Recent Publications

Problem Definition: Research in Operations Management has focused mainly on system-level load, ignoring the fact that service agents and customers express a variety of emotions that may impact service processes and outcomes. We introduce the concept of emotional load—the emotional demands that customer behaviors impose on service agents—to analyze how customer emotions affect service worker’s behavior.
Academic / Practical Relevance: Most theories in Organizational Behavior literature predict that emotions expressed by customers reduce agent’s cognitive abilities and therefore should reduce the agent’s speed (e.g. by increasing the service time required to serve an angry customer). We aim to shed light on the magnitude of that phenomenon while addressing important econometric challenges. We also investigate an important mechanism that drives this relation, namely, agent effort. We discuss practical opportunities that arise from measuring emotional load, and how it can be used to enhance productivity.
Methodology: We measure the emotional load of agents using sentiment analysis tools that quantify positive/negative customer emotion expressions in an online chat-type contact center, and link it to agent behavior: response time, and the length and number of messages required to complete a service request. Identifying a causal effect of customer emotion on agent behavior using observational data is challenging because there are confounding factors associated to the complexity of service requests, which are related to both customer emotions and agent behavior. Our identification strategy uses panel data and exploits the variation across messages within a focal request, using fixed effects to control for unobserved factors associated to case complexity. Instrumental variables are also used to address issues of measurement error and other endogeneity problems; the instruments are based on exogenous shocks to agent performance indicators that have been studied in the service operations literature.
Results: Analyses show that emotional load created by negative customer emotions increases agent response time (RT), the length of the agent messages (a measure of effort) and the required number of messages needed to complete a service request. Emotional load and agent RT reciprocally effect each other, with long agent RTs and a high number of messages producing more negative customer emotion.
Managerial Implications: We suggest that the emotional content in customer communications should be an important factor to consider when assigning workload to agents in a service system. Our study provides a rigorous methodology to measure the emotional content from customer text messages and objectively evaluate its associated workload. We discuss how this can be used to improve staffing decisions and dynamic workload routing through real-time monitoring of emotional load.

Problem definition: Developing estimators for delay distributions in a queueing network with a Fork-Join structure.
Academic / Practical Relevance: Delay announcements have become an essential tool in service system operations; they influence customer behavior and network efficiency. Most current delay announcement methods are designed for relatively simple call-center environments. Here we focus on a complex Fork-Join network of an Emergency Department. Such systems are mostly in a transient state and although queues for each station are fairly short, delays are long. These delays are the result of both resource scarcity and synchronization delays.
Methodology: We develop an exact analysis of the system. The analysis is based on recursively constructing the LaplaceStieltjes transform of the joint distributions, conditioning on customers’ movements in the network. We then examine the accuracy of the proposed approach on data of an Emergency Department.
Results: Estimations are provided throughout the customer’s stay in the network under Markovian assumptions; we then discuss (and provide) relaxations of both service time distribution and structural model assumptions. We provide evidence that the methodology developed is better than Last-to-Enter-Service (LES) estimators (which are based on snapshot principle arguments), reducing Root-Mean-Squared-Error (RMSE) by 25–30%. Our use case demonstrates the importance of incorporating speedup effects of service rates into delay estimators, which are naturally captured by our model, improving accuracy by 10%.
Managerial Implications: The results of our study allow management to implement delay announcements in complex Fork-Join networks. We also highlight the power of using exact approaches instead of relying on approximations, though the latter might be necessary for larger systems.

Abstract: Motivated by a large healthcare clinic that provides periodic health screening services, we study how to improve the system performance of an open-shop service system using priority and routing policies combined with information technologies. We consider the system performances from both macro system-level perspective, such as total wait, and micro station-level perspective, such as excessive waits at single stations. We empirically evaluate automation efforts done by the clinic to improve the streamline of patients using automatic routing and queue communication using SMS notifications and find that such practices have negligible impacts on system performance. We, therefore, use a stylized queueing model and simulations to study priority policies for open-shop networks that consider customers' profiles. We demonstrate that such policies may outperform the station-level FCFS policy by simultaneously improving both the macro- and micro-level performance measures. One of the challenges we uncover is the tension between pooling consid erations that promote last-minute decisions and behavioral considerations that promote early decisions to enable customers to smooth transfer from one station to another. To overcome this tension, we propose a buffer strategy that postpones the routing decisions, which leaves the system more exibility and gathers more information for decision making. We demonstrate the efficacy of this policy in improving and balancing these performance measures.

