Predicting online shopping cart abandonment with machine learning approaches

Abstract

Excessive online shopping cart abandonment rates constitute a major challenge for e-commerce companies and can inhibit their success within their competitive environment. Simultaneously, the emergence of the Internet’s commercial usage results in steadily growing volumes of data about consumers’ online behavior. Thus, data-driven methods are needed to extract valuable knowledge from such big data to automatically identify online shopping cart abandoners. Hence, this contribution analyzes clickstream data of a leading German online retailer comprising 821,048 observations to predict such abandoners by proposing different machine learning approaches. Thereby, we provide methodological insights to gather a comprehensive understanding of the practicability of classification methods in the context of online shopping cart abandonment prediction: our findings indicate that gradient boosting with regularization outperforms the remaining models yielding an F₁-Score of 0.8569 and an AUC value of 0.8182. Nevertheless, as gradient boosting tends to be computationally infeasible, a decision tree or boosted logistic regression may be suitable alternatives, balancing the trade-off between model complexity and prediction accuracy.

Keywords

classification e-commerce machine learning prediction shopping cart abandonment supervised learning

Introduction

To strengthen a company’s position within its competitive environment, marketers need to be able to precisely predict potential customers regarding their purchase and, further, non-purchase behavior. Considering this in the context of online shopping environment, customers frequently place items in their virtual shopping cart for reasons other than immediate purchase. This phenomenon is known as shopping cart abandonment and is particularly apparent in the context of e-commerce; it is the behavioral outcome of consumers placing item(s) in their online shopping cart without making a purchase by completing the checkout process during that online session (Huang et al., 2018; Kukar-Kinney & Close, 2010). Extant literature investigated the behavioral perspective of online shopping cart abandonment by identifying inhibitors to the purchase process: financial risks and concerns about delivery and return policies (Kukar-Kinney & Close, 2010), the usage of shopping carts as organization tools or for entertainment purposes (Kukar-Kinney & Close, 2010), and inhibitors at the checkout stage like perceived transaction inconvenience and privacy intrusion (Rajamma et al., 2009) are—inter alia—the main factors leading to online shopping cart abandonment.

With the spread of the Internet’s commercial usage, the ability to track consumers’ online activities allows companies to collect unbiased information about consumers’ behavior. The detailed records of past usage behaviors comprised by log files and resulting clickstream data can be analyzed by marketers to gain valuable insights. In this context, clickstream data have frequently been modeled to derive implications for website design or advertising efforts (see, for example, Chatterjee et al., 2003 and Montgomery et al., 2004) and further, to predict consumers’ future behaviors, e. g., regarding purchase (see, for example, Bucklin & Sismeiro, 2003 and Moe & Fader, 2004a).

Thus, the antecedents of online shopping cart abandonment are well understood by behavioral literature and clickstream data have been studied by methodological research to analyze consumers’ behavior. The rise of the Internet and the era of big data resulted in an excessive “datafication” (Kelly & Noonan, 2017; Lycett, 2013) of the organizational environment yielding the field of business intelligence comprising data analytics and predictive analytics approaches (Chen et al., 2012). However, despite the richness of clickstream data, prior shopping cart abandonment literature still lacks data-driven methods based on machine learning which make use of this information source to predict such abandoning customers. This might be due to the insufficient awareness of suitable intelligent approaches to extract knowledge from the steadily growing volumes of data (Fayyad et al., 1996).

To address this research gap, we utilize clickstream data of a leading German online retailer to train and subsequently compare different machine learning approaches for the prediction of online shopping cart abandonment (i.e., tree-based methods [more specifically, adaptive boosting, boosted logistic regression, decision tree, gradient boosting with regularization, gradient boosting, gradient boosting with dropout, random forest and stochastic gradient boosting], k-nearest neighbor, naïve bayes, multi-layer perceptron with dropout, and a support vector machine with radial basis kernel). We successfully implement these machine learning methods for online shopping cart abandonment prediction and compare them with logistic regression as a standard non-machine learning benchmark model regarding their predictive performance.

Our article makes several key contributions to the preceding literature. By combining the research fields of both shopping cart abandonment as well as clickstream data analysis with machine learning approaches, we particularly shed light on the practicability of machine learning methods in this application context, as this was neglected by prior research. Furthermore, we provide insights into the characteristics of customers abandoning their shopping cart based on clickstream data that is unsusceptible to self-selection, relatively unobtrusive, and easy to gather. We extensively review literature on classification methods to identify shopping cart abandonments and present validation procedures as well as performance metrics for such methods. Our findings can be useful both for marketing intelligence research by extending the field of machine learning applications in marketing contexts through automatically predicting online shopping cart abandoners and for practitioners to actively prevent such abandonments by several real-time reactions, for example, providing real-time purchase incentives and, moreover, to gain insights into machine learning methods.

The remainder of this article is organized as follows: the subsequent section describes the related work on online shopping cart abandonment and clickstream data. Furthermore, section “Machine learning approaches for classification” summarizes the background on machine learning approaches for classification. Section “Methodology” outlines the methodology comprising a preliminary data analysis and the research design. In sections “Findings” and “Discussion,” we present the findings and discuss both theoretical and practical implications, limitations, as well as directions for future research. Finally, section “Conclusion” draws a conclusion.

Related work

Online shopping cart abandonment

The online shopping cart abandonment phenomenon causes substantial losses of turnover for online retailers (Huang et al., 2018; Rajamma et al., 2009) resulting in a weakened position within their competitive environment. Therefore, extant marketing literature addressed this problem by drawing on a behavioral perspective to identify and understand essential determinants of online shopping cart abandonment; Rajamma et al. (2009) focused on potential inhibitors at the checkout stage and found increased perceived transaction inconvenience (e.g., long registration forms) and high-perceived risk (e.g., perceived security of information asked) to enhance online shopping cart abandonment. Partially, these findings seem to be applicable to new customers which are unfamiliar with the checkout process. Similarly, findings of Kukar-Kinney and Close (2010) indicate that privacy intrusion and security concerns rather lead to the consumers’ decision to buy the product from a stationary offline store. Furthermore, they found the entertainment value of shopping carts, the use of shopping carts as an organization tool, the wait for sale, and the concerns about costs to be antecedents of shopping cart abandonment (Kukar-Kinney & Close, 2010). Their identified determinants were supported by Close and Kukar-Kinney (2010) proving that customers’ tendencies to add items to the online shopping cart for reasons other than immediate purchase are—inter alia—due to organizational purposes. Huang et al. (2018) focused on mobile shopping cart abandonment in their study. They found intrapersonal (i.e., conflicts regarding mobile shopping attributes and low self-efficacy regarding mobile shopping) and interpersonal (i.e., discrepancies from the other’s attitudes to self-attitudes) conflicts to disturb consumers’ emotions during mobile shopping and, in turn, implying shopping cart abandonment. Overall, their findings indicate that the utilized device for online shopping might impact purchase behavior as well. Cho et al. (2006) proved that consumers’ confusion by information overload, high-value consciousness, negative past experiences, intention to conduct price comparisons, and unreliable websites are likely to trigger online shopping cart abandonment.¹

Clickstream data

Drawing on a more holistic perspective of online shopping behavior, further literature shifted away from explanatory behavioral approaches to data-driven methods predicting online purchase behavior in general. Typically, such predictions are based on clickstream data (see, for example, Moe & Fader, 2004a; Sismeiro & Bucklin, 2004; Van den Poel & Buckinx, 2005). Clickstream data model the navigation path a customer takes through the online shop (Montgomery, 2001; Montgomery et al., 2004) and can be extracted from log files which register all requests and information transferred between the customer’s computer and the company’s commercial web server (Bucklin & Sismeiro, 2003).

Examples for using clickstream data to predict online shopping behavior are—inter alia—Moe and Fader (2004a) who proposed a conversion model predicting each customer’s probability of making a purchase based on purchase and visit history. The same authors (Moe & Fader, 2004b) also developed a model for evolving visiting behavior and further, they examined the relationship between visiting frequency and purchasing propensity. They found consumers visiting an e-commerce site more frequently to have a greater propensity to buy (Moe & Fader, 2004b). Van den Poel and Buckinx (2005) predicted purchase behavior and investigated the contribution of different variables: they proved (1) general clickstream variables (i.e., number of days since last visit, and speed of clickstream behavior during last visit), (2) more detailed clickstream variables (i.e., number of accessories [and personal pages and products, respectively] viewed during last visit), (3) demographic variables (i.e., gender and the fact of supplying personal information), and (4) historical purchase behavior (i.e., number of days since last purchase and number of past purchases) to be meaningful predictors. Montgomery et al. (2004) set up different models to predict purchase conversion probability by modeling path information.

