Unsupervised Modeling of E-Customers’ Profiles: Multiple Correspondence Analysis with Hierarchical Clustering of Principal Components and Machine Learning Classifiers

1. Introduction

1.1. Background and Rationale

E-commerce is developing faster than ever, with an increase in the number of transactions [1,2]. This global trend allows retailers to expand their reach with as many customers as possible [3]. Additionally, the COVID-19 pandemic has further accelerated the change in consumer habits, with many consumers switching to online shopping [4,5] posing new challenges to e-tailers [6,7,8]. With the rise of online consumers, e-tailers are forced to improve customer experience throughout the purchasing process [9,10]. Similar trends are visible in the Serbian market, where e-shopping is becoming pronounced, emphasising the importance of e-commerce in modern business.

According to Ranđelović [11], despite the numerous benefits of e-commerce, consumers in Serbia still need to accept this form of commerce. One of the reasons for the slow adoption of e-commerce is consumer distrust, especially regarding the security of transactions [12]. Still, in 2023, the e-commerce market in Serbia grew by 34.5%, reaching a value of 955.7 million dollars, while it is predicted that by 2027, the market will grow to 1.65 billion dollars. The growth in e-commerce users is estimated to reach 4.36 million customers by 2027, making up 62.5% of Internet users in Serbia [13]. Regardless of these positive trends, Serbia faces challenges such as the preference for cash on delivery payments and the continued lack of trust in the security of online transactions [14].

Demographic data, such as gender, age, income and education, are vital in analysing consumer preferences and behaviour [15,16]. These variables enable market segmentation and adaptation of marketing strategies to the specific needs of different consumers [17]. For example, income often indicates purchasing power and can help target marketing campaigns to different socio-economic segments [18]. Education and age influence how consumers use technology and how they behave when shopping online [19]. Understanding these demographic factors allows e-tailers to identify target groups and optimise customer experience [20]. In Serbia, where e-commerce is growing, insight into demographic variables can help overcome obstacles and improve consumer acceptance of e-commerce. This provides valuable information for adjusting e-tailers' market strategies and improving user experience.

1.2. Literature Review

When analysing user profiles in e-commerce and the influence of demographics, limited studies include supervised or unsupervised learning models. Previous studies primarily rely on user descriptive profiles and inferential statistics for hypothesis testing [1,16,21,22], while others extend the analysis with multivariate regression [23] and factor models [24,25,26] whereas item answers are used as continuous values assuming a normal distribution. In addition, we have identified a sample of studies that use advanced mathematical modelling of e-customer profiles. For example, Hristoski, I. et al. [27] Using customer behaviour model graphs (CBMGs), identify 12 typologies of online shoppers. The CBMGs work as an N×N transitional probability matrix that denotes relative frequencies of invoking specific e-commerce functions. These frequencies are then used to map a unique transitional probability matrix of users with similar behaviour. Next, Swarnakar, P. et al. [28] use Logistic Regression (LR) and Artificial Neural Networks (ANN) to predict online purchase behaviours to help e-retailers develop suitable strategies. They first perform a statistical analysis of demographics and user preferences to identify relevant factors affecting online behaviour before using LR and ANN to predict user behaviour. In a vice versa approach, Chen, X. et al. [29] used deep learning algorithms to predict user demographics based on the multisource user characteristics and preferences data.

On the other hand, some have tried identifying user profiles through cluster modelling. For instance, Bellini, P. et al. [30] perform k-means clustering to identify user profiles based on demographics and preferences. The results show that after identifying different clusters of user profiles, a stimulus was generated for the customers, which increased buyers' purchases by 3.48%. The Kuruba, Y. et al. [31] Customer segmentation was successfully performed by relying on distributed clustering models. Lastly, a study by Hørlück, J. et al. [32] The study fails to identify clusters of user behaviour using an unsupervised Gaussian Mixture Model with Hierarchical Probabilistic Clustering. It neither falsifies nor supports the hypothesis of identifying clear user profiles based on buying behaviour.

