analytical model development & implementation experience from the field bhavani raskutti
TRANSCRIPT
Analytical Model Development & ImplementationExperience from the Field
Bhavani Raskutti
2
Topics to be covered
• Model development & implementation process
• Case Study 1: Corporate Customer Modelling at Telcos
• Case Study 2: Sales Opportunities for wholesalers
• Take-Home Points
3
Model Development & Implementation Process
Solution enabling
business to make
strategic & operational decisions
Business Problem
Data Acquisition & Preparation
DAP
AnalyticalProblem Definition
APD
D
Deployment
Presentation
P
Mathematical Modelling
(Algorithms)
Data Matrix
MM
Model Validation
MV
Decision-making by users• Insights via GUI• Automation• Training• Documentation• IT Support
Model Development• Iterative• 90% DAP
4
Topics to be covered
• Model development & implementation process
• Case Study 1: Corporate Customer Modelling at Telcos
• Case Study 2: Sales Opportunities for wholesalers
• Take-Home Points
5
Business Problem Large drops in margins & revenue in corporate customer base
Partial churn of some corporate customers to other telcos
Lack of understanding of customer’s needs
Project will target revenue improvement opportunities with an indicative $15 million in sales by:
undertaking a rapid analysis of Customer data from core systems, including front of house, customer satisfaction and marketing for customers with a spend greater than $100k, excluding state and local government
outcomes are to be validated using artificial intelligence tools and rigorous methodology by …
Verbatim from client’s presentation to stake holders
Using data analysis, increase revenue from corporate customers whose spend is > $100k
6
1. Analytical Problem Definition
• Increase revenue from corporate customers by- Win-back (database look-up)?- churn reduction? - Up-sell/cross-sell to an existing customer?
• Customer data- Relationship with customer
– Customer satisfaction survey data– Service assurance data (customer complaints)
- Demographic information about business customer– Industry segment information– Number of sites
- Revenue from customer– Quarterly revenue from different products
Create models to predict up-sell based on revenue data
1. Analytical Problem Definition
Using data analysis, increase revenue from corporate customers whose spend is > $100k
7
2. Data Acquisition & Processing
• Population:
- Customers in a segment who currently do not have the product being modelled
• Target or positive case definition:
- Customers in the segment who take up the product within a time period
• Predictors for modelling
Using revenue data, create models to predict customers likely to take up a specific product
2. Data Acquisition & Processing
8
Population and Target Definition• Let riP be the revenue from a customer on product P in billing
period i
• Population in period i includes all customers with r(i-1)P = 0
• Target or Product take-up in period i iff r(i-1)P=0 and riP >TUMIN
- TUMIN > 0 is the minimum take-up amount determined by the business
Predictors Labels
TRAIN: r(i-1)P = 0
Predict for riP = 0
i i+1
i-1 i
2. Data Acquisition & Processing
9
Low take-up rates: not enough targets
• Average number of take-ups for any product in any period is small- Large businesses
– Less than 20 take-ups in a period for 70 of the 100+ products– Less than 10 take-ups for 45 products
- Medium businesses– Less than 20 take-ups for 71 products– Less than 10 take-ups for 60 products
• Reasons- “niche” products- Saturated products
2. Data Acquisition & Processing
10
Low take-up rates (Cont’d)Impact of data aggregation
k=2 is useful
Large Businesses
70
4548
39
51
40
0
10
20
30
40
50
60
70
n=20 n=10
Minimun take-ups (n) for modelling
Nu
mb
er
of
un
mo
dell
ab
le p
rod
ucts
k=1
k=2
k=3
Minimum take-ups(n) for modelling
Medium Businesses
7166
54
71
5960
0
10
20
30
40
50
60
70
n=20 n=10
Minimun take-ups (n) for modelling
Nu
mb
er
of
un
mo
dell
ab
le p
rod
ucts
k=1
k=2
k=3
Minimum take-ups(n) for modelling
• Aggregate data over multiple billing periods k
• Product take-up in periods i to i+k-1 iff r(i-j)P=0 for j=1..k and j=0..k-1 r(i+j)P >(kTUMIN))
Predictors
Labels
i-3 i-2 i-1 i
TRAIN target: r(i-j)P = 0, j = 0..1
Predict if r(i+j)P = 0 or 1; j = 1..2
i-1 i i+1 i+2
2. Data Acquisition & Processing
11
Low take-up rates (cont’d)• Use of time interleaving
- Aggregate data with k=2- Generate 3 sets of data
moved forward by a period- Concatenate the 3 sets to get
3 times as much training data as for data aggregation with k=2
Impact of time interleaving
Time interleaving enormously enhances modellability
Large Businesses
70
4539
28
19
48
0
10
20
30
40
50
60
70
n=20 n=10Minimum take-ups (n) for modelling
Nu
mb
er o
f u
nm
od
ella
ble
pro
du
cts Raw
DA, k=2
TI
Medium Businesses
60
7166
5449