Moreover, clickstream data were frequently utilized by research to predict not only purchase behavior but further similar outcome variables. For instance, Bucklin and Sismeiro (2003) investigated drivers affecting the length of time spent viewing a website and the visitor’s decision to continue browsing or to exit the website. Sismeiro and Bucklin (2004) decomposed the purchase process into sequences that must be completed for a purchase to take place (i.e., completion of product configuration, input of personal information, and order confirmation with provision of credit card data) and predicted the probability of completion for each task with covariates of browsing behavior, repeat visitation, use of decision aids, input effort, and information gathering.

Machine learning approaches for classification

Overall, e-commerce as a research subject is suitable for the application of machine learning approaches as proposed by Kohavi and Provost (2001): online retailers can easily and inexpensively collect rich data with respect to the online behavior of customers (i.e., clickstream data) and, further, implement data mining and machine learning applications since political and social barriers are substantially lower than for traditional businesses. Consequently, typical problems for successfully applying machine learning (i.e., the need for a large volume of controlled and reliable data, data with sufficient descriptions, the ability to evaluate results, and to integrate applications successfully) are reduced by the characteristics of e-commerce environment (Kohavi & Provost, 2001).

Machine learning constitutes a new paradigm within data science research and emerged in the course of the artifical intelligence era, which, in turn, was first coined by Samuel (1959) describing it as “the programming of a digital computer to behave it in a way which, if done by human beings [. . .], would be described as involving the process of learning”. In this context, learning may be understood as the automatic search for more useful representations of data regarding a specific task (Chollet & Allaire, 2018). Machine learning algorithms and systems are consequently trained rather than explicitly programmed. During this process, these systems find statistical structure in given examples which are relevant to the task and derive rules for automating the task using guidance from a feedback signal (Cui et al., 2006). Thereby, classification algorithms are types of supervised learning approaches within machine learning, which predict a qualitative response for an observation, that is, they assign an observation to a category (James et al., 2013): Formally, let ${y_{k}, x_{k}}_{k = 1}^{N}$ be a training set, where $y_{k} \in {0, 1, 2, \dots, K - 1}$ is the class membership and $x_{k} = ℝ^{n}$ is the vector of predictor values, then the task is to learn a function to predict the class label $y_{k}$ from $x_{k}$ . Thereby, $K = 2$ in case of binary classification and $K > 2$ in case of multi-class classification tasks.

Drawing on the online shopping cart abandonment problem, the prediction of purchasers and non-purchasers (i.e., customers abandoning their shopping cart) can be considered a binary classification task. Common machine learning approaches for binary classification include—inter alia—tree-based methods, support vector machines, naïve bayes, k-nearest neighbor, and neural networks. The approaches are explained in detail hereinafter.

Tree-based approaches

One of the most common machine learning approaches are tree-based methods which descend from single decision trees, as proposed by Breiman et al. (1984). Basically, decision trees are flowchart-like structures that generate “if-else” rules and thereby allow for prediction of observation classes. Thereby, classification and regression tree models follow a recursive top-down approach in which binary trees aim to partition the predictor space with predictor variables x₁, . . . , x_k into subsets in which the distribution of the dependent variable $y$ is successively more homogeneous (Chipman et al., 1998).

Generally, single decision trees have the advantage of being easy to interpret and to understand (Moro et al., 2014). However, they frequently lead to overfitting, that is, the model learns to identify specific characteristics of the training data which are irrelevant or even obstructive for the classification of unknown data (Friedman, 2001; Srivastava et al., 2014). This results in drawbacks of predictive performance and less expressiveness of the models. Ensemble learning methods that construct several individually trained decision trees and combine their results into a classifier outperforming the single predictions (Opitz & Maclin, 1999; Rokach, 2010) may offer a solution to this problem. In this context, two widely used methods of aggregating trees are boosting and bagging.

In boosting, a family of algorithms converts weak learners (i.e., models that achieve accuracy just above random guessing) to strong learners with a powerful predictive capacity. The idea is to train weak learners sequentially with each weak learner trying to correct its predecessor (Schapire et al., 1998). Thus, each decision tree is built using feedback from previously grown trees (James et al., 2013). Popular boosting algorithms include adaptive boosting “AdaBoost” (Freund & Schapire, 1997), boosted logistic regression “LogitBoost” (Friedman et al., 2000), gradient boosting machines “GB” (Friedman, 2001, 2002), and stochastic gradient boosting “SGB” (Friedman, 2002).² For instance, AdaBoost as a basic boosting algorithm makes predictions by combining the output of weak learners to a weighted sum and putting higher weights on incorrectly classified instances

\hat{y} = s i g n (\sum_{m = 1}^{M} α_{m} h_{m} (x))

with the weak hypothesis $h_{m}$ detected by the weak learner and its importance $α_{m}$ .

In contrast to boosting, bagging (or bootstrap aggregating) grows successive trees independently from earlier trees, that is, each tree is constructed using a bootstrap sample of the data and, hence, a majority vote is taken for prediction (Breiman, 1996). Random forests add an additional layer of randomness to bagging and change how the trees are constructed; in standard decision trees, each node is split using the best split among all predictor variables, whereas, in random forests, the nodes are split using the best among a subset of predictors randomly chosen at that node (Breiman, 2001; Liaw & Wiener, 2002). Due to the recursive structure of tree-based methods, they often capture interaction effects between variables. However, since we focus on the performance of the models and not the importance of specific variables, we will not consider interaction effects further in our study.

Overall, tree-based methods have been found to outperform other established approaches across a variety of different classification tasks such as Internet protocol (IP) traffic flow classification (Williams et al., 2006), customer churn prediction (Vafeiadis et al., 2015), or—similar to our context—prediction of online purchase intention (Bogina et al., 2019; Boroujerdi et al., 2014; Zheng & Liu, 2018). They are particularly favorable since ensemble methods are able to reduce both bias and variance of the single learning algorithms: While individual models may get stuck in local minima, a weighted combination of several different local minima—produced by ensemble methods—are able to minimize the risk of choosing the wrong local minimum (Dietterich, 2002).

Support vector machines

Aside from tree-based methods, support vector machines are powerful tools for classification tasks (James et al., 2013). The basic support vector machine is solving pattern recognition problems by mapping data into a multidimensional input space and constructing an optimal hyperplane that separates the space into homogeneous partitions³ (Cortes & Vapnik, 1995; Vapnik, 1982). Predictions of new instances are then classified into those partitions. The support vector machine aims at constructing a classifier in the form of

\hat{y} = sign [\sum_{i = 1}^{N} α_{k} y_{k} ψ (x, x_{k}) + b]

where $α_{k}$ is positive real constant, $b$ is a real constant, and $ψ (\cdot, \cdot)$ represents the hyperplane (e.g., $ψ (x, x_{k}) = x_{k}^{T} x$ in case of a linear support vector machine; Suykens & Vandewalle, 1999). Aside from the linear case, Boser et al. (1992) proposed a non-linear classifier by applying the so-called kernel trick which allows the algorithm to fit the hyperplane in a transformed feature space.

We used a support vector machine with radial basis kernel for the comparison of machine learning models. However, support vector machines may become computationally infeasible on very large datasets like clickstream data (L’Heureux et al., 2017).

Naїve Bayes

The naїve bayes approach is a basic classifier based on applying the Bayes’ theorem with the naїve assumption that the attributes are conditionally independent (Duda et al., 1973). The classifier assigns a new case to a class label $\hat{y} = C_{k}$ by deriving the maximum a posteriori probability

\hat{y} = \underset{k ϵ {1, \dots, K}}{\arg \max} p (C_{k}) \prod_{i = 1}^{n} p (x_{i} | C_{k})

Naïve bayes as a generative classifier is frequently utilized for classification tasks due to its simplicity, efficiency, and efficacy (Muhammad & Yan, 2015).

K-nearest neighbor

Another basic approach, the k-nearest neighbor algorithm, classifies an observation by a majority vote of the observation’s neighbors (Cover & Hart, 1967). The underlying assumption of the algorithm is that observations which lay closely together within the predictor space (i.e., neighbors) will have the same class label. Thus, the classifier weights the class of the nearest neighbors strikingly high to predict the class label of an unclassified sample (Cover & Hart, 1967). The class is thereby assigned by taking the majority vote of the k nearest neighbors, with k being the number of neighbors that are considered during the classification task. The nearest neighbors are determined with the help of arbitrary distance functions (e.g., Euclidian distance $d (., .)$ ). For new observations $(y, x)$ the nearest neighbor $(y_{(1)}, x_{(1)})$ within the training set is defined by

d (x, x_{(1)}) = \min_{k} (d (x, x_{k}))

and $\hat{y} = y_{(1)}$ —the class of the nearest neighbor—is selected as prediction for $y$ . $x_{(j)}$ and $y_{(j)}$ describe the jth nearest neighbor of $x$ and its class membership $y$ .

K-nearest neighbor as a local learning approach may be suitable for online shopping cart abandonment prediction tasks since it is able to alleviate the challenge of imbalanced data (L’Heureux et al., 2017).