Furthermore, understanding that demographics play an important role in shaping user profiles, we provide a brief overview of demographic variables affecting user preferences. The literature reports that e-customers between the ages of 18 and 50 are the most active [33] with young adults being prevalent [34,35], mainly driven by convenience and interactivity [36]. Older adults tend to value product selection and post-consumption for repurchases [37], while younger adult. The studies on gender suggest an inconclusive contribution [38,39,40], reporting both significant and non-significant findings. However, some suggest that women are “shopping for fun” while males emphasise “quick shopping” behaviours [41]. Additionally, Naseri & Elliott [42] suggests that education significantly influences the adoption of online shopping [43,44,45,46], while also income status [47,48] plays an important role, where both income and education tend to have a higher probability of online shopping [49,50]. Lastly, we have identified that employment (work status) also determines purchase behaviours [51,52,53]. A more comprehensive list of demographics that may be of interest to the reader is described elsewhere [54,55].

The major challenges still exist. Although proposed typologies, clusters or classes of user profiles can be successfully mapped into different user profiles, the biggest issue is that most clusters and user profiles overlap and are not mutually exclusive. Also, from previous studies, there is a lack of evidence and methods describing the actual contribution of each individual variable (features) in describing contribution or shared variance. In this study, we first delve into unsupervised modelling of user profiles for obtaining cluster labels; after this, we use Machine Learning (ML) classification algorithms to predict user-profiles and identify the most important features.

1.3. Aims and Objectives

Considering the research problem's contextual settings, the study's primary idea is to switch from a traditional statistical exploratory analysis to an ML approach in identifying user profiles of e-customers. Firstly, we examine how demographic factors influence the online behaviour of e-customers in Serbia. Specifically, we are interested in how the demographics above affect four selected user behaviour factors: Purchase frequency (PURFREQ), Most important property when buying the first time (MIPB1T), Most important property before repeating the purchase (MIPBREP), and Reasons for quitting the online purchase (RFQ) using a hypothetical framework.

However, given that statistical testing highlights significant relationships between categories, it needs to provide an understanding of how these categories interact. As a response, we perform multiple correspondence analysis with hierarchical clustering on principal components (MCA-HCPC) as an unsupervised model for labelling user profiles. Secondly, after obtaining class categories, i.e., labels, we perform ML classification using a dataset from user profiles that includes demographics and user preferences. Lastly, we use the best-performing algorithms to allocate the most important features using Mean Dropout Loss (MDL) from the obtained classification results.

The rest of the study is structured as follows. The second section provides an in-depth description of the survey used for the study. In addition, we provide rigorous data analysis procedures, starting from a priori sample determination, study selection and proposition of hypotheses in the study. The third section provides information about descriptives, including demographics and user behaviour variables. Next, we provide significant and non-significant results to ensure the transparency and replicability of the findings. Also, given that existing association tests fail to provide exact interaction between categories, we perform MCA to provide a deeper understanding of the association between variables, including all variables in the study and reduction of variables to only those reported as statistically significant. The study provides a discussion section that informs about the findings obtained, and finally, it provides concluding remarks, limitations, and implications of the study in the last section.

2. Materials and Methods

2.1. Multistage Model of Data Workflow Framework

Data workflow framework consists of three stages (Figure 1). The first stage explains a priori sample sample determination, survey development and data collection. Also, this stage explains research hypothesis framework for the reduction of a raw dataset to a dataset comprising only of statistically significant association. Lastly, the first stage explains an in-depth statistical analysis regarding the statistically significant association between demographics and user preferences.

The second stage explains Multiple Correspondence Analysis (MCA) procedure for reducing dimensionality of a dataset into two dimensional space. Next, the study extracts relevant features by including only top two principal components that explain the most variation, i.e., inerta. Also, the Agglomerative Hierarchical Clustering (AHC) is performed on MCA’s principal components for obtaining clusters of user preferences. Following the procedure of identifying class labels, i.e., user profiles, we merge the class label vector with raw dataset.

The third segment proposes and explains utilised ML classifiers, i.e., ML algorithms, setting the parameters and loss functions used for the analysis. The second part of the third stage describes main classification metrics used for evaluating the model performance and ML classification performance metrics – Accuracy, Precision, Recall, F1-score and AUROC. Finally, the last part of the stage includes extraction of the most relevant features using Permutation Feature Importance (PFI) [56] that builds upon Mean Dropout Loss (MDL) metric [57].

