Data Mining and Knowledge Acquizition Chapter 9 Case Studies

1
Data Mining and Knowledge Acquizition Chapter
9 Case Studies

• Summer
• 2005

2
Chapter 9 Case Studies

• Churn Modeling in Wireless Comminications
• Web usage mining

3

• The choice of tool
• Type of model to build and which parameters to
  set
• Algorithm specific choices pruning decision
  trees
• How to segment the data for modeling
• The size and density of the model set
• How to handel the time element
• Which data to include and how to calculate the
  dreied variables

4
Segmenting the model set

• Club member two third high value
• Non club members
• Recent customers
• Joined in the previous eight or nine months
• Insufficient billing history

5
Churn Modeling in Wireless Comminications

• Predict which customers are likely to leave in
  the near future

6
The Wireless Telefone Industry

• Rapidly maturing
• Initial states
• Exponential growth
• Churn is not a problem
• Many new customers are joining then churning
• Many new customers for every churner
• Lost customers/new customers 15

7

• As the industriy matures
• Lost customer / new customer 80

8
Some Differences from ohter Industries

• Relatively high cost of acquizition
• Retaining an exisiting customer is much valuable
  then attracting a new one
• No direct customer contact
• Brand management and direct marketing
• Tremendous amount of data

9
The Business Problem

• The largerst company in a developing country
• Investing on DSS technologies
• Deregulated market
• Several enterents
• Market maturing (one third of population)
• Reactive marketing to proactive customer
  management
• Focus on exisiting customer
• How to keep them
• Make more profitable

10
Project Backround

• Parallel with development of data warehouse
• DSS based on relational OLAP by Microstrategy
• Slice and dice marketing and sales data
• Handset type, region, time of day
• Such questions
• MWhat is charen rate in April or May in club and
  non club members

11
Specifics about the Market

• About 5 millon customers
• Mostly in cities
• Type number percent churn rate
• Club 1,500,000 30 1.3
• Nonclub 3,500,000 70 0.9
• Average call per customer around 12
• No pay for incoming calls so data not collected
• Club members more valuable special promotion
  discount and coupons

12
What is Churn

• Involuntary not paying the bill for several
  months
• Predicting may reduce losses
• Voluntary every thing not involuntary
• Models for voluntary churn should not predict
  involuntary churn
• Involves
• Move out of service area
• Pass away
• Lured to other service provider
• Do not develop different models for each group

13
Why useful

• Churn prevension programs
• Discounts on air time
• Free incoming munites
• Other promotions
• Predict customers life time
• 1/(churn rate) expected lifetime
• Calculate lifetime value
• Prioritizing customer segments

14
Three Goals

• Near-term goal identify a list of probable
  churners
• Medium-term goal Build a churn management
  application
• Long-term goal Gomplete customer relationship
  management

15
Near-term goal identify a list of probable
churners

• The marketing department need top 10,000
• Not a score for each customer
• Club members only
• List by 24th of the month
• 10,000 club members most likely to churn by 24 th
  of each month

16
Medium-term goal Build a churn management
application CMA

• Aims
• Runing churn models
• Manage models
• Data analysis before and after modeling
• Import data and transform
• Export churn scores
• From building a data mining model to automating
  models as much as possible

17

• Auatomating new requirements
• incompatiChaning the modeling technique every
  month
• Manually pruning decision trees
• Incompatible with clustering
• Needs to have very reasonable defolt sets for
  modeling parameters
• Preclude some hybrid techniques first decision
  trees for variable selection then input into
  neural nets or lojistic regression

18
Approach to Building a churn model

• Define churn
• Inventory available data
• Customer information file age,ender ZIP
• Service account file activation data,
• Billing systems number of calls,
• Building models
• Decision trees
• Deploying scores

19

• Variables
• explain phenomena in the real world
• As opposedn to Mathematical transformations
• Drieved variables
• Growth rate of number of calls over time
• Proportion of calls of different type
• Change in proportions
• Call to customer service

20
Measure the scores against what realy happens

• How closes are estimated probabilities to the
  acutal churn probabilities for each group
• Are the churn scores relatively true
• Does a higher churn score imply a higher
  probability of churn

21
The Data

• Customer

Account
Usage by hour
Service
Monthly bill summaries
22

• Serviceaccess to the network provided to a
  single telepone number.
• Churn at the sevice level was primary interest
• Account multiple services may share the same
  account
• Several phone in a houshold
• Customer least defined of all

