Data-stories about (im)patient customers in tele-queues

Abstract

Credible queueing models of human services acknowledge human characteristics. A prevalent one is the ability of humans to abandon their wait, for example while waiting to be answered by a telephone agent, waiting for a physician’s checkup at an emergency department, or waiting for the completion of an internet transaction. Abandonments can be very costly, to either the service provider (a forgone profit) or the customer (deteriorating health after leaving without being seen by a doctor), and often to both. Practically, models that ignore abandonment can lead to either over- or under-staffing; and in well-balanced systems (e.g., well-managed telephone call centers), the “fittest (needy) who survive” and reach service are rewarded with surprisingly short delays. Theoretically, the phenomenon of abandonment is interesting and challenging, in the context of Queueing Theory and Science as well as beyond (e.g., Psychology). Last, but not least, queueing models with abandonment are more robust and numerically stable, when compared against their abandonment-ignorant analogues. For our relatively narrow purpose here, abandonment of customers, while queueing for service, is the operational manifestation of customer patience, perhaps impatience, or (im)patience for short. This (im)patience is the focus of the present paper. It is characterized via the distribution of the time that a customer is willing to wait, and its dynamics are characterized by the hazard-rate of that distribution. We start with a framework for comprehending impatience, distinguishing the times that a customer expects to wait, is required to wait (offered wait), is willing to wait (patience time), actually waits and felt waiting. We describe statistical methods that are used to infer the (im)patience time and offered wait distributions. Then some useful queueing models, as well as their asymptotic approximations, are discussed. In the main part of the paper, we discuss several “data-based pictures” of impatience. Each “picture” is associated with an important phenomenon. Some theoretical and practical problems that arise from these phenomena, and existing models and methodologies that address these problems, are outlined. The problems discussed cover statistical estimation of impatience, behavior of overloaded systems, dependence between patience and service time, and validation of queueing models. We also illustrate how impatience changes across customers (e.g., VIP vs. regular customers), during waiting (e.g., in response to announcements) and through phases of service (e.g., after experiencing the answering machine over the phone). Our empirical analysis draws data from repositories at the Technion SEELab, and it utilizes SEEStat—its online Exploratory Data Analysis environment. SEEStat and most of our data are internet-accessible, which enables reproducibility of our research.

Appendices

Appendix

A: Data repositories and EDA tools at the SEELab

SEELab is a research laboratory that opened in 2007 at the Technion [53]. (SEE stands for “Service Enterprise Engineering” [53].) At its beginning, SEELab focused on cleaning, archiving, and analyzing transaction-level call center data (event logs). Later, the data framework was extended to additional types of service systems, such as health care, internet, and face-to-face services.

Currently, SEELab databases include call-by-call data from four large call centers: a U.S. bank, two Israeli banks and an Israeli mobile-phone company. Three of the four databases cover periods of 2–3 years; the fourth one, which is presently the most active, has about 1.5 years data-worth.

The EDA environment of the SEELab is SEEStat—a software suite that enables real-time statistical analysis of service data at second-to-month resolutions. SEEStat implements many statistical algorithms: parametric distribution fitting and selection, fitting of distribution mixtures, survival analysis (mainly for estimating customers’ impatience), and more—all these algorithms interact seamlessly with all the databases. SEEStat interacts with SEEGraph, a pilot-environment for creating and displaying date-based (queueing) networks.

Two call center data-bases are publicly accessible, via the internet (at the SEELab server): from a small Israeli bank and a relatively large U.S. bank. The latter covers the operational history of close to 220 million calls; out of these, about 40 million were served by (up to 1,000) agents and the rest by a VRU (answering machine).

SEEStat Online: The connection protocol to SEELab data, for any research or teaching purpose, is simply as follows: go to the SEELab webpage http://ie.technion.ac.il/Labs/Serveng; then, either via the link SEEStat Online, or directly through http://seeserver.iem.technion.ac.il/see-terminal, and complete the registration procedure. Within a day or so, you will receive a confirmation of your registration, plus a password that allows you access to SEEStat, SEELab’s EDA environment, and via SEEStat to the above mentioned databases. Note that your confirmation email includes two attachments: a trouble-shooting document and a self-teaching tutorial. We propose that you print out the tutorial, connect to SEEStat and then let the tutorial guide you, hands-on, through SEEStat basics—this should take no more than 1 h.

B: Sources description for some figures in the paper

Some of our data sets are in the public domain and more will become available in the future. It is thus feasible, and of utmost significance, to provide references that enable other researchers to reproduce and validate the results of the present paper. Indeed, it is our strong belief that empirically-based research, specifically in Operations Research and Operations Management, should overcome the challenge of proprietary data, and strive for reproducibility, the latter being a well-established principle in traditional sciences. Accordingly, our data is publicly available [53], and Table 2 provides the sources of the figures that were created via SEEStat. (The names of the data sets coincide with their names in SEStat. Detailed instructions for reproducing our figures are provided in Nadjharov et al. [46].) Quoting the principle guiding [16]: “When we publish articles containing figures which were generated by computer, we also publish the complete software environment which generates the figures.”

Table 2 Source data for figures produced via SEEStat