2.2. Data Collection and Sample Size

The research was realised during the four weeks of November-December 2022 using two strategies for securing a representative sample. The survey was created on the Google Forms platform and distributed exclusively online. The target group was adult citizens of the Republic of Serbia who had experience ordering goods and durable products via the Internet and delivering them to a specific location. The survey included respondents with experience delivering goods such as electrical appliances, clothing, footwear, furniture, tools, small home and yard use items, sports equipment, and similar items. Before participating in the survey, respondents gave their consent to participate. They were informed about the research objectives and that their answers would be anonymous and analysed in groups. They were informed that they could stop participating at any time.

A priori sample size was determined as follows. Firstly, we conduct sample size estimation to estimate the minimum required sample size for performing χ² test statistics. To do so, we used G*Power (v.3.1.9.6) for calculating sample size per parameters: Effect size w = 0.3 (medium effect), α = 0.05, Power (1-β error probability) = 0.80 and df = 36 (determined as the most of 7 categories per variables j and k, such that df = (j-1)(k-1)). The output statistic shows non-centrality parameter λ = 26.46, with χ²_crit = 50.998, actual Power (1-β error probability) = 0.801, and minimum sample size of n = 294. Secondly, we rely on sample size calculation per Hamburg [58] – a commonly used sample size calculator can be found online (e.g., https://www.calculator.net/). The minimum required representative sample is n = 385. Lastly, given the sample size from previous similar studies [41,46] was higher than required, we have also stopped data collection at n = 906 respondents.

2.3. Research Hypothesis Framework

The research hypothesis framework (Figure 2) is designed to test the conditional dependencies, i.e., the existence of an association between different demographic properties of e-consumers and user preferences (and behaviour). To do so, we included variables primarily reported in previous studies [33,59] regarding demographic properties, e.g., gender, work status, age group, education, place of residence, and income. Next, we test the association with PURFREQ (Purchase frequency), MIPB1T (Most important property when buying the first time), MIPBREP (Most important property before repeating the purchase), and RFQ (Reasons for quitting the online purchase).

For hypothesis testing, Pearson’s χ² test statistic is chosen, described as follows:

$${\chi }^2=\sum \sum {\displaystyle \frac{{\left(O_{ij}-E_{ij}\right)}^2}{E_{ij}}},$$

(1)

such that Oij is observed frequencies, E_ij is expected frequencies computed as R_i × C_j/N, where R_i and C_j represent row and column marginals, while N is the total observations. However, we also included G² likelihood ratio:

$$G^2=2\sum O_{ij}log\left({\displaystyle \frac{O_{ij}}{E_{ij}}}\right),$$

(2)

as a more robust method for evaluating the goodness-of-fit and association between variables in contingency tables. Given that the χ² statistic is only approximated by the χ² distribution and worsens when expected frequencies are relatively small, which is a common controversy using χ² test statistic, we consider including G² because it provides more robust measurements for large dimensional tables. It is commonly discussed that G² advantage over traditional χ² is that G² for large contingency tables can be neatly decomposed into smaller components [60], which cannot be done by χ² test. Even so, as the sample size increases, statistics tend to converge. The degrees of freedom df is determined by df = (r-1)(c-1), where r and c represent classes of row and column profiles, respectively.

Furthermore, we analysed the diagnosticity of p values using the VS-MPR (Vovk-Sellke Maximum p-Ratio) calibration score. The score is computed as VS-MPR = -e × p × ln(p) and is commonly referred to as a lower bound to BF (favouring H₀ over H₁)[61]. The reason for including VS-MPR is that a large body of research [62,63,64] is calling into question the traditional threshold (p < 0.05) for deciding whether a model is statistically significant, i.e., rejects the null hypothesis. We consider the following labelling intervals as evidence in favour of the alternative over the null hypothesis: VS-MPR₁₀ = 1-3 anecdotal, VS-MPR₁₀ = 3-10 substantial, VS-MPR₁₀ = 10-30 strong, VS-MPR₁₀ = 30-100 very strong, and finally VS-MPR₁₀ > 100 as decisive evidence [65].

2.4. Multiple Correspondence Analysis Hierarchical Clustering of Principal Components

To understand MCA's description, let the initial raw dataset be a matrix X ∈ ℝ^n×m, where n is the number of observations and m is the number of variables that explain demographics and user preferences. To reduce the dimensionality of dataset X, we first perform hypothesis testing (see eq.2 and eq.3) to keep only statistically significant variables (p < 0.05, VS-MPR₁₀ > 3). After the removal of non-significant variables, the reduced X’ ⊂ X dataset comprises only significant variables. Hence, the X^’ is then subjected to MCA to further reduce dimensionality by transforming raw categorical data into Principal Components (PCs).

