A prescriptive analytics framework for efficient E-commerce order delivery
Introduction
E-commerce is contributing significantly to retail sales. In 2019, online retailing accounted for 14.1% of the retail sales, and this figure is estimated to reach 22% by 2023 [1]. Except for digital goods, all other products sold on e-commerce platforms require physical delivery. The last-mile delivery of goods, in the context of business-to-customer retailing, includes several challenges [2]. In fact, physical delivery is the most polluting, least efficient part of the logistics [3]. One of the problems with the physical delivery of goods is the presence of failed deliveries. Since failed order deliveries are reattempted, they incur additional costs in package handling, fuel, and emissions [3]. On the other hand, successful deliveries increase customer satisfaction and decrease the chance of product returns [4]. Therefore, it is highly desirable for logistics operators and platform owners to increase successful deliveries to make the delivery process efficient, and increase customer satisfaction.
The primary reason for failed deliveries is the absence of customers when deliveries are attempted [5]. Several last-mile delivery models aim to reduce missed deliveries. One possible solution is to deliver goods to pick-up points and instruct customers to collect them. Though this solution offers the potential to reduce delivery failures, it also inconveniences the customers because they have to physically visit the pick-up points [6]. It is found that the convenience offered by home delivery is a significant driver of e-commerce purchases [7]; therefore, this solution can negatively impact sales. The other solution is to deliver goods on pre-agreed time windows. Although this solution holds promise in increasing successful deliveries, time-windows are generally not offered for ordinary deliveries (goods other than grocery, furniture, heavy appliances) [8]. One reason is the associated difficulty in knowing the exact delivery date a priori. Deliveries often lead or lag the estimated time of arrival (ETA) [9], making it challenging to offer time slots. In the current e-commerce setting where home-delivery is the preferred delivery option [10,11] and offering delivery time-windows in advance is difficult, knowing the appropriate time periods for attended home delivery would be helpful to minimize missed deliveries and improve customer satisfaction.
Predictive models can help platforms learn appropriate delivery time-windows. Learning appropriate time windows allows platforms to plan delivery schedules that result in successful deliveries. Also, identifying orders that are likely to be missed allows companies to take proactive remedial measures. For instance, customers can be contacted before attempting the delivery to confirm their willingness to receive orders. Overall, predictive models can be leveraged in schedule planning to decrease delivery failures, reduce operational costs, and increase customer satisfaction.
So far, a few studies [[12], [13], [14], [15]] proposed data-driven solutions for increasing delivery successes; however, the approaches suffer from drawbacks. At the prediction task level, they employ sensitive customer-specific information such as half-hourly electricity consumption data, historical GPS coordinates, and smart home device data. These data sources are not readily available, and customers do not readily share them [[16], [17], [18]]. Also, there are legal issues concerning the security and privacy of customers. In fact, two of the studies [12,14] acknowledge the legal issues, and one study [14] acknowledges the non-availability of data sources and calls for research on using other kinds of data. Our work explores less sensitive, readily available data sources for building predictive models. We use alternate data sources such as aggregated location data and historical order delivery data that are widely available and propose novel predictors for building predictive models. At the schedule generation level, prior approaches integrate predictions using methods that take exponential time for computation. Though these methods are useful for smaller instances, e-commerce delivery hubs deliver hundreds of orders in a day making exponential time methods inadaptable. Our work builds on the available construction heuristics, which takes polynomial time for computation. Though heuristics usage reduces solution quality, the schedules can be generated in minutes against hours (or days).
Our approach is based on the fact that the success of a delivery attempt depends on both the order characteristics [3] and the location characteristics of a service region [19]. For instance, deliveries directed to office and business locations are more likely to succeed in the afternoon hours than deliveries directed to residential areas. Order features such as delivery-speed, value, and category can suggest the customers' willingness to wait. For example, customers choosing same-day or one-day delivery options are less willing to wait [9]; hence they may make arrangements for order reception resulting in high delivery success. Similarly, several order-related features and location-related features can become helpful in predicting delivery success.
In this paper, our first aim is to demonstrate that a service region's location features and order features can be used to predict the success or failure of a delivery attempt. To this end, we train several machine learning models on order delivery data supplied by our industry collaborator and location data extracted from Here technologies' location platform [20] and compare their predictive performance. Our second aim is to design a decision support framework that leverages delivery success predictive models and generates delivery schedules to service customers in the most appropriate time intervals. Accordingly, we propose a two-step decision support framework. In the first step, the framework uses machine learning models to generate order success profiles. Order success profiles map the probability of delivery success to the time of delivery. In the second step, for each order, the framework infers time windows appropriate for the delivery and uses the Vehicle Routing Problem with Time Windows (VRPTW) optimization scheme to generate the delivery schedules.