Artificial neural networks

Artificial neural networks are highly parallelized computer systems comprising process units (i.e., neurons) located on process layers with numerous weighted interconnections performing a learning process to create meaningful data representations (Jain et al., 1996). Regarding the concept of deep learning, artificial neural networks may use a number of hidden process layers (the depth of a network) between input and output layer containing non-linear operations in hierarchical architectures to learn characteristics and recognize patterns from given data (Bengio, 2009; Deng, 2011; Hinton et al., 2006). The concept of learning within deep learning (or artificial neural networks, respectively) describes a process of updating the network architecture and the weights of the neuron connections (Jain et al., 1996). To improve the performance, the optimizer is implementing a backpropagation algorithm to minimize the discrepancy between the actual and the target output vector (i.e., the loss score) by adjusting the weights (Rumelhart et al., 1986; Schmidhuber, 2015). To avoid overfitting, a regularization method called dropout can be integrated in the network which randomly sets a share of its output per layer to zero (Srivastava et al., 2014).

Concerning their connection structure (i.e., topology), neural network architectures can be distinguished between feedforward networks (e.g., multi-layer perceptrons; Deng, 2011; Q. Zhang et al., 2018) with neuron connections running to the output layer acyclically and recurrent networks (e.g., long short-term memories; Hochreiter & Schmidhuber, 1997) containing backward connections to build cyclic architectures (Jain et al., 1996; Schmidhuber, 2015). The most commonly used feedforward neural networks—multi-layer perceptrons—can be defined as

\hat{y} = β_{0} + \sum_{h = 1}^{H} β_{h} g (γ_{0 i} + \sum_{i = 1}^{I} γ_{h i} p_{i})

where $I$ denotes the number of inputs $p_{i}$ , $H$ is the number of hidden nodes in the network, the weights $ω = (β, γ)$ with $β = [β_{1}, \dots, β_{H}]$ and $γ = [γ_{11}, \dots, γ_{H I}]$ are for the hidden and output layer respectively, $g (\cdot)$ is the transfer function (e.g., sigmoid logistic), and $β_{0}$ as well as $γ_{0 i}$ are the biases of each node (G. Zhang et al., 1998).

Multi-layer perceptrons were found to outperform other machine learning approaches for purchase intention prediction only after balancing the class distribution with oversampling (Sakar et al., 2019), since deep learning approaches are frequently sensitive to class imbalance (L’Heureux et al., 2017).

Methodology

Preprocessing and preliminary data analysis

The purpose of this study is to predict shopping cart abandonment by making use of machine learning. The machine learning models explained in section “Machine learning approaches for classification” are compared to find the best classifier for this task. The clickstream data were gathered from server log files of a leading German online retailer, which primarily distributes fashion. The data were created by the online retailer through extracting the customers’ chronological online shop activities out of sequential log files. Each log file observation comprised one action or activity (e.g., a click) of a certain customer such as adding a product to the cart or clicking on a product to view its details. Subsequently, each customer’s activities during a session were assigned to summarizing variables. Hence, all activities of a customer were aggregated to one observation with different variables describing the session. Thereby, a session is a period of sustained web browsing or a sequence of the user’s page viewings until the user exits the online shop (Montgomery et al., 2004). The data comprise 3,511,037 observations or sessions between 1 February 2019 and 30 April 2019, that is, 3 months. Furthermore, the data contain 18 explanatory variables for each observation or session listed in Table 1 many of which are consistent with Van den Poel and Buckinx (2005) findings. We are only interested in visitors who made use of the virtual shopping cart during the session, that is, who placed item(s) in their cart. In line with Close and Kukar-Kinney (2010), shopping cart usage is thus defined as necessary precondition for shopping cart abandonment. Thus, we filtered out customers who did not add any items to their shopping cart during the session, so-called just-browsing customers, and 821,048 observations (23.38%) remained. We modeled the dependent variable—shopping cart abandonment—as a dummy variable using the information about the customer’s compiled and ordered shopping carts (variables BASKETS_BB and BASKETS) during the session

Y = {\begin{matrix} 1 if number of compiled shopping carts > 0 number of ordered shopping carts = 0; \\ 0 if number of compiled shopping carts > 0 number of ordered shopping carts > 0 . \end{matrix}

Table 1.

Variables of clickstream data.

Variable	Index	Description
Shopping cart abandonment	SCA	Dependent dummy variable capturing customer’s shopping cart abandonment $Y = {\begin{matrix} 1 & if customer abandoned; \\ 0 & otherwise . \end{matrix}$
Number of ordered shopping carts	BASKETS_BB	Metric predictor variable capturing the number of shopping carts ordered during the customer’s session
Number of compiled shopping carts	BASKETS	Metric predictor variable capturing the number of shopping carts compiled during the customer’s session
Number of logins	LOGS	Metric predictor variable capturing the number of logins during the customer’s session
Number of existing customers’ logins to the second step of the ordering process	LOGS_CUST_STEP2	Metric predictor variable capturing the number of logins of existing customers to the second step of the purchasing process during the customer’s session
Number of new customers’ logins to the second step of the ordering process	LOGS_NEWCUST_STEP2	Metric predictor variable capturing the number of logins of new customers to the second step of the purchasing process during the customer’s session
Number of overall page viewings	PIS	Metric predictor variable capturing the number of overall page viewings during the customer’s session
Number of shopping cart page viewings	PIS_AP	Metric predictor variable capturing the number of shopping carts page viewings during the customer’s session
Number of detailed product page viewings	PIS_DV	Metric predictor variable capturing the number of detailed product page viewings during the customer’s session
Number of category overview page viewings	PIS_PL	Metric predictor variable capturing the number of category overview page viewings (i.e., all products within a category) during the customer’s session
Number of department page viewings	PIS_SHOPS	Metric predictor variable capturing the number of department page viewings (i.e., all categories within a department) during the customer’s session
Number of detailed product page viewings using search function	PIS_SDV	Metric predictor variable capturing the number of detailed product page viewings after using the search function during the customer’s session
Number of search results page viewings	PIS_SR	Metric predictor variable capturing the number of overall search results page viewings during the customer’s session
Number of product types in shopping cart	POSITIONS	Metric predictor variable capturing different product types in the shopping cart during the customer’s session
Number of items in shopping cart	QUANTITY	Metric predictor variable capturing the number of items in the shopping cart during the customer’s session
Value of ordered shopping carts	VALUE_BB	Metric predictor variable capturing the value of shopping carts ordered during the customer’s session
New customer	NEW_CUST	Predictor dummy variable capturing new customers $X_{16} = {\begin{matrix} 1 & if new customer; \\ 0 & otherwise . \end{matrix}$
Accessing online shop through desktop	WEB_CUST	Predictor dummy variable capturing customers that access the online shop through desktop $X_{17} = {\begin{matrix} 1 & if accessing through desktop; \\ 0 & otherwise . \end{matrix}$
Accessing online shop through mobile phone	MOBILE_CUST	Predictor dummy variable capturing customers that access the online shop through mobile phone $X_{18} = {\begin{matrix} 1 & if accessing through mobile phone; \\ 0 & otherwise . \end{matrix}$

BASKETS: Number of Carts Compiled; LOGS: Number of Logins; LOGS_CUST_STEP2: Number of Existing Customers’ Logins to the Second Step of the Ordering Process; LOGS_NEWCUST_STEP2: Number of New Customers’ Logins to the Second Step of the Ordering Process; MOBILE_CUST: Customer Accessing through Mobile Phone; NEW_CUST: New Customer; PIS: Number of Overall Page Viewings; PIS_AP: Number of Shopping Cart Page Viewings; PIS_DV: Number of Detailed Product Page Viewings; PIS_PL: Number of Category Overview Page Viewings; PIS_SDV: Number of Detailed Product Page Viewings Using Search Function; PIS_SHOPS: Number of Department Page Viewings; PIS_SR: Number of Search Results Page Viewings; POSITIONS: Number of Product Types; QUANTITY: Number of Items; WEB_CUST: Customer Accessing through Desktop.

Our data contain 520,653 (63.41%) observations of shopping cart abandonments (or non-purchasers, respectively) and 300,395 (36.59%) observations of purchasers. Hence, the dataset is relatively balanced. We excluded the variable for the number of ordered shopping carts (BASKETS_BB) and the value of ordered shopping carts (VALUE_BB) further for prediction.⁴

Figure 1 illustrates the relationship between the page viewing and login variables by demonstrating the customer’s clickstream in the online shop; the customer typically starts browsing departments (PIS_SHOPS), then selects a certain category within a department (PIS_PL), and further, chooses a certain product within a category (PIS_DV). Optionally, the customer uses the shop’s search engine (PIS_SR) to look systematically for a specific product (PIS_SDV). To make a purchase, the customer can either directly sign in (LOGS) or check the items in the shopping cart (PIS_AP) first and then sign in and hence, proceed to the second step of the purchasing process (LOGS_CUST_STEP2 or LOGS_NEWCUST_STEP2). However, signing in to the second step of the purchasing process does not necessarily lead to a purchase of the customer.