The first step in MCA is to convert a raw categorical dataset X^’ into indicator matrix Z ∈ ℝ^n×q, where q is the number of class categories from the retained variable set. Next, from the defined indicator matrix Z, we perform centring by subtracting row, and column means to obtain centred matrix Z_C:

$${\mathit{Z}}_{C,ij}={\mathit{Z}}_{\mathit{i}\mathit{j}}-{\displaystyle \frac{1}{n}}\sum _{i=1}^n{\mathit{Z}}_{\mathit{i}\mathit{j}}-{\displaystyle \frac{1}{q}}\sum _{i=1}^n{\mathit{Z}}_{\mathit{i}\mathit{j}}+{\displaystyle \frac{1}{nq}}\sum _{i=1}^n\sum _{j=1}^q{\mathit{Z}}_{\mathit{i}\mathit{j}},$$

(3)

which adjusts for marginal distributions ensuring standardised matrix for PC extraction. Using Singular Value Decomposition (SVD) we decompose Z_C as follows:

$${\mathit{Z}}_{\mathit{C}}=\mathit{U}\mathsf{\Sigma }{\mathit{V}}^{\mathit{T}},$$

(4)

where U ∈ ℝ^n×k are singular vectors in the reduced space, Σ ∈ ℝ^k×k is a diagonal matrix with singular values of each PC, and V ∈ ℝ^q×k is the singular vector matrix representing variable contributions of each individual PC. The selection of top k components is performed based on the first two PCs that explain the most variance as PCs ∈ ℝ^n×k.

Finally, after obtaining reduced dataset PCs, an HCPC is performed for identifying clusters, i.e., labeling user profiles in a reduced space. For performing distance computation and obtaining distance matrix D, an Euclidian distance D(i,j) between each pair of observations in PCs are performed:

$$\mathit{D}\left(i,j\right)={‖{PCs}_i-{PCs}_j‖}^2,$$

(5)

which yields distance matrix D ∈ ℝ^n×n capturing similarities between respondents. For applying hierarchical agglomerative clustering on PCs, Ward’s method is selected for merging clusters and minimising within-cluster variance. The results are represented via dendrogram. Selection of clusters C is performed by linkage L assiging labels to observations C = {c₁, c₂, …,c_n}. As a last step, we merge cluster C labels with dataset X’ resulting in final dataset X_final = [X’|C], which is then subjected to classification by ML algorithms.

2.5. Machine Learning Classifiers

Machine learning classification is performed as follows. Let dataset X_final ∈ ℝ^n×(m+1) be the final dataset, where n is the number of observations and m represents the features (demographics and user preferences), while the last column, denoted as y ∈ {1, 2, …, c}, represents the cluster labels assigned via MCA-HCPC, such that c represents total number of selected clusters.

We first perform splitting the dataset into train/test dataset, such that X_train ∈ ℝⁿ_train^×m be training matrix with y_train ∈ {1, 2, …, c}ⁿ_train labels. Similarly, let X_test ∈ ℝⁿ_test^×m be training matrix with y_test ∈ {1, 2, …, c}ⁿ_test be the test dataset. Our goal is to train the ML classifier $ℳ$ : X_train → y by optimising loss function $ℒ$. For proposed ML classifiers – Gradient Boosting Machine (GBM), Decision Tree (DT), k-Nearest Neighbors (kNN), Gaussian Naïve Bayes (GNB), Random Forest (RF), and Support Vector Machine (SVM) – the loss functions are defined for GBM:

$${\mathcal{L}}_{GBM}=-\sum _{i=1}^n\sum _{j=1}^c{y}_{i,j}\mathrm{l}\mathrm{o}\mathrm{g}\left({\widehat{y}}_{i,j}\right),$$

(6)

where in multiclass problem y_i,j is a binary indicator (1 if label j, otherwise 0), and $\widehat{y}$_i,j is the predicted probability for class j. For DT and RF a Gini impurity is estimated as:

$${\mathcal{L}}_{Gini\left(t\right)}=1-\sum _{j=1}^c{p}_j^2,$$

(7)

and Entropy as:

$${\mathcal{L}}_{Entropy\left(t\right)}=-\sum _{j=1}^c{p}_j\mathrm{l}\mathrm{o}\mathrm{g}\left(p_j\right),$$