Our framework is developed in collaboration with one of the largest e-commerce platforms in India. Currently, the delivery agents deliver orders in the sequence that results in the shortest tour distance. In contrast, our framework generates schedules identifying the most appropriate time for order deliveries, thereby increasing the chance of delivery success. We establish the validity and generality of the approach by applying it in two diverse delivery hubs and comparing the operational costs and the required delivery attempts with that of the current process in practice. Specifically, we evaluate our approach against the baseline policy, where delivery schedules are optimized to produce the shortest tour through simulation experiments. Compared to the baseline policy, our approach reduces the delivery attempts and operational costs for both the hubs. The paper's contributions are as follows:
- •
It demonstrates the usage of diverse data sources such as e-commerce order delivery data and location data for building delivery success prediction models.
- •
It is the first paper to propose order success profiles and to define a procedure to generate them.
- •
It presents a practical decision support framework that generates appropriate time windows for order deliveries and uses them to develop schedules.
The remainder of this paper is organized as follows. In Section 2, we review the related literature. Section 3 describes the proposed decision support framework. The case study, data utilized, and the experimental settings are described in Section 4. The results of both predictive modeling and schedule generation are discussed in Section 5. In Section 6, we discuss our findings and provide concluding remarks.
Section snippets
Traditional order fulfillment methods
Reception boxes, controlled access systems, collection & delivery points (CDPs), attended home delivery (AHD) with time slot management are traditionally proposed in the literature to reduce missed deliveries [21]. Reception boxes are similar to letterboxes installed at customers' houses to accommodate order deliveries [2]. Since deliveries can happen irrespective of customers' presence, this solution reduces failed deliveries. However, due to the challenges it poses to business models and the
Proposed decision support framework
The proposed framework is intended to generate data-driven delivery schedules that could reduce failed deliveries and delivery costs. Fig. 1 illustrates our predictive and prescriptive analytics framework. We train several machine learning models and employ the best performing model in our framework. As a part of the training, data from two sources (historical order delivery data and location data) are utilized. The location data is captured from an external source; hence latitude and longitude
Case study
The proposed framework is intended to generate data-driven delivery schedules that could reduce failed deliveries and delivery costs. For this study, we consider an e-commerce delivery hub that fulfills customers' orders in a 9-h delivery shift. As a part of the fulfillment process, the delivery hub receives orders from the warehouse in the early hours; orders with premium service levels (1-day, 2-day deliveries) may arrive throughout the day. The delivery manager generates delivery routes or
Results for the delivery success prediction
Table 4, Table 5, Table 6, Table 7, Table 8 summarize the predictive modeling results. The first two tables report the AUC scores, from which we observe that the scores are reasonable and similar for all the models. This result suggests that models have good average performance in discriminating between delivery success and failure. However, to identify the best classifier, we employ statistical tests. For both hubs A & B, the observed differences between various classifiers are significantly
Discussion and conclusion
Predicting order delivery success is beneficial for e-commerce companies. Companies can plan deliveries accordingly, thereby increase successful deliveries. Though prior attempts were made, as discussed in Section 2, the approaches require sensitive customer-specific information that is not widely available. In this paper, we proposed a predictive and prescriptive analytics framework that schedules order deliveries in the most appropriate time periods. During the development, we generated novel
Credit author statement
The authors do not wish to include any author contribution statement for the paper.
Acknowledgments
This work was supported by the Indo-Dutch Grant [grant number 13(4)/2018-CC&BT] provided by the Ministry of Electronics & IT (MeitY), India, in collaboration with the Netherlands Organization for Scientific Research (NWO).
We thank Flipkart India Pvt. Ltd. for providing necessary data for the research work. In particular, we thank Chandrasekhar K, Shailendra Kumar, Sreeparna Chatterjee, Muthusamy Chelliah for providing advice and sharing the data.
Shanthan Kandula is a doctoral candidate of Information Systems at the Indian Institute of Management Ahmedabad. He holds a bachelor of technology degree in Electrical & Electronics Engineering. His research interests are in e-commerce, business analytics, classical machine learning, natural language processing, and deep reinforcement learning. He also has IS practitioner experience working at banking firms.
References (62)
- et al.
The role of physical distribution services as determinants of product returns in Internet retailing
J. Oper. Manag.
(2014) - et al.
Acceptability of collection and delivery points from consumers’ perspective: a qualitative case study of Christchurch city
Case Stud. Transport Policy.
(2017) - et al.
The delivery problem: optimizing hit rates in e-commerce deliveries
Transp. Res. B Methodol.
(2018) - et al.
Improving home delivery efficiency by using principles of address intelligence for B2C deliveries
Transport. Res. Proc.
(2016) - et al.