Nevertheless, with respect to the descriptive statistics in Table 2, we find that existing customers (or new customers, respectively) which subsequently make a purchase sign in to the second step of the ordering process approximately 5.93 times (or 4.46 times, respectively) more often than non-purchasers. Generally, purchasers sign in more often (1.03 logins on average) than non-purchasers (0.93 logins on average). This might indicate that the cause for shopping cart abandonment frequently occurs before the customer proceeds to the checkout stage.

Table 2.

Descriptive statistics of clickstream data.

Variable	Observations of shopping cart abandonments (n = 520,653)					Observations of purchasers (n = 300,395)
Variable	Mean	SD	Median	Min	Max	Mean	SD	Median	Min	Max
BASKETS	0.99	0.11	1	0	2	1.09	0.40	1	0	49
LOGS	0.93	0.27	1	0	2	1.03	0.21	1	0	2
LOGS_CUST_ STEP2	0.06	0.23	0	0	1	0.32	0.47	0	0	1
LOGS_NEWCUST_ STEP2	0.02	0.13	0	0	1	0.07	0.26	0	0	1
PIS	22.26	27.47	12	1	513	46.45	37.06	35	2	593
PIS_AP	1.05	2.17	0	0	71	3.06	3.42	2	0	57
PIS_DV	3.42	7.48	0	0	200	6.53	9.76	3	0	203
PIS_PL	3.99	11.37	0	0	279	8.75	17.00	1	0	315
PIS_SHOPS	7.68	17.55	1	0	405	15.87	25.14	6	0	396
PIS_SDV	1.40	3.92	0	0	142	3.13	5.46	1	0	127
PIS_SR	2.82	7.48	0	0	222	5.71	10.18	2	0	208
POSITIONS	2.88	3.31	2	1	66	3.38	3.31	2	1	111
QUANTITY	2.95	3.55	2	1	143	3.48	3.49	2	1	143
	Counts		Proportion (%)			Counts		Proportion (%)
NEW_CUST	48,839		9.38			15,387		5.12
WEB_CUST	214,455		41.29			171,789		57.19
MOBILE_CUST	238,694		45.85			84,401		28.1

Furthermore, the number of purchasers’ overall page viewings is 2.09 times higher than of non-purchasers on average. Overall, customers abandoning their shopping cart browse less pages than purchasers—regardless of the pages’ type. Particularly, the median reveals that there are significant differences regarding the number of page viewings between purchasers and abandoners: the median of abandoners’ overall page viewings is 12, 1 for department viewings, and 0 for all other types of page viewings. In contrast, purchasers’ median for overall page viewings is 35, 6 for department viewings, and for example, 2 for shopping cart viewings.

On average, purchasers add more items and different product types (3.48 and 3.38, respectively) to their shopping cart than non-purchasers (2.95 and 2.88, respectively).

There is a larger absolute (48,839) and relative (9.38%) proportion of new customers among the observations of shopping cart abandonments than among those making a purchase (15,387 observations or 5.12%, respectively). Moreover, there is a larger proportion of mobile shoppers among customers abandoning their shopping cart (45.85%) compared to the observations of purchasers (28.1%). The latter descriptive findings are consistent with the results of preceding (behavioral) research: for example, as argued earlier, Huang et al. (2018) proved that online shopping cart abandonment occurs more frequently for customers using a mobile device due to high emotional ambivalence. Moe and Fader (2004a) found that—among new customers—online conversion rate is lower as purchasing thresholds and perceived risks are high for unexperienced visitors.

Experimental setup

Since each machine learning approach and its subsequent refinements and modifications exhibit individual strengths and weaknesses in dependence of the underlying data and the requested task, it is highly recommended in the machine learning literature to compare and test different algorithms (Moro et al., 2014; Razi & Athappilly, 2005). Thus, we compared different models of those proposed in section “Machine learning approaches for classification” to predict shopping cart abandonment for our data, listed in Table 3. In addition, we included a standard logistic regression model in our comparison serving as a non-machine learning benchmark method.

Table 3.

Machine learning approaches for comparison.

Approach	Description
Adaptive boosting (AdaBoost)	Ensemble of weak learners, algorithm puts higher weights on incorrectly classified instances
Boosted logistic regression (LogitBoost)	Algorithm applies logistic regression techniques to the AdaBoost method by minimizing the logistic loss
Decision tree (DT)	Algorithm recursively partitions the predictor space into subsets in which the distribution of the dependent variable is successively more homogeneous
Gradient boosting (linear base learner) with L1 and L2 regularization (GBReg)	Ensemble of weak learners (with linear base learners), algorithm applies L1 (Lasso Regression), and L2 (Ridge Regression) regularization
Gradient boosting (tree base learner) (GBTree)	Ensemble of weak learners (with tree base learners), algorithm minimizes the model’s loss by adding weak learners sequentially using a gradient descent like procedure
Gradient boosting (tree base learner) with dropout (GBDropout)	See GBTree, but the algorithm randomly drops boosting tree members
k-Nearest neighbor (KNN)	Algorithm classifies an observation by assigning it to the class most common among its $k$ nearest neighbors
Multi-layer perceptron network with dropout (MLPDropout)	Feedforward neural network with dropout regularization technique
Naïve bayes (NB)	Algorithm is based on the Bayes’ theorem and classifies an observation by deriving the maximum a posteriori probability
Random forest (RF)	Ensemble of decision trees, algorithm predicts new data by aggregating the predictions of the trees
Stochastic gradient boosting (SGB)	Algorithm fits base learner at each iteration on the subsample of the data—instead of the full—drawn at random without replacement
Support vector machine with radial basis kernel (SVMRadial)	Support vector machine implementation with radial basis kernel

To estimate and, hence, validate the models, we randomly partitioned the data into a training and a test subset in a 67/33 ratio, that is, 67% (or 550,098 observations, respectively) of the data are used as training data and 33% (or 270,950 observations, respectively) are used as test data.

We performed $k$ -fold cross-validation with the training data to fit the models and optimized their hyperparameters, respectively (Geisser, 1975; Stone, 1974): the sample, that is, the training data, is randomly split into $k$ equal sized subsamples $D_{1}, D_{2}, . . ., D_{k}$ . Of the $k$ subsamples, one single subsample is retained as validation data to test the fitted model subsequently and the remaining $k - 1$ subsamples are used as training data to fit the model. This step is repeated $k$ times with each of the $k$ subsamples serving as validation data once. Drawing on machine learning literature, $k = 10$ is frequently utilized since it provides an adequate trade-off between method’s variance and method’s bias (i.e., trade-off between the estimated parameter’s expected value and the estimated value) (Bradley, 1997; Breiman, 1996; Kohavi, 1995; Tibshirani & Tibshirani, 2009; P. Zhang, 1993). Thus, we applied 10-fold cross-validation.

Furthermore, to validate and evaluate our models’ performance, we considered different performance metrics that indicate the models’ predictive ability. In a binary decision problem, the classifier labels observations as either positive or negative. Consequently, the classification procedure yields four different outputs in a $2 \times 2$ confusion matrix: the sample is either correctly classified as positive (true positive [TP]), correctly classified as negative (true negative [TN]), falsely classified as positive (false positive [FP] or Type II error), or falsely classified as negative (false negative [FN] or Type I error). Thereby, accuracy is one of the most commonly used measures for classification performance due to its simplicity (see, for example, Kohavi, 1995). It is the ratio between correctly classified samples to the total number of samples

Accuracy = \frac{TP + TN}{P + N}

However, recent research shifted away from solely presenting accuracy results since accuracy assumes balanced class distribution and equal error costs (i.e., Type I errors are equivalent to Type II errors) which is rarely the case in real world applications (Davis & Goadrich, 2006; Provost & Fawcett, 1997). To address these problems, a receiver operating characteristics (ROC) curve and thus, the area under the ROC curve (AUC)⁵ has been increasingly used by the machine learning community since they are insensitive to changes in class distributions and scale-invariant (Bradley, 1997; Fawcett, 2006). A ROC graph is a two-dimensional depiction of classification performance to measure different classifiers’ performances and captures the trade-off between benefits (i.e., true positives) and costs (i.e., false positives) (Fawcett, 2006). It is created by plotting the true positive rate (TPR) (or sensitivity or recall, respectively) against the false positive rate (FPR) (or $1 - specificity$ , respectively) (Bradley, 1997; Fawcett, 2006; Hand, 2009; Provost & Fawcett, 2001)

TPR = Sensitivity = Recall = \frac{TP}{P}; FPR = 1 - Specificity = \frac{FP}{N}; Specificity = \frac{TN}{N}

The classifier’s AUC value is a portion of the area of the unit square and its value ranges from 0.0 to 1.0 (perfect classification). It should be higher than 0.5 which equals the AUC of an uninformative classifier (Bradley, 1997; Fawcett, 2006). An important statistical property of the AUC is that a classifier’s AUC is equivalent to the probability that the classifier will rank a randomly chosen positive observation higher than a randomly chosen negative observation (Fawcett, 2006).