(8)

where algorithms select the feature that maximises the reduction in Gini (or Entropy). For Gaussian Naïve Bayes, the classifier calculates posterior probability for each class:

$${\mathcal{L}}_{GNB}=\sum _{i=1}^{n}logP\left(y_i\right)+\sum _{j=1}^m\mathrm{l}\mathrm{o}\mathrm{g}\left({\displaystyle \frac{1}{\sqrt{2\pi {\sigma }_j^2}}}\mathrm{e}\mathrm{x}\mathrm{p}\left(-{\displaystyle \frac{{\left(X_{i,j}-{\mu }_j\right)}^2}{2{\sigma }_j^2}}\right)\right),$$

(9)

where P(y_i) is the prior probability of class y_i, while µ is the mean and σ² is the variance of feature j conditioned on y_i. The SVM loss function is determined per Hinge loss:

$${\mathcal{L}}_{SVM}={\displaystyle \frac{1}{n}}\sum _{i=1}^n\max \left(\mathrm{0,1}-y_i\left(\mathbf{w}\bullet X_i+b\right)\right)+\lambda {‖\mathbf{w}‖}^2,$$

(10)

where y_i ∈ {-1, 1} represents the true class label, w is the weight vector, whilee X_i is the feature vector, b is the bias, and λ is the regularisation parameter controlling the margin. Given that we face classification problem, the selection of highest performing algorithm is conducted per classification evaluation metrics of Accuracy:

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}},$$

(11)

where TP = True Positive, TN = True Negative, FP = False Positive, FN = False Negative. The Recall is calculated as:

$$Recall={\displaystyle \frac{TP}{TP+FN}}.$$

(12)

The Precision is estimated as:

$$Precision={\displaystyle \frac{TP}{TP+FP}},$$

(13)

and finally F1-score is calculated as:

$$F1-score={\displaystyle \frac{2TP}{2TP+FP+FN}},$$

(14)

for estimating the classification performance of ML algorithms. Additionally, given that we encounter a multi-class classification problem, we also provide Receiver Operating Characteristic (ROC) curve, and estimate Area Under ROC (AUROC) by adopting it to multi-class problem.

For assuring transparency and replicability of the results, the following parameters are set for the ML models. For GBM we set the Shrinkage = 0.1, with 1.0 interaction depth with minimum number of observations per node = 10. The training used per tree is 50%. The number of trees is optimised such that maximum number of trees is 100. For DT algorithm, the minimum number of observations for split is 20 with minimum number of observations in terminal = 7, with maximum interaction depth = 30. The DT tree complexity is optimised with maximum complexity penalty = 1.0. The kNN algorithm settings are set: weights = Rectungular with Euclidian distance metric used. The number of nearest neighbors is optimised such that maximum allowed nearest neighbors = 10. For RF algorithm the training data used per tree is 50%, while features are split automatically with optimised number of trees set to maximum of 100 trees. The SVM algorithm is set with the following parameters: Weights = Linear, Tolerance of termination criterion = 0.001, with function ξ = 0.01. The costs of constraints violation is optimised with maximum violation cost = 5. Lastly, there were no smoothing parameter set for GNB algorithms, however, all algorithms contain scaled features and seed is set 1234.

3. Results

3.1. Descriptive Statistics

The characteristics (Figure 3) show the following. The study comprises 906 participants, ranging from 21 to 77 years old (Median = 34.90, SD = 12.67). Most were female (63%, n = 575), while <1% of participants selected “preferred not to disclose” (n = 8). Most of the respondents had completed high school (44.92%, n = 407), followed by a Bachelor’s degree (37.86%, n = 343) and a Master’s degree (13.02%, n = 118). The respondents were primarily situated in towns (75.5%, n = 684), followed by rural (13.36%, n = 121) and suburban areas (11.15%, n = 101). The employment characteristics show that 58% (n = 526) of respondents are employed. The income status is classified according to monthly earnings of RSD (Republic Serbia Dinar) in categories as Low Income (<50.000 RSD), Mid Income (50.000-100.000 RSD), Mid-High Income (100.000 – 200.000 RSD), High Income (>200.000 RSD). The results show that most respondents had Mid-High Income (29.47%, n = 267), 25.17% (n = 228) did not want to say their income, followed by Mid Income (22.85%, n = 207), High Income (13.91%, n = 126), and Low Income (8.61%, n = 78). Lastly, the age distribution shows average AVGAge = 35.914 with SDAge = 12.671, ranging from 21 to 77 years old. The age distribution is then coded as follows: Early Adulthood (18-24 years), Young Adulthood (25-34 years), Early Midlife (35-44 years), Midlife (45-54 years), Late Midlife (55-64 years), Older Adulthood (>65 years). As per coded categories, the descriptives show that Early Adulthood (28.04%, n = 254) is the most dominant category, followed by Midlife (24.94%, n = 226) and Young Adulthood (24.72%, n = 224), Early Midlife (14.68%, n = 133), Late Midlife (6.84%, n = 62), Older Adulthood (0.77%, n = 7).