Abstract: Service systems are often stochastic and preplanned by appointments, yet implementations of their appointment systems are prevalently deterministic. At the planning stage of healthcare services, for example, customer punctuality and service durations are often assumed equal their means - and this gap, between planned and reality, motivated our research. Specifically, we consider appointment scheduling and sequencing under a time-varying number of servers, in a data-rich environment where service durations and punctuality are uncertain. Our data-driven approach, based on infinite-server queues, yields tractable and scalable solutions that accommodate hundreds of jobs and servers. We successfully test our approach against near-optimal algorithms (which exist for merely single servers). This entails the development of a data-driven robust optimization approach with novel uncertainty sets. To test for practical performance, we leverage a unique data set from a cancer center that combines real-time locations, electronic health records, and appointments log. Focusing on one of the center’s infusion units (roughly 90 daily appointments, 25+ infusion chairs), we reduce cost (waiting plus overtime) on the order of 15%–40% consistently, under a wide range of experimental setups.

Problem definition: We study the problem of designing a dynamic invitation policy for proactive service systems with finite customer patience under scarce capacity. In such systems, prior knowledge regarding customer value or importance is used to decide whether the company should offer service or not. Academic/practical relevance: Proactive service systems are becoming more popular, as data availability and machine-learning techniques are developed to forecast customer needs. However, very little is known about the efficient use of such tools to promote and manage service systems. Methodology: We use fluid approximation and the Filippov convex method to analyze system dynamics and develop approximations for important performance measures. Results: We show that whereas prioritizing customers in descending order of their r-µ ranking (as long as there are idle servers in the system) is optimal on the fluid level, refinements are necessary in the presence of abandonment on the stochastic level. We propose an r-µ-N policy to account for customer patience. Managerial implications: Our policy can be used to promote service effectivenes and allow decision makers the means to trade off service level against costs in such systems explicitly. Using a case study of a transportation service provider, we show that such a policy can triple revenue compared with random arrivals (nonproactive policy).

Abstract: We explore whether customers are loss averse in time and how delay information may impact such reference-dependent behavior using observational and field-experiment data from two call centers of an Israeli bank. We consider settings with no announcements and announcements of different accuracy levels. We face two key challenges: (1) we do not directly observe the reference points customer use as any other studies using field data; and (2) it is difficult to separate the reference-dependent behavior from the potential non-linear waiting cost of customers. To address these challenges, we develop a dynamic decision model with consumer learning, through which we infer the reference point each customer used during any given call. The reference points may be different across different customers and evolve across different calls of the same customers. We also exclude the alternative explanation by showing that our main reference-dependent models better explain the observed customer abandonment than models where customers have non-linear waiting cost. Our results indicate that customers are loss averse regardless of the availability or accuracy of the announcements, when their waiting time is relatively long (90s or longer). While delay announcements do not alter the nature that customers are loss averse, accurate announcements may affect customers' belief about the offered waiting time and thus impact the reference points. Through counterfactual studies, we demonstrate that providing delay announcements improves the call center performance given the loss aversion behavior observed in our data. Interestingly, as customers become more loss averse, the value of providing delay announcements decreases.

Abstract: We highlight two psychological aspects of the load in service work – cognitive load (amount of information customers present) and emotional load (emotions customers present), and examine their effects on response time of service agents, in service conversations conducted using text-based chats. Using operational data of 145,995 chat service conversations, we show that cognitive load and emotional load increase agent response time both between and within service conversations. Our analyses unpack common assumptions that number of customers is identical to amount of work load, and shed light on customer-agent dynamics both between and within service conversations. In studying text-based service communication, which is rapidly expanding and insufficiently studied, we open up exciting opportunities for further research.