An alternate performance measure is the $F_{1}$ -Score comprising both precision and recall

Precision = \frac{TP}{TP + FP}; F_{1} = 2 \cdot \frac{Precision \cdot Recall}{Precision + Recall}

Ideally, the performance measure is chosen by properly reflecting the investigation’s aims to avoid misleading conclusions. Since our data are relatively balanced, it seems reasonable to consider accuracy as a basic performance metric. However, as we intend to convert customers abandoning their shopping carts into purchasers our main aim is to correctly classify actual positives (i.e., observations of shopping cart abandonments) by minimizing the Type I error. Consequently, the higher the recall the less false negatives (i.e., shopping cart abandonments classified as purchasers) have been predicted. Besides, we intend to maximize the proportion of actual positives among the predicted positives by minimizing the Type II error, that is, purchasing customers should not be classified as non-purchasers. Thus, the higher the precision the less false positives have been predicted. The $F_{1}$ -Score considers the trade-off between recall and precision. Therefore, we determined the $F_{1}$ -Score, recall, and precision as our main performance metrics for the test data. In addition, to yield valid results, we considered the ROC curve or the AUC, respectively, as a performance metric since it is a common measure of separability capturing the trade-off between both TPR (or sensitivity or recall respectively, analogous to $F_{1}$ -Score) and FPR (i.e., how many negative instances are falsely classified as positive among the negative instances). For the training data, the best classifier during hyperparameter optimization was automatically chosen based on the AUC values.

Although prediction accuracy (i.e., AUC, F1-Score, or accuracy) is frequently the main decision criterion when comparing different machine learning models, the models’ complexity in terms of computation time and computation effort (e.g., numbers of hyperparameters to be optimized) is of similar importance regarding the application in practice and should therefore be considered as well (Doshi-Velez & Kim, 2017; Guidotti et al., 2019; Tambe et al., 2019).

Findings

Drawing on the training results in Table 4, gradient boosting with regularization outperformed the remaining approaches with an AUC of 0.9008. The final gradient boosting model’s fitted hyperparameters did not include the lasso regression technique (L1 regularization) but made use of the ridge regression technique (L2 regularization). The gradient boosting with tree base learners and random forest yielded comparable results (AUC of 0.8953 and 0.8954, respectively) whereas naïve bayes and boosted logistic regression realized the lowest AUC values (0.8218 and 0.8381, respectively).

Table 4.

Training data results.

Model	Fitted parameters		AUC	Estimation time (s)^a
Model	Parameter	Fitted value	AUC	Estimation time (s)^a
Logistic regression			0.8003	20.3
AdaBoost	Number of trees	50	0.8698	703,903.9
AdaBoost	Method	Adaboost.M1	0.8698	703,903.9
LogitBoost	Number of boosting iterations	21	0.8381	380.0
DT	Complexity parameter	0.0129	0.7988	225.07
GBReg	Number of boosting iterations	150	0.9008	4,021.28
	L2 regularization	0.1
	L1 regularization	0
	Learning rate	0.3
GBTree	Number of boosting iterations	150	0.8953	6,701.14
	Maximum tree depth	3
	Shrinkage	0.4
	Minimum loss reduction	0
	Subsample ratio of columns	0.8
	Minimum sum of instance weight	1
	Subsample percentage	1
GBDropout	Number of boosting iterations	150	0.8952	49,794.27
	Maximum tree depth	3
	Shrinkage	0.4
	Minimum loss reduction	0
	Subsample ratio of columns	0.8
	Minimum sum of instance weight	1
	Subsample percentage	0.75
	Fraction of trees dropped	0.01
	Probability of skipping dropout	0.95
KNN	Maximum number of neighbors	30	0.8828	127,773.4
	Distance	2
	Kernel	Optimal
MLPDropout	Number of hidden units	768	0.8807	218,894.0
	Dropout rate	0.35
	Batch size	64
	Learning rate	0.000006
	Rho	0.2
	Learning rate decay	0
	Activation function	Sigmoid
	Epochs	30
NB	Laplace correction	0	0.8218	5,757.49
	Distribution type	Kernel Density Estimation
	Bandwidth adjustment	0.3
RF	Number of randomly selected predictors	14	0.8954	171,587.7
	Splitting rule	Gini
	Minimal node size	35
SGB	Number of boosting iterations	150	0.8800	2,033.17
	Maximum tree depth	3
	Shrinkage	0.1
	Minimum terminal node size	10
SVMRadial	Sigma	0.1818	0.8808	1,306,838.6
SVMRadial	Cost	0.5	0.8808	1,306,838.6

The highest AUC value is marked in bold. AdaBoost: Adaptive Boosting; DT: Decision Tree; GBDropout: Gradient Boosting with Dropout; GBReg: Gradient Boosting with L1 and L2 Regularization; GBTree: Gradient Boosting with Tree Base Learners; KNN: k-Nearest Neighbor; LogitBoost: Boosted Logistic Regression; MLPDropout: Multi-Layer Perceptron Network with Dropout; NB: Naïve Bayes; RF: Random Forest; SGB: Stochastic Gradient Boosting; SVMRadial: Support Vector Machine with Radial Basis Kernel.

With 40 GB RAM.

Regarding estimation time, the benchmark logistic regression, decision tree and boosted logistic regression performed the fastest 10-fold cross validation to optimize the hyperparameters (20.3, 225.07, and 380.0 s, respectively). The support vector machine and adaptive boosting were the most time-consuming models to estimate (1,306,838.6 and 703,903.9 s, respectively). Gradient boosting with regularization yielded a moderate estimation time (4,021.28 s) and thus, provides an appropriate trade-off between AUC and estimation time.

Since we are rather interested in the fitted models’ performances on new and unknown data, the test data result in Table 5 exhibit a higher practical relevance than the preceding results: similarly to the training data results, the gradient boosting model with regularization was superior to the remaining models regarding the test data. It yielded the best AUC (0.8182) and accuracy (82.29%) results. In line with these findings, the $F_{1}$ -Score (0.8569) proves that the model is the most suitable approach in our comparison to balance the trade-off between precision and recall. With respect to its confusion matrix in Appendix 1, the gradient boosting model classified 28,209 abandonments falsely as purchasers (16.42% of all abandonments) and 19,767 purchasers as abandonments, respectively (19.94% of all purchasers). This is further reflected by the model’s precision (0.8790) and recall (0.8358), that is, there is a high proportion of both correctly predicted abandonments among all correctly and falsely predicted abandonments (87.90%) and correctly predicted abandonments among all actual abandonments (83.58%).

Table 5.

Test data results.

Model	Performance metrics
Model	AUC	Accuracy	Precision	Recall	$F_{1}$ -Score
Logistic regression	0.8012	78.94%	0.6677	0.8454	0.7461
AdaBoost	0.7516	78.54%	0.8024	0.8777	0.8384
LogitBoost	0.7623	77.19%	0.8349	0.7981	0.8161
DT	0.7741	74.78%	0.9015	0.6762	0.7728
GBReg	0.8182	82.29%	0.8790	0.8358	0.8569
GBTree	0.8105	81.78%	0.8701	0.8377	0.8536
GBDropout	0.8123	81.84%	0.8731	0.8350	0.8536
KNN	0.7962	80.5%	0.8585	0.8290	0.8435
MLPDropout	0.7911	80.36%	0.8503	0.8378	0.8440
NB	0.5022	63.56%	0.6351	0.9996	0.7767
RF	0.8108	81.75%	0.8711	0.8359	0.8531
SGB	0.7902	80.08%	0.8521	0.8299	0.8409
SVMRadial	0.7956	81.23%	0.8479	0.8578	0.8528

For each column, the highest value is marked in bold. AdaBoost: Adaptive Boosting; DT: Decision Tree; GBDropout: Gradient Boosting with Dropout; GBReg: Gradient Boosting with L1 and L2 Regularization; GBTree: Gradient Boosting with Tree Base Learners; KNN: k-Nearest Neighbor; LogitBoost: Boosted Logistic Regression; MLPDropout: Multi-Layer Perceptron Network with Dropout; NB: Naïve Bayes; RF: Random Forest; SGB: Stochastic Gradient Boosting; SVMRadial: Support Vector Machine with Radial Basis Kernel.

Although naïve bayes realized an extremely high recall (0.9996), its precision (0.6351) is just slightly better than random guessing. This is due to its negligible Type I error (i.e., 68 abandonments classified as purchasers [0.0004% of all abandonments]) and its substantial Type II error (i.e., 98,677 purchasers classified as abandonments [99.52% of all purchasers]). Consequently, by focusing exclusively either on precision or recall, one could draw misleading conclusions regarding model selection. The $F_{1}$ -Score of the naïve bayes model (0.7767) reveals that it constitutes a suboptimal choice.