Supplementary items such as “What communication channel would you prefer to have with the supplier?” labelled as Communication, and “Used parcel locker?“ show that most users prefer SMS (49.56%, n = 449) and Mobile App (40.95%, n = 371) and uses parcel locker (24.06%, n = 218), respectively. The main preferences, such as MIPB1T, show that most users consider positive reviews (32.67%, n = 296) when buying the first time, followed by “To have a secure shopping certificate” (28.59%, n = 259), and “Fast and secure delivery” (22.30%, n = 202). The MIPBREP property suggests that “Satisfaction with products purchased” (68.32%, n = 619) is the most dominant reason for repeating the purchase, followed by “Satisfaction with the online shopping (12.69%, n = 115), “Satisfaction with the delivery of purchased products” (11.81%, n = 107), and others. The RFQ shows that negative reviews (30.13%, n = 273), followed by “Because I want to see the product live” (22.96%, n = 208) and “Long delivery time” (13.80%, n = 125) are the most common reasons why users quiet their purchase. Lastly, the PURFRE show that most users buy every six months (31.02%, n = 281), followed by “Once a month” (29.91%, n = 271).

We construct contingency tables before performing the statistical inferential statistics and χ2 tests to investigate the association between variables. However, due to the limited length and extensive number of tables (24 tables) for each model comparison, the complete analysis with tables is provided in supplementary material in cases the reader is interested in following the complete analysis. In the following, we provide the final results of each model comparison using both χ² and G² test statistics, including the effect size (for χ² test) and VS-MPR ratio.

3.2. Hypothesis Testing

After performing hypotheses testing as a proposed framework, we have rejected 8/24 hypotheses regarding the conditional dependencies, i.e., the association between e-customer demographics and user preferences (Table 1). Namely, we have only considered cases where both χ² and G² suggest statistically significant association (p < 0.05) and where there is at least moderate, i.e., substantial evidence in favour of the alternative hypothesis (VS-MPR₁₀ > 3). In such instances, we consider that results will be more robust and reliable than previous findings. We additionally provided all other non-significant test results in the appendices (Table A1).

Surprisingly, no evidence suggested that demographic properties are conditionally dependent, i.e., associated with MIPB1T (p > 0.05). There is, however, anecdotal evidence (VS-MPR₁₀ = 2.386) suggesting that Age (χ² = 37.455, p = 0.052) tends to be associated with MIPB1T but fails to reject the null (p > 0.05). Hence, we can conclude that we fail to reject the null that indicates a user preference describing “Most important property when buying from a webshop for the first time?”.

The results investigating the association of demographics to the PURFRE variable suggests that Age (χ² = 50.519, G² = 54.838, p = 0.001), Education (χ² = 38.498, G² = 39.004, p = 0.001), Income (χ² = 52.421, G² = 49.376, p = 0.001), and Work status (χ² = 51.892, G² = 50.487, p = 0.001), ranging from very strong (VS-MPR₁₀ > 30) in Education, to decisive evidence (VS-MPR₁₀ > 100) of Age, Income and Work Status in associations with PURFRE. Next, the analysis of associations with RFQ suggests that Work status provides very strong evidence (χ² = 50.697, G² = 54.134, p = 0.001, VS-MPR₁₀-χ² = 47.147), followed by substantial evidence regarding Residence (χ² = 24.599, G² = 25.710, p = 0.017, VS-MPR₁₀-χ² = 5.349) and Income (χ² = 41.019, G² = 41.087, p = 0.017, VS-MPR₁₀-χ² = 5.413). Lastly, the test considering MIPBREP suggests only the existence of association to Education with substantial (G² = 30.371, p = 0.016, VS-MPR₁₀ = 5.516) to strong (χ² = 33.84, p = 0.006, VS-MPR₁₀ = 12.456) evidence in favour of the alternative.

