Beyond Spam: ORMS Modeling Opportunities for Email Response Management

1
Beyond Spam OR/MS Modeling Opportunities for
Email Response Management

• Ramesh Sharda
• Robert A. Greve, Ashish Gupta, Manjunath Kamath,
  Mohan R. Chinnaswamy
• Oklahoma State University
• Ramesh.Sharda_at_okstate.edu

2
Managing Email

• Pull the plug!
• Spam control
• Email filtering and organization
• Effective management policies and strategies
• Organizational and Individual level
• Modeling opportunities

3
Our Projects

• Routing and priority decisions in an email
  contact center
• Simulation and queuing theory
• How often should we process our emails
• Simulation
• Which email messages to process
• Stochastic Programming with Recourse

4
Queuing and Simulation Models For Analyzing
Customer Contact Center Operations
5
Inbound E-mail Contact Center Issues

• Operational planning
• Number of agents
• Agents schedule
• Routing policies and priorities
• Routing to agents
• Processing order
• Performance measures
• Response/Resolution time
• Agent utilization
• Organizational behavior
• Human factors

6
Call/Contact Center Literature

• Koole and Talim 2000
• Exponential approximation in the design of call
  centers
• Koole and Mandelbaum 2002
• Queueing models of call centers
• Koole, Pot and Talim 2003
• Performance of call center with skill-based
  routing
• Armony and Maglaras 2002, 2003
• Optimal staffing policy
• Estimation scheme for the response time
• Whitt 2002
• Challenges and research directions in the design
  of customer contact centers

7
Contact Center Description
START
RECEIVE E-MAIL
IDENTIFY E-MAIL TYPE USING SOFTWARE
DIRECT E- MAIL TO AN AGENT
YES
PRE-PROCESS E-MAIL
FORWARD E-MAIL ?
PROCESS E-MAIL
DELAY
NO
END
RESOLVED ?
NO
YES
E-MAIL HANDLING LOGIC
8
Model Details

• Poisson arrivals
• Agents process e-mails according to a FCFS
  discipline
• For an unresolved problem, the e-mail enters the
  system with a delay independent of the prior
  processing
• The pre-processing time follows a uniform
  distribution
• The processing time follows a general
  distribution
• Erlang, Exponential, and Hyperexponential
• 2 types of e-mail and 3 agents

9
Open Queuing Network Model

• Nodes represent agents
• Customers represent e-mails
• Model parameters
• Number of nodes and the number of servers at each
  node
• Markovian routing probability matrix
• Mean and SCV (Variance/Mean2) service time at
  each node
• Arrival rate and SCV for new e-mails

10
Numerical Experiments

• Queueing model was solved using the Rapid
  Analysis of Queueing Systems (RAQS) package
• RAQS is a software package for analyzing general
  queueing network models based on a two-moment
  framework http//www.okstate.edu/cocim/raqs/
• Simulation results obtained using a model in
  Arena 7.0
• Replications 10
• Run length 9,240 hours
• Warm up 840 hours

11
EXPERIMENTAL DESIGN
Numbers refer to priorities assigned to new
email, previously processed (by the same
agent)
email, and previously processed (by the a
different agent) email, respectively. First Come
First Served (FCFS) gives priority only based on
arrival time.
12
EXPERIMENTAL DESIGN

• PERFORMANCE MEASURES
• Purchase Inquiry Response Time
• Purchase Inquiry Resolution Time
• Problem Resolution Request Response Time
• Problem Resolution Request Resolution Time

13
RESULTS (high utilization)

• Prioritization schemes that gave last priority to
  new email messages result in longer response and
  resolution times.
• By routing incoming messages to the agent with
  the fewest messages waiting for processing, the
  load is balanced across the agents.
• Routing messages to the agent who previously
  processed the message may result in disparity in
  individual agent utilizations, causing a gap
  between the best and worst performance.

14
CONCLUSIONS

• Simulation, which has been used for modeling
  customer call centers, can also be used to model
  the unique characteristics of customer contact
  centers
• Management decisions regarding routing and
  priority schemes can impact performance
• The queuing model consistently underestimates the
  system performance measures.

15
Scheduling Email Processing to Reduce Information
Overload and Interruptions
16
Prior relevant research on Email Overload
interruptions research

• First reported by Peter Denning (1982),Later by
  Hiltz, et al. (1985), Whittaker, et al.(1996) and
  many
• According to distraction theory, interruption is
  an externally generated, randomly occurring,
  discrete event that breaks continuity of
  cognitive focus on a primary task (Corragio
  1990 Tétard F. 2000).
• Research done in HCI is rich but in MS/OR???
• Research that looks at the problem of information
  overload and interruptions simultaneously is
  scarce. (Speier et al.1999, Jackson, et al.,
  2003, 2002, 2001), Venolia et al. (2003)

17
Research Model
Utilization Probability of a knowledge worker
being busy (?/µ)
18
Our approach- SIMULATION
Policies that we are comparing - Triage
(C1-morning, C1-Afternoon) Scheduled (C2, C4,
C8(Jackson et al. 2003)) Flow (continuous) C

• Phases of task processing
• (Miyata Norman, 1986)-
• Planning
• Execution
• Evaluation

19
Notations used

• i Task types- simple (S), complex (C) ,
  email(E) thus, i S, C, E
• Prim Primary task, which is either a simple or a
  complex task.
• Prim S, C
• ? Minimum utilization
  of knowledge worker
• ?i arrival rate for task of type i
• µi Service rates for task of type i
• P Planning phase of a task
• Exe Execution phase of task
• Eval Evaluation phase of task
• Stage Current stage of task processing. Thus
  Stage P, Exe, Eval
• IPrim-Stage Interruption lag for a primary task
  at a particular processing stage.
• RPrim-Stage Resumption lag for a primary task at
  a particular processing stage.
• RPrim, IPrim Mean Resumption lag Mean
  Interruption lag for a primary task