Similarly, albeit the decision tree classified a high proportion of purchasers correctly and only 12,688 (i.e., 12.80% of all purchasers) wrong, it categorized 55,634 cart abandonments as purchasers (i.e., 32.38% of all abandonments). Thus, due to its high Type I error, its recall is extremely low (0.6762), but it realized the highest precision value of all models (0.9015).

Generally, our results indicate a substantial predictive ability of the most tree-based methods (i.e., gradient boosting with regularization [and linear base learners], gradient boosting [with tree base learners], gradient boosting with dropout [and tree base learners], and random forest) compared with the remaining machine learning approaches. The latter were outperformed by tree-based models with regard to all relevant performance metrics (AUC, accuracy, and $F_{1}$ -Score).⁶

Logistic regression as a non-machine learning benchmark approach yielded the lowest $F_{1}$ -Score but realized a higher AUC value than several other machine learning approaches like boosted logistic regression, k-nearest neighbor, multi-layer perceptron, naïve bayes and support vector machine. Nevertheless, it did not perform better than the tree-based methods (except for adaptive boosting, decision tree, and stochastic gradient boosting) with regard to AUC.

Moreover, the k-nearest neighbor algorithm as a basic machine learning approach outperformed more sophisticated algorithms like the multi-layer perceptron, the stochastic gradient boosting, and adaptive boosting with respect to its AUC value (0.7962).

Discussion

Our findings contribute to a deeper understanding regarding the successful implementation of machine learning methods for predicting online shopping cart abandoners with a strong forecast performance to apply marketing techniques in real-time to convert them to purchasers. Thus, we discuss our findings’ theoretical contribution and practical implications in this section. We also discuss limitations and propose suggestions for future research.

Theoretical contribution

Overall, we fill a research gap by identifying suitable machine learning approaches for online shopping cart abandonment prediction not only in terms of accuracy but, further, in terms of practicability. Thereby, we contribute to literature in several ways. First, we are able to characterize customers abandoning their shopping cart descriptively with our data. Preceding literature on shopping cart abandonment (e.g., Close & Kukar-Kinney, 2010; Huang et al., 2018; Kukar-Kinney & Close, 2010) primarily shed light on behavioral aspects of the abandonment process with experimental designs. In contrast, our research deals with unbiased clickstream data comprising an exceptionally high number of observations. Our data indicate that there is a higher proportion of new customers and mobile shoppers among customers abandoning their shopping carts compared to purchasers, whereas the latter add more items to their shopping cart and view an increased number of pages on average.

Second, we contribute to literature by proposing a broad range of machine learning models to compare their performance regarding online shopping cart abandonment prediction and, thus, to predict future customers abandoning their shopping carts in real-time. Prior literature either drew on a behavioral perspective to understand the antecedents of shopping cart abandonment or predicted—more generally—purchase behaviors with conservative approaches and less observations (see, for example, Huang et al., 2018; Kukar-Kinney & Close, 2010; or Sismeiro & Bucklin, 2004). For our data, the gradient boosting with regularization yielded the highest accuracy (82.29%). However, with respect to our main aim, to minimize the Type I error (i.e., abandoners falsely classified as purchasers) and the Type II error (i.e., purchasers falsely classified as abandoners), we focused on the $F_{1}$ -Score capturing the trade-off between precision and recall. Consistent with the accuracy results, the gradient boosting with regularization outperformed the remaining models regarding the $F_{1}$ -Score (0.8569). In addition, it realized the highest AUC value (0.8182) compared to the other models.

Overall, we found tree-based methods to be superior to the remaining machine learning approaches and logistic regression as a benchmark non-machine learning approach aligning with prior research comparing machine learning approaches in different application fields like customer churn prediction or phishing detection (Abu-Nimeh et al., 2007; Caruana & Niculescu-Mizil, 2006; Vafeiadis et al., 2015) and—similar to our context—prediction of online purchase intention (Bogina et al., 2019; Boroujerdi et al., 2014; Zheng & Liu, 2018). Thus, we complement the literature on machine learning comparisons in a marketing context.

Moreover, despite the striking importance of prediction accuracy as a decision criterion for appropriate machine learning approaches, the models’ practicability with respect to modeling complexity as an essential criterion is of particular importance (Doshi-Velez & Kim, 2017; Guidotti et al., 2019; Tambe et al., 2019) but, at the same time, is often neglected by current research. Thus, we considered the models’ complexity in terms of computation time and computation effort (e.g., numbers of hyperparameters to optimize) to add to literature. Thereby, the decision tree approach and boosted logistic regression yielded only slightly worse AUC results compared to gradient boosting with regularization and, simultaneously, their complexity in terms of both computation effort and time was rather low. Hence, in case of online shopping cart abandonment prediction, a decision tree model and boosted logistic regression perform well in balancing the trade-off between accuracy and complexity. Furthermore, as stated by prior literature, we found the support vector machine approach to be extremely computationally infeasible (L’Heureux et al., 2017) despite its acceptable prediction accuracy.

Practical implications

Our research may help to gather a comprehensive understanding of machine learning approaches for prediction or classification, particularly with regard to online shopping cart abandonment prediction. More specifically, our research provides multifold practical implications for decision makers.

Since research about advanced machine learning approaches in marketing contexts is still in its infancy (e.g., Cheung et al., 2003 and Cui et al., 2006), we reviewed relevant literature to provide an introduction to such models, its potential applications, as well as performance metrics, and common methods for validation: for machine learning models, $k$ -fold cross-validation is a common method to optimize the models’ hyperparameters. Decision makers should draw on either recall as a performance measure if their main aim is to correctly classify abandonments or precision if they intend to avoid falsely classified purchasers. The $F_{1}$ -Score considers the trade-off between both. Besides, the AUC is a common measure of separability since it is insensitive to skewed class distributions. Overall, tree-based approaches and particularly boosting methods are superior to the remaining machine learning models regarding forecast accuracy within online shopping cart abandonment prediction. Random forest yields comparable results but is rather time-consuming to estimate (171,587.7 s estimation time). The support vector machine and adaptive boosting are computationally intensive with estimation times of 1,306,838.6 and 703,903.9 s, respectively.

Aside from pointing out methodological aspects, we drew on an economical perspective to enhance an organization’s turnover; with regard to our data, the mean value of purchasers’ ordered shopping carts (VALUE_BB) is €271.73 and they add 3.479 items into their shopping cart on average and thus, we expect the online retailer’s sales loss for each shopping cart abandonment to be around €230 with 2.945 items in their shopping cart on average. Therefore, we determined a suitable approach to correctly identify shopping cart abandonments as well as purchasers; our findings indicate that gradient boosting with regularization outperformed the remaining approaches. Organizations can implement this method to predict non-purchasers in real-time when a sufficient amount of information about the customer’s activities during the session has been collected. Overall, we found particularly tree-based machine learning approaches such as random forest or gradient boosting to outperform traditional classification approaches such as logistic regression and decision tree, which are frequently utilized by practitioners.

Drawing on an overall practicability perspective, decision makers may take a slight loss in prediction accuracy into account if, instead, the model’s complexity in terms of computation time and effort is substantially lower: in our application context, decision tree and boosted logistic regression yielded acceptable prediction results and their computation effort was substantially lower compared to gradient boosting methods.

Limitations and future research

Our research is subject to limitations which stimulate further research. First, the set of useful variables for prediction was limited. With respect to extant literature (see, for example, Bucklin & Sismeiro, 2003; Moe & Fader, 2004a; Van den Poel & Buckinx, 2005), we expect, for example, demographic variables, historical purchase behavior, or the time customers spend on the single pages to be informative variables. Furthermore, we did not have information about the customers’ identity and thus, could not determine whether there were recurring customers. However, this information could be of great interest for analyzing online behavior and predicting shopping cart abandonment. For instance, Huang et al. (2018) anticipated that some customers might use the mobile phone for initial purchase stages (i.e., browsing and collecting information) and then switch to the computer for completing the purchase. However, such customers are listed as two distinct sessions in the current data. Another missing information concerns the value of abandoned shopping carts. While there is a variable that indicates the value of ordered carts (i.e., VALUE_BB), the value of abandoned carts can only be estimated. In line with extant literature on shopping cart abandonment (e.g., Close & Kukar-Kinney, 2010; Kukar-Kinney & Close, 2010), it can be assumed that the value of ordered items influences abandoning rates and, thus, could aid the prediction of such. Moreover, if detailed information about spent time and further, the chronological order of customers’ actions in the online shop would be available, we could decompose the session into sequences or segments. Then, we could determine a critical point in the customer’s session in which abandonment can be predicted reliably with the $F_{1}$ -Score or the AUC exceeding a defined threshold (see, for example, Sismeiro & Bucklin, 2004). Hence, future research could replicate this study with more detailed data, for example, between-site clickstream data (i.e., panel data collected by media measurement company), that are typically more comprehensive and frequently used in clickstream analyses (see, for example, Moe & Fader, 2004a).