Abstract: A significant challenge in declarative approaches to scheduling is the creation of a model: the set of resources and their capacities and the types of activities and their temporal and resource requirements. In practice, such models are developed manually by skilled consultants and used repeatedly to solve different problem instances. For example, in a factory, the model may be used each day to schedule the current customer orders. In this work, we aim to automate the creation of such models by learning them from event data. We introduce a novel methodology that combines process mining, timed Petri nets (TPNs), and constraint programming (CP). The approach learns a sub-class of TPN from event logs of executions of past schedules and maps the TPN to a broad class of scheduling problems. We show how any problem of the scheduling class can be converted to a CP model. With new instance data (e.g., the day’s orders), the CP model can then be solved by an off-the-shelf solver. Our approach provides an end-to-end solution, going from event logs to model-based optimal schedules. To demonstrate the value of the methodology we conduct experiments in which we learn and solve scheduling models from two types of data: logs generated from job-shop scheduling benchmarks and real-world event logs from an outpatient hospital.

Abstract: Analysing performance of business processes is an important vehicle to improve their operation. Specifically, an accurate assessment of sojourn times and remaining times enables bottleneck analysis and resource planning. Recently, methods to create respective performance models from event logs have been proposed. These works have several limitations, though: They either consider control-flow and performance information separately, or rely on an ad-hoc selection of temporal relations between events. In this paper, we introduce the Temporal Network Representation (TNR) of a log. It is based on Allen’s interval algebra, comprises the pairwise temporal relations for activity executions, and potentially incorporates the context in which these relations have been observed. We demonstrate the usefulness of the TNR for detecting (unrecorded) delays and for probabilistic mining of variants when modelling the performance of a process. In order to compare different models from the performance perspective, we further develop a framework for measuring performance fitness. Under this framework, TNR-based process discovery is guaranteed to dominate existing techniques in measuring performance characteristics of a process. In addition, we show how contextual information in terms of the congestion levels of the process can be mined in order to further improve capabilities for performance analysis. To illustrate the practical value of the proposed models, we evaluate our approaches with three real-life datasets. Our experiments show that the TNR yields an improvement in performance fitness over state-of-the-art algorithms, while congestion learning is able to accurately reconstruct congestion levels from event data.

Abstract: Using call center field data, we examine the reference-dependent impact of delay announcements on both customer abandonment behavior and time spent in service. Delay announcements induce a reference point telling customers how long they must wait for service. Customers behave differently before and after this reference point. Our novel contributions include demonstrating the effects of delay announcements on service time, as well as simultaneously considering both reference-dependent abandonment behavior and reference- dependent service times. One additional minute spent waiting after the reference point increases the hazard rate of abandonment by up to 52%, a much larger effect than the hazard rate decrease from having one additional minute to go before the reference point. For those customers who do not abandon, we also observe reference-dependent behavior during service. Conditional on the same total wait in queue, one additional minute spent waiting after the reference point triggers up to 21 extra seconds in service. While longer total waiting time in queue also leads to longer service time, its effect is much smaller than this reference-dependent effect. Reference-dependent behaviors in abandonment and service impact system-level congestion in opposite directions. Therefore, managers should consider both effects when considering operational decisions, including staffing and customer prioritization.

Abstract: Call centers’ managers are interested in obtaining accurate point and distributional forecasts of call arrivals in order to achieve an optimal balance between service quality and operating costs. We present a strategy for selecting forecast models of call arrivals which is based on three pillars: (i) flexibility of the loss function; (ii) statistical evaluation of forecast accuracy; and (iii) economic evaluation of forecast performance using money metrics. We implement fourteen time series models and seven forecast combination schemes on three series of daily call arrivals. Although we focus mainly on point forecasts, we also analyze density forecast evaluation. We show that second-moment modeling is important for both point and density forecasting and that the simple seasonal random walk model is always outperformed by more general specifications. Our results suggest that call center managers should invest in the use of forecast models which describe both first and second moments of call arrivals.

Abstract: Many service centers offer a “call-back” option, in which customers entering the queue are informed of the anticipated (on-line) wait and can choose to wait either online or off-line till an agent contacts them. We show that such a policy has the potential of both improving service performance and server utilization, by balancing the load between overloaded and underloaded periods. However, our analysis suggests that companies need not offer that service at all times, and that the delay guarantees proposed should be planned according to the anticipated load throughout the day. In order to optimize the operation of such a system,we develop an Iterative Simulation Algorithm to determine what delay guarantees the company should offer in a time-varying environment. Those guarantees depend on service level targets the company wishes to provide and the delay sensitivity of the online vs. off-line customers.