20
Mathematical conditions and equations

• Following conditions were implemented in the
  simulation model

21
Mathematical conditions equations
22
Mathematical conditions equations
23
Model Implementation
Sn, Cn- new simple complex task Si, Ci
interrupted simple complex task E Email
(Interrupt)
24
Profile plots
Results

• Policy C4 resulted in
• minimum percentage increase in utilization
• minimum of interruptions per simple task
• minimum of interruptions per complex tasks
• Result holds under
• The work environment requires high, medium or
  low utilization of knowledge worker, or
• The work environment requires processing of
• either more simple or more complex tasks, or
• For both arrival patterns (Pattern I when all
  email arrived during office hrs, Pattern II when
  80 emails arrived during office hrs).

RU
RU
of interruptions per simple or complex task
increase in utilization
25
Practical implications

• If other tasks are more important and email
  communication is secondary !
• Process emails 4 times a day with each processing
  block not exceeding 45 min.
• Is timely email processing a survival issue for
  your kind of organization?
• Use flow (continuous) policy

26
A Stochastic Programming Approach to Managing
Email Overload
27
Email Overload

• Inability to respond to all email in a timely
  manner
• The knowledge worker must not only take into
  consideration the current email that is in need
  of processing and the timeliness of this email,
  but he or she must also consider what future
  email demands may be on the horizon.
• Stochastic Programming takes possible FUTURE
  scenarios into consideration
• All other efforts consider only the present state
  

28
An Illustrative Example Optimizing Email
Processing

• With respect to email processing, the
  optimization involves maximizing the utility or
  value of the emails that are processed.
• The optimal solution must take into consideration
  that the utility of a processed email may
  decrease with time.
• The optimal solution must also consider the
  potential arrival of different types of email in
  the future.
• The decision variables correspond to whether or
  not to process an email in a given stage (time
  frame).
• The stochastic parameters include the potential
  arrival of various types of emails.

29
An Illustrative Example Optimizing Email
Processing

• Beginning Inbox (i type, j age)

30
An Illustrative Example Optimizing Email
Processing

• Utility of email processed (i type, j age)

31
An Illustrative Example Optimizing Email
Processing

• Arrival scenarios (number of type i email
  arriving)

32
An Illustrative Example Optimizing Email
Processing

• Time needed to process email (days)

33
Formulations

• LP Single-period
• LP Multi-period
• SP Perfect Information
• SP Here and Now

34
Formulation

• Sets and Indices
• T is the set of the different days under
  consideration
• I is the set of possible types of email messages
• J is the set of possible ages of an email
  message in days
• Q is the set of possible arrival scenarios
• t 1..4 denotes the day under consideration
• i 1..2 denotes the type of email message
• j 1..5 denotes the age of an email message
• q 1..64 denotes the arrival scenario

35
SP Formulation(Here and Now)

• Parameters
• Nt 1,i,j,q This represents the number of email
  of type i that are j
• days old on day one. This represents the
  beginning inbox.
• At,i,q This represents the number of arriving
  email of type i, given scenario q.
• Ui,j This represents the utility or value of an
  email of type i,
• having an age of j.
• Pq This represents the probability of scenario
  q.
• Di This represents the time needed, in days, to
  process an email of type i.

36
SP Formulation (cont.)(Here and Now)

• Variables
• Xt,i,j,q This represents the number of email
  that are processed on day t, that are of type i
  and have an age of j, given scenario q.
• Nt,i,j,q This represents the number of email of
  type i that are j days
• old on day t, given scenario q.
• Objective Function
• Max SqSiSj Pq Xt,i,j,q Ui,j

37
SP Formulation (cont.)(Here and Now)

• Constraints
• Nt,i,j,q Nt-1,i,j-1,q Xt-1,i,j-1,q t gt 1, i
  1..2, j gt 1, q 1..64
• Nt,i,j,q At,i,q t gt 1, I 1..2, j 1, q
  1..64
• Xt,i,j,q lt Nt,i,j,q t 1..4, i 1..2, j lt
  5, q 1..64
• Xt,i,j,q Nt,i,j,q t 1..4, i 1..2, j 5,
  q 1..64
• Si Sj Xt,i,j,q Di lt 1 t 1..4, i 1..2, j
  1..5, q 1..64
• Xt,i,j,q Xt,i,j,q1 t 1, i 1..2, j
  1..5, q lt 63
• Xt,i,j,q Xt,i,j,q1 t 2, i 1..2, j
  1..5, q lt 63
• Xt,i,j,q Xt,i,j,q1 t 3, i 1..2, j
  1..5, q lt 63
• Xt,i,j,q Xt,i,j,q1 t 4, i 1..2, j
  1..5, q lt 63

38
Sample Results
39
Extensions

• More realistic modeling of the problem needed
• Differences in service times for different email
  classes
• Identification of utilities
• Automatic identification of email categories
• Real time solution of the SPR problem before the
  Inbox is shown to the user
• Another optimization challenge

40
Future research

• Perform these studies in experimental or field
  settings.
• Use measures of Perceived Information Overload
  (NASA-TLX, SWAT)
• More realistic modeling by incorporating email
  characteristics as well as knowledge worker
  differences
• Single vs. multi-user settings/Network modeling
• Nonlinear formulations
• Stochastic knapsack
• A really rich domain for OR/MS modeling!!!