Mastering Data Mining: Chapter 10

1
Waste Not, Want Not Improving Manufacturing
Process
2004. 3. 23. 2000-12008 ???, 2004-20971 ???
2
Introduction

• CRM is the killer application for data mining
• Another fruitful application for data mining is
  cost reduction through industrial process
  improvement
• The efficiency of many manufacturing processes
  depends on hundreds or thousands of variables
  whose interactions are not well understood
• We can find how variables are tied together
  through the process using data mining

3
Data Mining to Reduce Cost at R.R.Donnelley

• R. R. Donnelley and Sons
• The largest printing company in the United States
• Having sales of around 5 billion yearly
• Having 49 manufacturing plants
• One of these plant is the setting for this
  chapters first case study
• The company has three principal strategies for
  enhancing customer and shareholder value
• The first strategy as driving continuous cost
  reduction and productivity improvement by
  creating process improvements, optimizing our
  manufacturing operations and enhancing
  supply-chain management

4
The Problems of R.R.Donnelley

• The technical problem
• High-speed printing is done on rotogravure
  presses
• There has been a problem called cylinder
  banding
• The symptom of cylinder banding is a streak of
  ink running across the printed image, ruining the
  print job
• The business problem
• The business problem was learning how to avoid
  costly interruptions to print runs by isolating
  the conditions under which certain defects appear

5
The Data of R.R.Donnelley

• Data mining depends on data
• In order to build a decision tree capable of
  explaining anything, there must be a training set
  consisting of observations of relevant variables
  for both positive and negative outcomes
• At the start, no such data existed
• They needed to choose a data mining approach that
  would help explore the data and identify a
  handful of the most important variables
• The team set out to collect data on humidity, ink
  temp., ink viscosity, acidity, voltage level,
  blade pressure, paper type, etc.

6
The data of R.R.Donnelley

• Deciding on the right attributes
• The final set of input variables included
• ESA anode distance in millimeters, chrome
  solution ratio, ESA current density in amperes
  per square decimeter, plating tank, viscosity,
  etc.
• Total 17 classes
• Data preparation for continuous inputs
• What was needed was a way to pick splits that
  captured the notion that there is one range that
  affects the outcome in one direction, a second
  range that affects the outcome in the opposite
  direction, and a neutral range separating the
  first two
• Defining the target class
• How long a run is long enough to be considered
  free of banding?

7
Including Rules for Cylinder Bands

• The goal of this project were different from the
  goals of most of the data mining projects
• In this case, there was no plan to use the
  decision trees as predictive models
• Instead, the trees were used to generate a set of
  practical, prescriptive rules that could be
  applied on the shop floor
• We starts out with a set of heuristics
• Example higher values of ink temperature, lower
  values of humidity increase likelihood of banding
• Making split with the ID3 decision tree algorithm
• A set of operating guidelines to try on the shop
  floor
• Keep the chrome solution ratio high
• Keep the ink temperature low
• Keep the ink viscosity high

8
Change on the Shop Floor

• Figure 14.3 shows the incidence of banding over
  time
• It is not hard to pinpoint the time when the new
  guidelines came into effect!
• It took some time for confidence in the new
  guidelines to spread through the print crews
• Long-term impact
• Before the data mining, 538 banding incidents
  caused more than 800 hours of downtime
• After the data mining, only 21 such incidents
  resulting in 30 hours of downtime

9
Business Problem of Time-Warner

• Decreasing expenses
• To control paper cost is to buy when the price is
  right
• To control paper costs is to use less paper
• Relationship of Time Inc. to the Printing Plants
• Printing plants have no direct incentive to be
  frugal with paper
• The contractual limits on the amount of waste
• Performance Variation between plants
• To reduce the waste is valuable

10
The Data of Time-Warner

• Invoice
• Useful information for a shipment of paper
• Stock Status
• Information on paper in inventory including
  transit damage and in-plant damage to paper rolls
• Transfer
• Sometimes paper is transferred between printing
  plants
• Usage
• To track the magazine or department that gets
  charged for a press run

11
The Data of Time-Warner

• Press Run
• This was the most important table. It includes
  the press start time, the linear feet of paper
  used after makeready, the press stop time, etc.
• Spoilage Detail
• More information about waste, including
  calculated allowances for blanket-wash waste and
  bindery waste and the weight of paper printed
  after the correct number of pages has been
  reached.
• Shipment
• Information about how the paper got to the plant,
  and the carriers that transported it down to
  which door of a two-door box car from which it
  was unloaded

12
Approach to the Problem of Time-Warner

• Hypothesis Testing
• Waste as a function of press type
• Waste as a function of paper age
• Waste as a function of basis weight
• Waste as a function of time of day
• Waste as a function of presses equipped with
  automatic blanket washers
• Waste as a function of number of rolls in press
  run

13
Types of Waste

• Types of Waste
• Wrapper waste
• Strip waste
• Make-ready waste
• Running waste
• Core waste
• Bindery waste
• Trim Waste
• Overrun
• Addressable Waste
• This is avoidable waste
• Inducing rules for addressable waste

14
Data Transformation

• Transformation on the raw data
• Changing the format
• Changing the type
• Convenience Fields
• Several convenience fields were added to the data
• Two types of derived fields
• Continuous variables ? Categorical variables
• Creating derived fields that contain information
  from two or more other fields
• Classification Target
• Data Characterization and Profiling
• Statviz produces a small graph for each variable

15
Decision Trees

• Classification versus Explanation
• The goal is not to classification
• What is the most important factors affecting
  waste
• Extracting Rules for Addressable Waste
• To increase the minimum number of records allowed
  in a node
• To look only at nodes classified as high waste
  ignoring those which describe more usual runs
• Other rules
• Association Rules

16
Putting It All Together

• Actionable information that they could
• There is correlation between paper age and
  running waste
• Print runs using paper from multiple mills had
  slightly higher running waste percentages
• A fifth color in addition to the usual four,
  something that happens fairly often with the
  cover of Time magazine, leads to increased
  running waste

17
Lessons Learned

• The projects used data mining to find
  prescriptive rules that could be used to improve
  the production process
• The projects succeeded due to the constant
  involvement of subject matter experts who
  understood paper and printing inside out and were
  willing to provide guidance to the data miners
• Implementation of the new policies suggested by
  data mining requires the active cooperation of
  the people on the plant floor
• Data mining does not always require huge volumes
  of data-The Donnelley study used only a few
  hundred records
• Where huge volumes of data are available, as at
  Time Inc., data mining can help sense of it.