
n
iiii xzfs
1
2)),,((min
yyyyp
yyyypyyLp ˆ if )ˆ()1(
ˆ if )ˆ()ˆ,(
IBM Research
© 2006 IBM Corporation89
Quantile Regression Tree – Local or Global? Motivation:
– Identify a locally optimal definition of neighborhood
– Inherently nonlinear
Adjustments of M5/CART for Quantile prediction:
– Predict the quantile rather than the mean of the leaf
– Empirically, splitting/pruning criteria do not require adjustment
Industry = ‘Banking’
Sales<100K
IBM Rev 2003>10K
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
no
yes
no
no
yes
yes
Industry = ‘Banking’
Sales<100K
IBM Rev 2003>10K
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
Fre
qu
ency
IBM Sales
Wallet Estimate
no
yes
no
no
yes
yes
IBM Research
© 2006 IBM Corporation90
Aside: Log-Scale Modeling of Monetary Quantities Due to exponential, very long tailed typical
distribution of monetary quantities (like Sales and Wallet), it is typically impossible to model them on original scale, because e.g.:
– Biggest companies dominate modeling and evaluation
– Any implicit homoskedasticity assumption in using fixed loss function is invalid
Log scale is often statistically appropriate, for example if % change is likely to be “homoskedastic”
Major issue: models ultimately judged in dollars, not log-dollars…
IBM Research
© 2006 IBM Corporation91
Empirical Evaluation: Quantile Loss
Setup
– Four domains with relevant quantile modeling problems:direct mailing, housing prices, income data, IBM sales
– Performance on test set in terms of 0.9th quantile loss
– Approaches: Linear quantile regression, Q-kNN, Quantile trees, Bagged quantile trees, Quanting (Langrofd et al. 2006 -- reduces quantile estimation to averaged classification using trees)
Baselines
– Best constant model
– Traditional regression models for expected values, adjusted under Gaussian assumption (+1.28)
IBM Research
© 2006 IBM Corporation92
Performance on Quantile Loss
Conclusions
– Standard regression models are not competitive
– If there is a time-lagged variable, LinQuantReg is best
– Otherwise, bagged quantile trees (and quanting) perform best
– Q-kNN is not competitive
IBM Research
© 2006 IBM Corporation93
Residuals for Quantile Regression
Total positive holdout residuals: 90.05% (18009/20000)
IBM Research
© 2006 IBM Corporation94
Graphical Model for SERVED(?) Wallet Estimation
Customer’s Firmographics (X)
Customer’s IT Wallet (W)
Customer’s Spendingwith IBM (S)
Customer’s Relationshipwith IBM (Z)
View 1
View 2
Two conditionally independent views !
IBM Research
© 2006 IBM Corporation95
Generic Problem Setting
Unsupervised learning scenario:
Unobserved target variable
Observations on multiple predictor variables
Domain knowledge suggesting that the predictors form multiple conditionally independent views
Goal: To predict the target variable
IBM Research
© 2006 IBM Corporation96
Summary of Results on Generic Problem Analysis of a relevant class of latent variable
models– Markov blanket can be split into conditionally independent
views
– For exponential linear models, the maximum likelihood estimation reduces to convex optimization problem
Solution approaches for Gaussian likelihoods– Reduction to single linear least squares regression
– ANOVA for testing conditional independence assumptions
Empirical evaluation– Comparable to supervised learning with significant amount of
training data
– Case study on wallet estimation
IBM Research
© 2006 IBM Corporation97
Discriminative Maximum Likelihood Inference Given: Directed graphical model and parametric form of the
conditional distributions of nodes given their parents
Goal: Predict the target W using the parameter estimates that are most likely given the observed data and the graphical model:
Where = (0, 1) is the parameter vector for the parametric conditional likelihoods, and D is our data
Solution: Expectation-Maximization (EM) algorithm
– Converges to a local optimum in general
Estimating W: Mean or mode of “posterior”
dwww Z),|(SPX)|(Plogmax Z)X, | (SP logmax10D,
*
Z)X,|W(P *
IBM Research
© 2006 IBM Corporation98
General Theoretical Result: Exponential Models
Theorem: When the conditional distributions p(W|X) and p(S|W,Z), correspond to exponential linear models with matching link functions, the incomplete discriminative log-likelihood: LD() = log PD,(S|X,Z)is a concave function of the parameters
Maximum likelihood estimation reduces to a convex optimization problem
EM algorithm converges to the globally optimal solution
IBM Research
© 2006 IBM Corporation99
Gaussian Likelihoods and Linear Regression Assume both discriminative likelihoods P(W|X)
and P(S|W,Z) are linear and gaussian:
wi - txi = iw ~ N(0, w2) i.i.d
si - wi - tzi = is ~ N(0, s2) i.i.d
Previous theorem says that EM would give ML solution MLE= (MLE, MLE)
But if we add equations up we eliminate W:
si - txi - tzi = (is+ iw) ~ N(0, s2+ w
2) i.i.d
Maximum likelihood solution of this problem is linear regression and gives solution LS= (LS, LS)
– Are the two solutions the same?