The results suggest statistical dependency between demographics and user preferences (excluding MIPB1T). However, more needs to be understood about how specific underlying classes associate with each other. Hence, MCA uses the χ² distance better to understand the relationship between investigated demographics and user preferences.

3.3. Multiple Correspondence Analysis

The MCA analysis suggest improved understanding of the association between class categories. Namely, the explained inertia is 11.797% in the first two components and 16.1% in the first three components (Figure 4A). Next, the η² correlation coefficient (Figure 4B), which represents the degree of association between variables and principal components, suggests that all three components explain the AGE variable well. The second component mainly captures AGE (η² = 0.516) and WORKSTAT (η² = 0.607), while user preferences include PURFRE (η² = 0.141) and RFQ (η² = 0.201). Lastly, AGE (η² = 0.499), EDU (η² = 0.119) and WORKSTAT (η² = 0.175) consider demographic variables explained by the PC3, with PURFRE (η² = 0.135), RFQ (η² = 0.134) and MIPBREP (η² = 0.181) user preferences captured. Most of the information is well-explained by the first two components, with the MIPBREP variable slightly higher η². Hence, the association between variables is discussed within the first two PCs.

From the MCA biplot (Figure 4C), we can infer three latent projections that describe demographics and user preferences. Namely, the vertical mainly suggests late midlife to older adults, low to mid-income, and with purchasing frequency from 6 to 12 months, who prefer person-to-person purchase, i.e., seeing the product live, which is mainly described by the PC2. Observing the negative side of PC1(-) and PC2(-), i.e., left bottom quadrant, we can infer that this mainly describes students, high school education, early adults, unemployed and part-time employed respondents, rural areas, and with purchasing frequency from once a month to once in six months. This can also be supplemented by the v-test (Figure 4D) score, as it mainly quantifies the class category distance from the average. Hence, PC1-PC3 can offer similar information on class categories in factor plots.

Lastly, the bottom right quadrant, captured mainly by the PC1(+) inertia, describes the respondents as early midlife and employed mainly with an MSc degree, with town residence and high income. These respondents are characterised by user preferences describing purchasing power from several times a week (CTR = 1.78, cos² = 0.04, v-test = 6.20) to several times a month (CTR = 3.77, cos² = 0.11, v-test = 10.6), while most crucial property before repeating the purchase (MIPBREP) being the satisfaction with the delivery (CTR = 1.13, cos² = 0.029, v-test = 5.12) and satisfaction with the online shopping process (CTR = 0.474, cos² = 0.012, v-test = 3.33) while the main reason for quitting the purchase is long delivery (CTR = 1.80, cos² = 0.05, v-test = 6.52).

3.4. Classification Results

The training (and holdout set) classification accuracy (Table 2) suggests that highest classification accuracy is obtained via SVM (0.950), GBM (0.939) and RF (0.928). The SVM suggests that classification metrics are mostly consistent across Precision, Recall, F1-Score and AUROC. However, the GNB, instead of RF, show high scores of Precision, Recall and F1-Score, while AUROC results suggests that GBM (0.994) and RF (0.994) are the highest.

From a cluster of user profiles, the labelling shows significant imbalance of a dataset, threrefore, an accuracy results can be misleading. Also, although precision and recall metrics are useful when false positive and false negatives are problematic, respectively. We consider F1-score and AUROC particularly useful since they adress the case of imbalanced dataset. Therefore, the results from holdout set show that SVM (0.949), GBM (0.936), GNB (0.936) and RF (0.920) are the highest performing classifiers.

The classification results (Figure 5A) show that GBM (0.994), RF (0.994) and SVM (0.902) are the highest performing classifiers. Although we are interested in PFI rankings obtained through MDL score of highest performing algorithms, we provide MDL scores of all ML classifiers. Overall, we can conclude that Age plays a crucial role in ML classifiers, followed by Work status, Education and Income status. Still, the highest performing ML classifiers suggests that Age (GBM, RF, and SVM) is the most important feature, followed by Work status (GBM, RF), Education (GBM, RF, SVM) and Residence (SVM). Hence, ensemble learners (GBM, RF) offer similar results, while SVM suggests similar results in decision boundary.