Second, we excluded just-browsing customers from our investigation. A possible direction for future research could be to conduct a multi-class classification by differentiating between purchasers, abandonments, and just-browsing customers, similar to the cluster analysis of Moe (2003).

Third, the models’ performance strongly depends on the optimized hyperparameters which may be a time-consuming procedure for some of the models. Therefore, we considered only a limited range of possible hyperparameter values. Moreover, other values of $k$ in cross-validation could lead to different results.

Finally, a real-time implementation requires a certain amount of data to be collected before the model can make a reliable decision.

By implementing these models, companies may detect shopping cart abandoners in real-time and, subsequently, convert some of them into purchasers by making use of targeted marketing measures such as individual chat pop-ups, coupons or special discounts. For instance, Close and Kukar-Kinney (2010) suggest human-human interactions (i.e., live chats with employees or other online shoppers) to avoid shopping cart abandonment. These could pop-up on the website if the online user is predicted to abandon by the machine learning model. Therefore, future research is recommended to test whether pop-up messages and offers impact customers’ online shopping behavior and can prevent online shopping cart abandonment.

Conclusion

Online shopping cart abandonment can inhibit corporate growth and hence, harm a company’s success within its competitive environment. Simultaneously, the emergence of the Internet’s commercial usage leads to the ability to track consumers’ online activities and online behavior resulting in clickstream data.

Thus, to identify online shopping cart abandoners by extracting valuable knowledge from such clickstream data, we proposed different machine learning approaches. We analyzed data of a German online retailer comprising 821,048 observations and fitted the models using 10-fold cross validation. Thereby, our article contributes to extant literature by combining research fields of both online shopping cart abandonment and clickstream data with machine learning approaches.

Our data indicate that among customers abandoning their shopping carts there is a higher proportion of new customers and mobile shoppers compared to purchasers, whereas the latter add more items to their shopping cart and have a higher number of page viewings on average. Moreover, our comparison results prove that gradient boosting with regularization is a suitable method to distinguish between abandonments and purchasers yielding an AUC of 0.8182, an $F_{1}$ -Score of 0.8569, and an accuracy of 82.29%. Nevertheless, a decision tree or boosted logistic regression may be suitable alternatives yielding only slightly less accurate prediction results and being computationally more feasible.

Nevertheless, research on clickstream data combined with machine learning approaches is still in its infancy—particularly in a marketing context. Thereby, machine learning will be inevitable for e-commerce businesses to be successful in the long-term and the analysis provided in this article shall stimulate further research on this topic.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Theresa Maria Rausch

Nicholas Daniel Derra

Notes

Appendix

Appendix 1:

confusion matrices

Model	Prediction	Actual
Model	Prediction	0 (Purchaser)	1 (Abandonment)
Logistic regression	0 (Purchaser)	83,817	41,722
Logistic regression	1 (Abandonment)	15,335	130,076
AdaBoost	0 (Purchaser)	62,009	21,005
AdaBoost	1 (Abandonment)	37,143	150,793
LogitBoost	0 (Purchaser)	72,036	34,692
LogitBoost	1 (Abandonment)	27,116	137,106
DT	0 (Purchaser)	86,464	55,634
DT	1 (Abandonment)	12,688	116,164
GBReg	0 (Purchaser)	79,385	28,209
GBReg	1 (Abandonment)	19,767	143,589
GBTree	0 (Purchaser)	77,662	27,875
GBTree	1 (Abandonment)	21,490	143,923
GBDropout	0 (Purchaser)	78,294	28,352
GBDropout	1 (Abandonment)	20,858	143,446
KNN	0 (Purchaser)	75,687	29,383
KNN	1 (Abandonment)	23,465	142,415
MLPDropout	0 (Purchaser)	73,803	27,869
MLPDropout	1 (Abandonment)	25,349	143,929
NB	0 (Purchaser)	475	68
NB	1 (Abandonment)	98,677	171,730
RF	0 (Purchaser)	77,903	28,197
RF	1 (Abandonment)	21,249	143,601
SGB	0 (Purchaser)	74,409	29,217
SGB	1 (Abandonment)	24,743	142,581
SVMRadial	0 (Purchaser)	72,724	24,427
SVMRadial	1 (Abandonment)	26,428	147,371

AdaBoost: Adaptive Boosting; DT: Decision Tree; GBDropout: Gradient Boosting with Dropout; GBReg: Gradient Boosting with L1 and L2 Regularization; GBTree: Gradient Boosting with Tree Base Learners; KNN: k-Nearest Neighbor; LogitBoost: Boosted Logistic Regression; MLPDropout: Multi-Layer Perceptron Network with Dropout; NB: Naïve Bayes; RF: Random Forest; SGB: Stochastic Gradient Boosting; SVMRadial: Support Vector Machine with Radial Basis Kernel.

References

Abu-Nimeh

Nappa

Wang

Nair

(2007, October). A comparison of machine learning techniques for phishing detection. In Proceedings of the Anti-Phishing Working Groups 2nd Annual eCrime Researchers Summit (pp. 60–69). ACM.

Bengio

(2009). Learning deep architectures for AI. Now Publishers.

Bogina

Mokryn

Kuflik

(2019). Will this session end with a purchase? Inferring current purchase intent of anonymous visitors. Electronic Commerce Research and Applications, 34, Article 100836.

Boroujerdi

E. G.

Mehri

Sadeghi Garmaroudi

Pezeshki

Mehrabadi

F. R.

Malakouti

Khadivi

(2014, May 27–29). A study on prediction of user’s tendency toward purchases in websites based on behavior models [Conference session]. 2014 6th Conference on Information and Knowledge Technology (IKT), Shahrood, Iran.

Boser

B. E.

Guyon

I. M.

Vapnik

V. N.

(1992, July). A training algorithm for optimal margin classifiers. In Proceedings of the Fifth Annual Workshop on Computational Learning Theory (pp. 144–152). ACM.

Bradley

A. P.

(1997). The use of the area under the ROC curve in the evaluation of machine learning algorithms. Pattern Recognition, 30(7), 1145–1159.

Breiman

(1996). Bagging predictors. Machine Learning, 24(2), 123–140.

Breiman

(2001). Random Forests. Machine Learning, 45(1), 5–32.

Breiman

Friedman

Olshen

Stone

(1984). Classification and regression trees. Chapman & Hall.

10.

Bucklin

R. E.

Sismeiro

(2003). A model of web site browsing behavior estimated on clickstream data. Journal of Marketing Research, 40(3), 249–267.

11.

Caruana

Niculescu-Mizil

(2006). An empirical comparison of supervised learning algorithms. In Cohen

(Ed.), Proceedings of the 23rd international conference on Machine learning (pp. 161–168). ACM.

12.

Chatterjee

Hoffman

D. L.

Novak

T. P.

(2003). Modeling the clickstream: Implications for web-based advertising efforts. Marketing Science, 22(4), 520–541.

13.

Chen

Chiang

R. H. L.

Story

V. C.

(2012). Business intelligence and analytics: From big data to big impact. MIS Quarterly, 36(4), 1165–1188.

14.

Cheung

K.-W.

Kwok

J. T.

Law

M. H.

Tsui

K.-C.

(2003). Mining customer product ratings for personalized marketing. Decision Support Systems, 35(2), 231–243.

15.

Chipman

H. A.

George

E. I.

McCulloch

R. E.

(1998). Bayesian CART model search. Journal of the American Statistical Association, 93(443), 935–948.

16.

Cho

C.-H.

Kang

Cheon

H. J.

(2006). Online shopping hesitation. Cyberpsychology & Behavior, 9(3), 261–274.

17.

Chollet

Allaire

J. J.

(2018). Deep learning with R. Manning.

18.

Climent

Momparler

Carmona

(2019). Anticipating bank distress in the Eurozone: An extreme gradient boosting approach. Journal of Business Research, 101, 885–896.

19.

A. G.

Kukar-Kinney

(2010). Beyond buying: Motivations behind consumers’ online shopping cart use. Journal of Business Research, 63(9–10), 986–992.

20.

Cortes

Vapnik

(1995). Support-vector networks. Machine Learning, 20(3), 273–297.

21.

Cover

Hart

(1967). Nearest neighbor pattern classification. IEEE Transactions on Information Theory, 13(1), 21–27.

22.

Cui

Wong

M. L.

Lui

H.-K.

(2006). Machine learning for direct marketing response models: Bayesian networks with evolutionary programming. Management Science, 52(4), 597–612.

23.

Davis

Goadrich

(2006, June). The relationship between precision-recall and ROC curves. In Proceedings of the 23rd International Conference on Machine learning (pp. 233–240). ACM.

24.

Deng

(2011, October). An overview of deep-structured learning for information processing. In Proceedings of Asian-Pacific Signal & Information Processing Annual Summit and Conference (APSIPA-ASC) (pp. 1–14). NIPS.

25.

Dietterich

T. G.

(2002). Ensemble learning. In Arbib

M. A.

(Ed.), The handbook of brain theory and neural networks (pp. 110–125). The MIT Press.