IBM Research
© 2006 IBM Corporation100
Equivalence and Interpretation
Equivalence Theorem: When U=[X,Z] is a full column rank matrix, the two estimates are identical: MLE =LS
Consistency of LS and unbiasedness of resulting W estimates
Can make use of linear regression computation and inference toolsIn particular: ANOVA to test validity of assumptions
Some caveats we glossed over
– In particular, full rank requirement implies cannot have intercept in both gaussian likelihoods!
IBM Research
© 2006 IBM Corporation101
ANOVA for Testing Independence Assumptions
ANOVA: Variance-based analysis for determining the goodness of fit for nested linear models
Example of nested models:
– Model A: Linear model with only variables in X, Z and no interactions
– Model B: Allow interactions only within X and Z
– Model C: Allow interactions between variables in X and Z
Key Idea: if model C is statistically superior to model B conditional independence and/or parametric assumptions are rejected
IBM Research
© 2006 IBM Corporation102
Some Simulation Results
Z
z
IBM Research
© 2006 IBM Corporation103
Wallet Case Study Results
Modeling equations: (monetary values log scale)
log(wi) = f(xi) + cw + iw, iw~ N(0, σ2)
log(si) − log(wi) = g(zi) + cs + is, is~ N(0, σ2)
(cw, cs are intercepts, f, g are parametric forms)
Data is 2000 IBM customers in finance sector
ANOVA results consistent with cond. independence:
IBM Research
© 2006 IBM Corporation104
Market Alignment Project (MAP): Background
MAP - Objective:
– Optimize the allocation of sales force
– Focus on customers with growth potential
– Set evaluation baselines for sales personal
MAP – Components:
– Web-interface with customer information
– Analytical component: wallet estimates
– Workshops with Sales personal to review and correct the wallet predictions
– Shift of resources towards customers with lower wallet share
IBM Research
© 2006 IBM Corporation105
MAP Tool Captures Expert Feedback from the Client Facing Teams
Transaction Data
D&BData
Wallet models: Predicted
Opportunity
ResourceAssignments
Expert validated
Opportunity
Analytics and Validation
Data Integration
Insight Delivery and Capture
Post-processing
MAP Interview Team Client Facing Unit (CFU) Team
Web Interface
MAP interview process – all Integrated and Aligned Coverages
The objective here is to use expert feedback (i.e. validated revenue opportunity) from from last year’s workshops to evaluate our latest opportunity models
IBM Research
© 2006 IBM Corporation106
MAP Workshops Overview Calculated 2005 opportunity using naive Q-kNN
approach
2005 MAP workshops
– Displayed opportunity by brand
– Expert can accept or alter the opportunity
Select 3 brands for evaluation: DB2, Rational, Tivoli
Build ~100 models for each brand using different approaches
Compare expert opportunity to model predictions
– Error measures: absolute, squared
– Scale: original, log, root
IBM Research
© 2006 IBM Corporation107
Initial Q-kNN Model Used
Distance metric
– Identical Industry
– Euclidean distance on size (Revenue or employees)
Neighborhood sizes 20
Prediction
– Median of the non-zero neighbors
– (Alternatives Max, Percentile)
Post-Processing
– Floor prediction by max of last 3 years revenue
Indu
stry
Employees Reven
ue
Universe of IBM customers with D&B information
Neighborhood of target company
Target company i
IBM Research
© 2006 IBM Corporation108
0
2
4
6
8
10
12
14
16
18
20
0 2 4 6 8 10 12 14 16 18 20
Expert Feedback
MODEL_OPPTY
Expert Feedback (Log Scale) to Original Model (DB2)
Experts reduce opportunity to 0(15%)
Experts acceptopportunity (45%)
Experts changeopportunity (40%)
Increase (17%)
Decrease (23%)
IBM Research
© 2006 IBM Corporation109
Observations
Many accounts are set for external reasons to zero
– Exclude from evaluation since no model can predict this
Exponential distribution of opportunities
– Evaluation on the original (non-log) scale suffers from huge outliers
Experts seem to make percentage adjustments
– Consider log scale evaluation in addition to original scale and root as intermediate
– Suspect strong “anchoring” bias, 45% of opportunities were not touched
IBM Research
© 2006 IBM Corporation110
Evaluation Measures
Different scales to avoid outlier artifacts
– Original: e = model - expert
– Root: e = root(model) - root(expert)
– Log: e = log(model) - log(expert)
Statistics on the distribution of the errors