Analytics meets port logistics: a decision support system for container stacking operations
Decis. Support. Syst.
(2019) - et al.
Employees recruitment: a prescriptive analytics approach via machine learning and mathematical programming
Decis. Support. Syst.
(2020) - et al.
An investigation of bankruptcy prediction in imbalanced datasets
Decis. Support. Syst.
(2018) A decision support framework for robust R&D budget allocation using machine learning and optimization
Decis. Support. Syst.
(2019)- et al.
Ensemble machine learning models for aviation incident risk prediction
Decis. Support. Syst.
(2019) - et al.
A survey on machine learning for data fusion
Inform. Fusion
(2020)
Feature selection with multi-view data: a survey
Inform. Fusion
Technical analysis and sentiment embeddings for market trend prediction
Expert Syst. Appl.
Choosing the proper autoencoder for feature fusion based on data complexity and classifiers: analysis, tips and guidelines
Inform. Fusion
Deep learning in business analytics and operations research: models, applications and managerial implications
Eur. J. Oper. Res.
Study of the impact of resampling methods for contrast pattern based classifiers in imbalanced databases
Neurocomputing.
E-commerce Share of Total Global Retail Sales from 2015 to 2023
Solving the last mile issue: reception box or delivery box?
Int. J. Phys. Distrib. Logist. Manag.
Characteristics and typology of last-mile logistics from an innovation perspective in an urban context
Simulation of B2C e-commerce distribution in Antwerp using cargo bikes and delivery points
Eur. Transp. Res. Rev.
On risk, convenience, and internet shopping behavior
Commun. ACM
E-commerce in Hungary: a market analysis
Theory Methodol. Pract.
Commonly used e-commerce supply chains for fast moving consumer goods: comparison and suggestions for improvement
Int J Log Res Appl
Delivery service, customer satisfaction and repurchase: evidence from an online retail platform
SSRN Electron. J.
Unlocking the failed delivery problem? Opportunities and challenges for smart locks from a consumer perspective
Res. Transp. Econ.
Using customer-related data to enhance e-grocery home delivery
Ind. Manag. Data Syst.
Efficient parcel delivery by predicting customers’ locations*
Decis. Sci.
Smart home devices and B2C E-commerce: A way to reduced failed deliveries
Location Data Sharing: What are the Concerns Around Privacy?, Geospatial World
The Trust Opportunity: Exploring Consumer Attitudes to the Internet of Things
The Consumer-data Opportunity and the Privacy Imperative
Here Places API, Here Developer Documentation
Cited by (20)
Efficient hybrid oversampling and intelligent undersampling for imbalanced big data classification
2024, Expert Systems with ApplicationsMore to tip, or tip more? Examining consumers' preservice tipping behavior in the on-demand supermarket delivery context
2024, Decision Support SystemsStochastic modeling of integrated order fulfillment processes with delivery time promise: Order picking, batching, and last-mile delivery
2024, European Journal of Operational ResearchIs online-to-offline customer care support essential for consumer service?
2023, Journal of Retailing and Consumer ServicesA data-driven decision support system for service completion prediction in last mile logistics
2023, Transportation Research Part A: Policy and PracticeIOT data Fusion framework for e-commerce
2022, Measurement: SensorsCitation Excerpt :Furthermore, the regularity of the consumer survey was modest, making it difficult to respond more effectively to the experimental requirements. Forensic investigators must apply evidence collecting techniques and technology in every category of digital forensics and may develop new Internet-specific investigative procedures to make use of the vast and varied data acquired & maintained by the omnipresent Internet of Things services [9]. Many problems remain outstanding although this study produced many theoretical process models to address the special qualities of the Internet of Things [18].
Shanthan Kandula is a doctoral candidate of Information Systems at the Indian Institute of Management Ahmedabad. He holds a bachelor of technology degree in Electrical & Electronics Engineering. His research interests are in e-commerce, business analytics, classical machine learning, natural language processing, and deep reinforcement learning. He also has IS practitioner experience working at banking firms.
Srikumar Krishnamoorthy is currently a faculty in the Information Systems Area at the Indian Institute of Management Ahmedabad, India. His key research interests include data mining, text mining, social media analytics, and personalization in e-commerce. He received his doctorate in IT and Systems Management from the Indian Institute of Management Lucknow, India.
Debjit Roy is a professor in production and quantitative methods at the Indian Institute of Management in Ahmedabad, India. His research interests are to estimate the performance of logistical and service systems, such as container terminals, automated distribution centers, vehicle rental, trucking, and restaurant systems. He holds a Ph.D. in Industrial Engineering (with a major in Decision Sciences/Operations Research and a minor in Computer Science) and an MS in Manufacturing Systems Engineering from the University of Wisconsin-Madison.