Abstract: This paper develops an aggregate stochastic model of an emergency department (ED) based on a careful study of data on individual patient arrival times and length of stay in the ED of the Rambam Hospital in Israel, which was used in a large-scale exploratory data analysis by Armony et al. (2015). This data set is of special interest because it has been made publicly available, so that the experiments are reproducible. Our analysis confirms the previous conclusions about the time-varying arrival rate and its consequences, but we also find that the probability of admission to an internal ward from the ED and the patient length-of-stay distribution should be time varying as well. Our analysis culminates in a new time-varying infinite-server aggregate stochastic model of the ED, where both the length-of-stay distribution and the arrival rate are periodic over a week.

Abstract: Service processes, for example in transportation, telecommunications or the health sector, are the backbone of today׳s economies. Conceptual models of service processes enable operational analysis that supports, e.g., resource provisioning or delay prediction. In the presence of event logs containing recorded traces of process execution, such operational models can be mined automatically.
In this work, we target the analysis of resource-driven, scheduled processes based on event logs. We focus on processes for which there exists a pre-defined assignment of activity instances to resources that execute activities. Specifically, we approach the questions of conformance checking (how to assess the conformance of the schedule and the actual process execution) and performance improvement (how to improve the operational process performance). The first question is addressed based on a queueing network for both the schedule and the actual process execution. Based on these models, we detect operational deviations and then apply statistical inference and similarity measures to validate the scheduling assumptions, thereby identifying root-causes for these deviations. These results are the starting point for our technique to improve the operational performance. It suggests adaptations of the scheduling policy of the service process to decrease the tardiness (non-punctuality) and lower the flow time. We demonstrate the value of our approach based on a real-world dataset comprising clinical pathways of an outpatient clinic that have been recorded by a real-time location system (RTLS). Our results indicate that the presented technique enables localization of operational bottlenecks along with their root-causes, while our improvement technique yields a decrease in median tardiness and flow time by more than 20%.

Abstract: Information systems have been widely adopted to support service processes in various domains, e.g., in the telecommunication, finance, and health sectors. Information recorded by systems during the operation of these processes provides an angle for operational process analysis, commonly referred to as process mining. In this work, we establish a queueing perspective in process mining to address the online delay prediction problem, which refers to the time that the execution of an activity for a running instance of a service process is delayed due to queueing effects. We present predictors that treat queues as first-class citizens and either enhance existing regression-based techniques for process mining or are directly grounded in queueing theory. In particular, our predictors target multi-class service processes, in which requests are classified by a type that influences their processing. Further, we introduce queue mining techniques that derive the predictors from event logs recorded by an information system during process execution. Our evaluation based on large real-world datasets, from the telecommunications and financial sectors, shows that our techniques yield accurate online predictions of case delay and drastically improve over predictors neglecting the queueing perspective.

Abstract: Information systems have been widely adopted to support service processes in various domains, e.g., in the telecommunication, finance, and health sectors. Recently, work on process mining showed how management of these processes, and engineering of supporting systems, can be guided by models extracted from the event logs that are recorded during process operation. In this work, we establish a queueing perspective in operational process mining. We propose to consider queues as first-class citizens and use queueing theory as a basis for queue mining techniques. To demonstrate the value of queue mining, we revisit the specific operational problem of online delay prediction: using event data, we show that queue mining yields accurate online predictions of case delay.

Abstract: We model the decision-making process of callers in call centers as an optimal stopping problem. After each waiting period, a caller decides whether to abandon a call or continue to wait. The utility of a caller is modeled as a function of her waiting cost and reward for service. We use a random-coefficients model to capture the heterogeneity of the callers and estimate the cost and reward parameters of the callers using the data from individual calls made to an Israeli call center. We also conduct a series of counterfactual analyses that explore the effects of changes in service discipline on resulting waiting times and abandonment rates. Our analysis reveals that modeling endogenous caller behavior can be important when major changes (such as a change in service discipline) are implemented and that using a model with an exogenously specified abandonment distribution may be misleading.