– Mean of e2
– Mean of |e|
Total of 6 criteria
IBM Research
© 2006 IBM Corporation111
Model Comparison Results
Model Rational DB2 Tivoli
Displayed Model (kNN) 6 6 4 5 6 6
Max 03-05 Revenue 1 1 0 3 1 4
Linear Quantile 0.8 5 6 2 4 3 5
Regression Tree 1 3 2 4 1 2
Q-kNN 50 + flooring 2 3 6 6 4 6
Decomposition Center 0 0 3 5 0 4
Quantile Tree 0.8 0 1 2 4 1 4
(Anchoring)
(Best)
We count how often a model scores within the top 10 and 20 for each of the 6 measures:
IBM Research
© 2006 IBM Corporation112
MAP Experiments Conclusions Q-kNN performs very well after flooring but is
typically inferior prior to flooring
80th percentile Linear quantile regression performs consistently well (flooring has a minor effect)
Experts are strongly influenced by displayed opportunity (and displayed revenue of previous years)
Models without last year’s revenue don’t perform well
Use Linear Quantile Regression with q=0.8 in MAP 06
IBM Research
© 2006 IBM Corporation113
MAP Business Impact MAP launched in 2005
– In 2006 420 workshops held worldwide, with teams responsible for most of IBM’s revenue
Most important use is segmentation of customer base
– Shift resources into “invest” segments with low wallet share
Extensive anecdotal evidence to success of process
– E.g., higher growth in “invest” accounts after resource shifts
MAP recognized as 2006 IBM Research Accomplishment
– Awarded based on “proven” business impact
IBM Research
© 2006 IBM Corporation114
Summary
Wallet estimation problem is practically important and under-researched
Our contributions:
– Propose Wallet definitions: SERVED and REALISTIC
– Offer corresponding modeling approaches:
• Quantile estimation methods• Graphical latent variable model
– Evaluation on simulated, public and internal data
– Implementation within MAP project
We are interested in extending both theory and practice to other domains than IBM
IBM Research
© 2006 IBM Corporation115
References Marketing Science
– R. Rust, K. Lemon and V. Zeithaml, “Return on Marketing: Using Customer Equity to Focus Marketing Strategy”, J. of Marketing, 2004.
– P. Kotler, Marketing Management. Millennium Ed., Prentice-Hall, 2000.
Cost-sensitive Learning– P. Domingo, Meta-Cost: A general method for making classifiers cost-sensitive,
The 5th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 1999.
– B. Zadrozny, J. Langford and N. Abe, Cost-sensitive learning by cost-propotionate example weighting, in IEEE International Conference on Data Mining (ICDM), 2003.
– N. Abe, B. Zadrozny and J. Langford, An Iterative Method for Multi-class Cost-sensitive Learning, The Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, August 2004.
MDP and Reinforcement Learning– R. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, MIT Press,
Cambridge, MA, 1998. – L. P. Kaelbling, M. L. Littman, A. W. Moore, Reinforcement Learning: A Survey,
Journal of Artificial Intelligence Research, 1996.
IBM Research
© 2006 IBM Corporation116
References Bayesian Networks and Causal Networks
– K. Murphy, A brief introduction to Bayesian Networks and Graphical Models, http://www.cs.berkeley.edu/~murphyk/Bayes/bayes.html
– D. Heckerman, A tutorial on learning with Bayesian Networks, Microsoft Research MSR-TR-95-06, March 1995.
– J. Pearl, Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000.
– P. Spirtes, C. Glymour and R. Scheines, Causation, Prediction, and Search, 2nd Edition (MIT Press), 2000.
Case Study: Lifetime Value Modeling
– N. Abe, N. Verma, C. Apte and R. Schroko, Cross channel optimized marketing by reinforcement learning, The Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, August 2004.
– E. Pednault, N. Abe, B. Zadrozny, H. Wang, W. Fan and C. Apte, Sequential cost-sensitive decision making by reinforcement learning, The Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, August 2002.
IBM Research
© 2006 IBM Corporation117
References Case Study: Customer Wallet Estimation
– S. Rosset, C. Perlich, B. Zadrozny, S. Merugu, S. Weiss and R. Lawrence, Customer Wallet Estimation. 1st NYU workshop on CRM and Data Mining, 2005.
– S. Merugu, S. Rosset and C. Perlich, A New Multi-View Regression Method with an Application to Customer Wallet Estimation. The Twelfth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, August 2006.
– R. Koenker, Quantile Regression. Econometric Society Monograph Series,
Cambridge University Press, 2005.