40
0
10
20
30
40
50
60
70
n=20 n=10Minimum take-ups (n) for modelling
Nu
mb
er o
f u
nm
od
ella
ble
pro
du
cts
Raw
DA, k=2
TI
i-5 i-4 i-3 i-2
PredictorsPrediction
LabelsTRAIN
i-4 i-3 i-2 i-1
i-3 i-2 i-1 i
i-1 i i+1 i+2
2. Data Acquisition & Processing
12
Predictors for Modelling
• Revenue predictors used- r(i-3)Q – revenue for all products in billing period i-3- Change in revenue from period i-3 to i-2, r(i-3)Q - r(i-2)Q
- Projected revenue for period i-1, 2r(i-3)Q - r(i-2)Q
• All revenue predictors used both as raw values, and normalised by total customer revenue
• Binary predictors indicating churn/take-up in period i-2
• All continuous predictors converted to binary using 10 equisize bins- Overcomes the negative impact of large variance in revenues- Allows generation of non-linear models using linear techniques
Predictors Labels
i-3 i-2 i-1 i
TRAIN target: r(i-j)P = 0, j = 0..1
2. Data Acquisition & Processing
13
3. Mathematical Modelling• Imbalance in class sizes
- Large businesses– 51 products with < 0.5% take-up on average– 25 products with < 0.1% take-up
- Medium businesses– 74 products with < 0.5% take-up on average– 54 products with < 0.1% take-up
• Maximisation of total take-up revenue - Identifying new high value customers is a priority- Extent of variance
– Take-up amounts range from TUMIN to over a million dollars– Take-up amounts are not correlated with total revenue in
previous billing periods
3. Mathematical Modelling
14
Imbalance in class sizes• Use of Support Vector Machines (SVMs) instead of decision
trees, neural nets or logistic regression
- Based on Vapnik’s statistical learning theory- Maximises the margin of separation between two classes
• Two different SVM implementations
- SVMstd : equal weight to all training examples
- SVMbal : class dependent weights so all take-ups have a higher weight than all non-take-ups
m
mCC
• m+ and m- : number of +ve and -ve examples
• C+ and C- : weight of +ve and -ve examples
3. Mathematical Modelling
15
Identifying high value take-up
• SVMval: SVM with different weights for different positive (take-up) training examples
- All take-up examples have a higher weight than all the non-take-up examples (as for SVMbal)
- Each take-up training example has a weight proportional to the amount of take-up
MINTU
iTU
m
mCiC
2
)()(
• m+ and m- : number of +ve and -ve examples
• C- : weight of -ve examples
• TU(i) : Take-up amount of the ith +ve example
• C+(i) : weight of the ith +ve example
3. Mathematical Modelling
16
4. Model Validation• Model assessment
- Two tests for assessing quality of models (~4,000 models)– 10-fold cross validation tests to determine the best of the 3 SVMs– Tests in production setting to evaluate time interleaving
- All tests on 30 product take-up prediction problems in 4 segments - Performance measures on unseen test set
– Area under receiver operating characteristic curve (AUC)• Measures quality of sorting• Decision threshold independent metric
– Value weighted AUC (VAUC)• Indicates potential revenue from the sorting
• SVMval with time interleaved data is used for generating models
- SVMval significantly more accurate than the other two
- Time interleaving produces more stable models
4. Model Validation
17
Model Validation by Business
• Predictive models identify more sales opportunities than that identified manually- 3 times as many in large businesses segment- 5 times as many in medium businesses segment
• Results for 2 different regions in medium businesses- Region 1: Predictions for just 5 products generated 9 new
opportunities with an increase in revenue of ~400K A$- Region 2: Predictions identified opportunities that were
already being processed by sales consultants
• Predictive modelling spreads the techniques of good sales teams across the whole organisation
4. Model Validation
18
5. Presentation
• Output in Excel Spread Sheet automatically generated
• One customer list per segment with:
- Take-up likelihood for all modelled products- Last quarter revenue for all products
5. Presentation
19
6. Deployment• Implementation in Matlab & C with output in Excel
• Automatic quarterly updates of model after consolidated revenue figures are available
• Models for ~50 products for each of the 4 business segments
• Output delivered to business analytics group
- Different cut-offs for different products/regions- Superimposition of other data for filtering/sorting
• Use of output by sales consultants for renegotiating contracts with customers
6. Deployment
20
Project Timeline
• Initial approach to data availability for pilot: 12 weeks
• Data to pilot: 6 weeks
• Model validation by business: 12 weeks
• Pilot deployment (5 products, 1 segment): 6 weeks
• Acceptance by business teams: over 9 months
• Final deployment: 4 weeks
• In operation for more than 8 years!!