26.

Doshi-Velez

Kim

(2017). Towards a rigorous science of interpretable machine learning. arXiv: 1702.08608.

27.

Duda

R. O.

Hart

P. E.

Stork

D. G.

(1973). Pattern classification and scene analysis. Wiley.

28.

Fawcett

(2006). An introduction to ROC analysis. Pattern Recognition Letters, 27(8), 861–874.

29.

Fayyad

Piatetsky-Shapiro

Smyth

(1996). From data mining to knowledge discovery in databases. AI Magazine, 17(3), 37–54.

30.

Freund

Schapire

R. E.

(1997). A decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 55, 119–139.

31.

Friedman

(2001). Greedy function approximation: A gradient boosting machine. The Annals of Statistics, 29(5), 1189–1232.

32.

Friedman

(2002). Stochastic gradient boosting. Computational Statistics & Data Analysis, 38(4), 367–378.

33.

Friedman

Hastie

T. H.

Tibshirani

(2000). Additive logistic regression: A statistical view of boosting. The Annals of Statistics, 28(2), 337–407.

34.

Geisser

(1975). The predictive sample reuse method with applications. Journal of the American Statistical Association, 70(350), 320–328.

35.

Guidotti

Monreale

Ruggieri

Turini

Giannotti

Pedreschi

(2019). A survey of methods for explaining black box models. ACM Computing Surveys, 51(5), 1–42.

36.

Hand

D. J.

(2009). Measuring classifier performance: A coherent alternative to the area under the ROC curve. Machine Learning, 77(1), 103–123.

37.

Hinton

G. E.

Osindero

Teh

Y.-W.

(2006). A fast learning algorithm for deep belief nets. Neural Computation, 18(7), 1527–1554.

38.

Hochreiter

Schmidhuber

(1997). Long short-term memory. Neural Computation, 9(8), 1735–1780.

39.

Huang

G.-H.

Korfiatis

Chang

C.-T.

(2018). Mobile shopping cart abandonment: The roles of conflicts, ambivalence, and hesitation. Journal of Business Research, 85, 165–174.

40.

Jain

A. K.

Mao

Mohiuddin

K. M.

(1996). Artificial neural networks: A tutorial. Computer, 29(3), 31–44.

41.

James

Witten

Hastie

Tibshirani

(2013). An introduction to statistical learning. Springer.

42.

Kelly

Noonan

(2017). The doing of datafication (and what this doing does): Practices of edification and the enactment of new forms of sociality in the Indian public health service. Journal of the Association for Information Systems, 18(2), 872–899.

43.

Kohavi

(1995). A study of cross-validation and bootstrap for accuracy estimation and model selection. In International Joint Conference on Artificial Intelligence (IJCAI), 2, 1137–1145.

44.

Kohavi

Provost

(2001). Applications of data mining to electronic commerce. In Provost

Kohavi

(Eds.), Applications of data mining to electronic commerce (pp. 5–10). Springer.

45.

Kukar-Kinney

A. G.

(2010). The determinants of consumers’ online shopping cart abandonment. Journal of the Academy of Marketing Science, 38(2), 240–250.

46.

L’Heureux

Grolinger

Elyamany

H. F.

Capretz

M. A. M.

(2017). Machine learning with big data: Challenges and approaches. IEEE Access, 5, 7776–7797.

47.

Liaw

Wiener

(2002). Classification and regression by random forest. R News, 2(3), 18–22.

48.

Lycett

(2013). “Datafication”: Making sense of (big) data in a complex world. European Journal of Information Systems, 22(4), 381–386.

49.

Moe

W. W.

(2003). Buying, searching, or browsing: Differentiating between online shoppers using in-store navigational clickstream. Journal of Consumer Psychology, 13(1–2), 29–39.

50.

Moe

W. W.

Fader

P. S.

(2004a). Capturing evolving visit behavior in clickstream data. Journal of Interactive Marketing, 18(1), 5–19.

51.

Moe

W. W.

Fader

P. S.

(2004b). Dynamic conversion behavior at E-commerce sites. Management Science, 50(3), 326–335.

52.

Montgomery

A. L.

(2001). Applying quantitative marketing techniques to the Internet. Interfaces, 31(2), 90–108.

53.

Montgomery

A. L.

Srinivasan

Liechty

J. C.

(2004). Modeling online browsing and path analysis using clickstream data. Marketing Science, 23(4), 579–595.

54.

Moro

Cortez

Rita

(2014). A data-driven approach to predict the success of bank telemarketing. Decision Support Systems, 62, 22–31.

55.

Muhammad

Yan

(2015). Supervised machine learning approaches: A survey. ICTACT Journal on Soft Computing, 5(03), 946–952.

56.

Opitz

Maclin

(1999). Popular ensemble methods: An empirical study. Journal of Artificial Intelligence Research, 11, 169–198.

57.

Provost

Fawcett

(1997). Analysis and visualization of classifier performance: Comparison under imprecise class and cost distributions. In Proceeding of the 3rd International Conference on Knowledge Discovery and Data Mining (pp. 43–48). ACM.

58.

Provost

Fawcett

(2001). Robust classification for imprecise environments. Machine Learning, 42(3), 203–231.

59.

Rajamma

R. K.

Paswan

A. K.

Hossain

M. M.

(2009). Why do shoppers abandon shopping cart? Perceived waiting time, risk, and transaction inconvenience. Journal of Product & Brand Management, 18(3), 188–197.

60.

Razi

M. A.

Athappilly

(2005). A comparative predictive analysis of neural networks (NNs), nonlinear regression and classification and regression tree (CART) models. Expert Systems with Applications, 29(1), 65–74.

61.

Rokach

(2010). Ensemble-based classifiers. Artificial Intelligence Review, 33(1–2), 1–39.

62.

Rumelhart

D. E.

Hinton

G. E.

Williams

R. J.

(1986). Learning representations by back-propagating errors. Nature, 323(6088), 533–536.

63.

Sakar

C. O.

Polat

S. O.

Katircioglu

Kastro

(2019). Real-time prediction of online shoppers’ purchasing intention using multilayer perceptron and LSTM recurrent neural networks. Neural Computing and Applications, 31(10), 6893–6908.

64.

Samuel

A. L.

(1959). Some studies in machine learning using the game of checkers. IBM Journal of Research and Development, 3(3), 210–229.

65.

Schapire

R. E.

Freund

Bartlett

Lee

W. S.

(1998). Boosting the margin: A new explanation for the effectiveness of voting methods. The Annals of Statistics, 26(5), 1651–1686.

66.

Schmidhuber

(2015). Deep learning in neural networks: An overview. Neural Networks, 61, 85–117.

67.

Sismeiro

Bucklin

R. E.

(2004). Modeling purchase behavior at an E-commerce web site: A task-completion approach. Journal of Marketing Research, 41(3), 306–323.

68.

Srivastava

Hinton

Krizhevsky

Sutskever

Salakhutdinov

(2014). Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15, 1929–1958.

69.

Stone

(1974). Cross-validatory choice and assessment of statistical predictions. Journal of the Royal Statistical Society: Series B (Methodological), 36(2), 111–133.

70.

Suykens

J. A. K.

Vandewalle

(1999). Least squares support vector machine classifiers. Neural Processing Letters, 9(3), 293–300.

71.

Tambe

Cappelli

Yakubovich

(2019). Artificial intelligence in human resources management: Challenges and a path forward. California Management Review, 61(4), 15–42.

72.

Tibshirani

R. J.

Tibshirani

(2009). A bias correction for the minimum error rate in cross-validation. The Annals of Applied Statistics, 3(2), 822–829.

73.

Vafeiadis

Diamantaras

K. I.

Sarigiannidis

Chatzisavvas

(2015). A comparison of machine learning techniques for customer churn prediction. Simulation Modelling Practice and Theory, 55, 1–9.

74.

Van den Poel

Buckinx

. (2005). Predicting online-purchasing behaviour. European Journal of Operational Research, 166(2), 557–575.

75.

Vapnik

V. N.

(1982). Estimation of dependences based on empirical data. Springer.

76.

Williams

Zander

Armitage

(2006). A preliminary performance comparison of five machine learning algorithms for practical IP traffic flow classification. ACM SIGCOMM Computer Communication Review, 36(5), 5–16.

77.

Zhang

Patuwo

B. E.

M. Y.

(1998). Forecasting with artificial neural networks: The state of the art. International Journal of Forecasting, 14(1), 35–62.

78.

Zhang

(1993). Model selection via multifold cross validation. The Annals of Statistics, 21(1), 299–313.

79.

Zhang

Yang

L. T.

Chen

(2018). A survey on deep learning for big data. Information Fusion, 42, 146–157.

80.

Zheng

Liu

(2018, January 12–14). A scalable purchase intention prediction system using extreme gradient boosting machines with browsing content entropy [Conference session]. 2018 IEEE International Conference on Consumer Electronics (ICCE), Las Vegas, NV, United States.