23

• Billing system to warehouse

24
Historical Churn Rates

• Predictors of churn in the near future is the
  recent history of churn in verious dimensions
• Handset churn rate
• Demographic churn rate
• Gender,age group,geographic area
• For a total of several hundred combinations
• ZIP code churn rate

25
Customer and Account Level

• Social security number
• Market id splits service area into different
  marketing regions
• Age and gender not so accurate but social
  security number can be used to drive
• Inaccurate data is self-reported
• Income occupation not used

26
Service Level

• Activation data and reason for activation
• Deactivation data for churning customers
• Features ordered by the customer
• Billing plan
• Handset type
• Dealer where the sercice was activated
• More accurate accept for dealers

27
Billing History

• Monthly summary for nine months
• Total amount billed,late charege and amount
  overdue
• All calls (numer of calls and amount billed)
• Oversee calls (number and amount billed)
• Fee paid services
• Directory assistace charges
• Provides several time series

28
Rejecting some Variables

• Variabes that cheat
• Future deactivation dates
• Identifiers customer id,social security num.
  Phone num.
• Very high skewalmost all values are identical
• Categories with too many values
• Group into larger units dealer location into
  market area
• Additional lookup information weight,manufacturer
  
• Historical churn rates included

29

• Absolute dates to relative dates
• Number of days to present
• Number of Activation days rather then activation
  data
• Seasonality information
• Store year or month or day of month
• Untrustworthy values
• Customer information is colleced by sale force
  when customer sign up for service
• No insentive for data collection
• Salary, occupation

30
Drived Variables

• Make sense
• Can be explained to business users
• Combination of variables
• Even they do not make apperent sense
• From billing system
• Summation of variable for all months and theri
  variance
• The ratio of each month value to the total
• Ratios between succesive months and between the
  first and last months
• Ratios within a month such as domestic or
  overseas usae to total number of calls within a
  month

31

• Some are redundent
• if number of calls in a month 0 then
• All ratio are 0
• Additional variables
• Age of customer, lenght of services
• What portion of customer life she is a customer
  length/age
• Rough estimate of customers worth
• Length of serviceaverge revenue per month

32
Lessons about building churn models

• Finding the most significant variables
• History is the best predictor of the future
• Churn rate of the handset
• Churn by demographics or ZIP code
• Number of different telephone by a customer
• Customer with multiple telephone much less likely
  to churn
• Number of change of feautures,age,market type
• Declien to 0 usage in the most recent month

33
Listening to the business users

• Before starting
• Assing a churn score for each customer
• Marketing department need top 10,000 churners
  that are club members
• Initially by using a single model
• Two different model for club and non-club members

34
Listening to the data

• Not all customer have 6 month of history
• Develop another model for recent customers
• much more accurate
• Lift value is close the theoretical prediction
• Billing data not available?
• Variables
• Handset is one
• Billing plan is family basic
• Handset has a high churn rate
• Exisiting customer if they leave handset and join
  to famiiy plans get discounts

35
Including historical churn rates

• Past is the best predictor of the future
• Variables
• Handset model
• Demographic age,income,..
• By area ZIP or market area
• Ussage patterns
• Breaking into several dimensions
• Quantiles for total billing,total number of
  calls,average duration

36
Composing the model set

• I one month of history
• Training set has 7 times higher
• IIseven months of history
• Seven month of history
• Churners are in one month
• Rich in history but only for one month september
• May overfit data
• III 4 months of history for predicting 3 months

37

• When historical data is limited
• When data warehouse is recently build
• Customer growth is so rapid
• Have little history
• Size and density
• Prefer 20-40 percent churners
• Model set as large as possible
• High oversampling rate
• Actual rate are in the order of 1

38
Build a model for churn management application

• Model automatically be rebuild
• Avoiding manual pruning and other manual
  processes
• Much effort to adjust parameter for not one model
  but

39
Listening data to determine model parameters

• Many models were build
• High density is requierd for churners
• Oversampling
• 30 found to be better

40

• Defining churn
• Interesting customers who leave for a competitor
• Uninteresting involuntary for not paying
• How the churn results will be uses
• A model for predicting life time value
• Listing high value customer for a camping
• Identifying data requirements
• Include historical churn rates for handsets and
  demographic rates
• Models slide in time windows