6. Deployment
21
Key Success Factors• Willingness of stake-holders to try non-standard solutions
• Innovative solution: Paper published in KDD 2005 - Target definition using multiple overlapping time periods to boost
the number of rare events for modelling- Use of support vector machines for customer analytics
• Being lazy - Scope change from 4 to 50 products- Scope change from 2 to 4 segments- Development of ~200 predictive models in one shot - No stale models in production
• Working with business analysts to instigate change:- Product-centric modelling to customer-centric product packaging
22
Topics to be covered
• Model development & implementation process
• Case Study 1: Corporate Customer Modelling at Telcos
• Case Study 2: Sales Opportunities for wholesalers
• Take-Home Points
23
DAP
APD
DP
MM
MV
- Sales demand - Similar products
@ similar outlets have similar demand to sales relationship
- Anomaly may be due to lack of stock
Increase wholesale sales
into major retailers
- Quantify demand - Define normalised
sell-rate - Define a long term
in-stock measure - Define products &
outlets that are similar
- Weekly SOH & sales for each store & SKU
- SKU master
- Store master
Simple univariate regression in SQL
Perform comparisons & find anomalies
with stock issues
- Self-serve report for each sales rep
- Presents list of products with sales opportunities
- Click thru’ to detailed graphs
Case Study: Wholesale Sales
- Absolute error
- Validate with retail
24Demand
In-s
tock
%
· R1· R2
Demand
Sel
l Rat
e
Sell rate vs Consumer Demand plot • Each point is a store• R1 & R2 are comparable retailers• Values for the same product
Possible reasons for difference• Competing product at R2• Pricing at R2 vs R1• Lack of stock at R2
Case Study: Wholesale Sales (Cont’d)
25
DAP
APD
DP
MM
MV
- Sales demand - Similar products
@ similar outlets have similar demand to sales relationship
- Anomaly may be due to lack of stock
Increase wholesale sales
into major retailers
- Quantify demand - Define normalised
sell-rate - Define a long term
in-stock measure - Define products &
outlets that are similar
- Weekly SOH & sales for each store & SKU
- SKU master
- Store master
Simple univariate regression in SQL
- Self-serve report for each sales rep
- Presents list of products with sales opportunities
- Click thru’ to detailed graphs
- SQL & Cognos
- Automatic weekly updates
- Training by corporate training team
- Support from IT helpdesk
Perform comparisons & find anomalies
with stock issues
Case Study: Wholesale Sales (Cont’d)
- Absolute error
- Validate with retail
26
Topics to be covered
• Model development & implementation process
• Case Study 1: Corporate Customer Modelling at Telcos
• Case Study 2: Sales Opportunities for wholesalers
• Take-Home Points
27
Take-home points
• Data acquisition & processing phase forms 80-90% of
any analytics project
• Business users are tool agnostic
- R, SAS, Matlab, SPSS, … for statistical analysis
- Tableau, Cognos, Excel, VB, … for presentation
• Business adoption of analytics driven by
- Utility of application
- Validation of results by using real-life cases
- Ease of decision-making from insights
- Ability to explain insights