41
Segmentation

• Profilingusing data to profile or describe a
  group of customers or prospects
• Segmentation spliting the database into
  different sections and segments
• Market driven
• Data driven
• Demographic
• Psycographic
• Buying behavior
• Risk pattern
• Levels of profitibility

42
RMF Recency, Frequency, Monertary Value

• Recency number of months from the last purchase
• Predicting response to a subsequent offer
• if you have recently purchase something from the
  company, you are more likely to make another
  purchase
• Frequency number of purchases
• Total or
• Within a specified period
• Monetary valuetotal dollar amount
• Total or withi a specified period

43
How to create RMF scores

• Sort the data three times by each variable
• Each list is divided into equal slices
• Quantiles
• People in the top segment has a score of 5
• In the second segment has a score of 4
• Construct the RFM cube
• Customers in the same cell has the same score

44

• Customers who have made the most recently
  purchases, buy frequently and spend lots of money
  are in the 555 cell of the cube
• A customer who has made a recent purchase, buys
  frequently but does not spend lots of money might
  be in 542

45

• Each of the buckets for each variable may be
  recoded
• Bucket_1 coded as 9
• Bucket_2 coded as 6
• Bucket_3 as 3
• Weigth are given to each of three variabels
• Usually recency has the highest weight
• Ex recency 5, frequency 3, monetary 1
• A score is calculated for each customer
• Based on the encodings and weights

46
Example RFM socres

• Customer Recency frequency monetary
• Smith 09.2001 10 322
• Jones 10.2000 2 25
• Jonson 10.1999 4 120
• Frequency and monetary values are for the lost 24
  months
• Weights
• recency 5,
• frequency 3
• monetary 2

47

• Rules of encoding
• Recency 20 points if in last 3 months
• 10 p 6
• 5 p 9
• 3 p 12
• 1 p 24
• Frequencynumber of purchase within 24 months
• X 4 points (maximum 20 points)
• Monetary valeu spent lost 24 monts X 0.10
• (maximum 20 points)

48

• Customer Recency frequency monetary
• Smith 20 20 20
• Jones 3 8 2.5
• Jonson 1 16 12
• Customer Recency frequency monetary
• Smith 20x5 20x3 20x2 200
• Jones 3x5 8x3 2.5x2
  44
• Jonson 1x5 16x3 12x2 77

49
Disadvantages of RFM

• Based on only the three variables
• Other more valuable information may exist in the
  datawarehouse
• Recodings and weights of variables are arbitary
• Not statisitcally based
• Dificult the track customer movements from
  segments
• Customers move because of the actions of other
  customers

50
Web Mining

• Web content mining
• Web structure mining
• Web usage mining

51
Web Usage Mining
52
Describtion of data

• Click stream 1,148.6 MB
• 777480 raws
• First quarter of 2000
• Data has
• Session
• Assertment
• Content
• Product
• demographic

53
The problem

• Clustering sessions
• Transformations
• K-means used
• 0-1 normalization for
• Categorical variables
• order or see product
• Yes no to 0 1

54
Clustering Usege Sessions

• A session episode of interraction between the
  web user and the web server
• Ended when the user leave the web side
• Cluster the session based on their common
  proeperties
• Applications in education, e-commerce
• Another problem clustering web users
• Requier user data

55
Variables

• Avgtime agerage session time
• Totalclic total number of clicks in a session
• AvAsLev average assertment level
• Ass_UB assertment unique butique
• Seasonal,brand order,sale assertment, life
  style,M

56

• ________Clusters____________
• Variables 1 2 3 4 5 6
• Avgtime 47 43 53 39 41 50
• Tot_clicks 18 6 5 25 33 7
• Ass_avg 3.96 3.24 1.98 3.24 2.58 3.12
• Pr_yn yes no no yes yes yes
• Or_yn yes no no no yes no

57
Classification Problem

• 1. predict whether a visitor wil visit the
  legcare product subcategory in the subsequent
  three clicks
• 2. predict whether a visitor wil visit the
  legware product subcategory in the subsequent
  three clicks
• 3. predict whether a visitor wil click onther
  page or leave the session

58

• Functionalites classification
• Algorithm decision trees
• Tool Answer Tree