4. Discussion

4.1. Hypothesis Testing Results

We obtain the following conclusions based on the association between demographic variables and user preferences. Age indicates a significant association with Purchase Frequency (p = 0.001, V = 0.118). From a general trend (Figure 6A), Early (18-24 years) and Young Adults (25-34 years) dominate in purchasing from at least once a month to once every six months, presumably due to a combination of comfort with digital technology. However, there is a significant increase in purchasing of Early Midlife (35-44 years) and Midlife (45-54 years) several times a month and even several times a week, which was quite surprising since most of the prior literature report the dominance in frequent purchases of young adults (21-30 years). This may be attributed to the cause that parents tend to shift priorities and behaviour (e.g., career, family, investments).

Analysing the association between Education and Purchase Frequency (Figure 6B) the evidence suggests that higher educational attainment corresponds to a more selective behaviour. Namely, there is an increase in purchases from several times a month to once a month, while there is an increase in users with small purchases from once in six months to once in twelve months. We assume that Master’s and PhD degree holders tend to shop less frequently but with more deliberate timing and thoughtful behaviour in their purchases. Bachelor’s degree holders are slightly more active in online shopping than High school graduates, probably due to higher income levels and better comfort with digital content. The discussion on Elementary school education is inconclusive since only four subjects participated in the questionnaire.

The work status (Figure 6C) exhibits similar behaviour patterns of unemployed respondents, students and part-time employees. The evidence shows a significant increase in purchase frequency among employed participants. On average, there is a 25.7% increase in purchases several times per month and a 16.5% increase in purchase frequency once a month. In contrast, 6.8% and 4.9% drop in purchases “once a month” and “once every six months”, respectively. Lastly, there is a significant drop from 12-25% in purchasing from several times a week to once a month, and a significant increase in the purchases of “once every six months” (6.9%) and “once every 12 months” (37%) of the retirees.

The results comparing income status and purchase frequency (Figure 6D) show the highest effect (χ² = 52.421, V = 0.120, VS-MPR₁₀ = 3392.513, G² = 49.376, VS-MPR₁₀ = 1221.5) among tested variables, suggesting extreme evidence (VS-MPR₁₀ > 100). There is, on average, a 29.97% increase in purchase frequency “several times a week” comparing High with Low to Mid-High Income respondents. Also, there is a 14.02% increase in purchasing “several times a month” compared to Low to Mid-High income respondents. There is a significant drop of 2.91%, 9.7% and 12.13% in purchase frequency “once a month”, “once every six months”, and “once every twelve months”, respectively.

The association between Residence and RFQ (Figure 7A) suggests that suburbans' main reason for quitting is to see the product live (20% higher). At the same time, rural areas state that the main reasons for quitting are inappropriate or hidden information (8-16% higher) and complicated searches on the website (4-20% higher). The prevalent factor across all groups, particularly in rural areas, is the desire to see the product in person, while inappropriate/hidden information are main reasons for quitting the purchase.

Regarding the dependency between Income and RFQ (Figure 7B) the evidence suggests that negative reviews (30%) are the most common reasons for quitting. However, higher-income respondents cite long delivery times (29.5%) and website design (29.9%) as the most common reasons for quitting, suggesting that convenience and user experience are more critical. In comparison, the need for a better price is a minor concern (12.8%) for these users. The low-income respondents cite the need for a better price (28.4%) and reflect a more price-sensitive group, while also the need to see the product live (25.3%). Paradoxically, the respondents who preferred not to disclose their income cite that the highest dissatisfaction rate is due to inappropriate or hidden information (27.4%) about the product as a potential concern over transparency and trustworthiness. Lower-income users prioritise affordability and assurance, while higher-income respondents prioritise efficiency and convenience.

The analysis of Work Status and RFQ (Figure 7C) shows that unemployed respondents and students demonstrate more risk-averse attitudes, whereas unemployed respondents cite negative reviews and long delivery times as primary concerns for quitting. At the same time, students also cite negative reviews in addition to a need for a better price, which is also the main reason (32.5%) emphasised in responses of part-time employees. The employed e-consumers cite hidden information and website design as the primary reasons, aligning somewhat with part-time workers. Still, although this suggests that individuals generally have more disposable income, they expect a high standard of service.

Lastly, the results regarding the association between Education and MIPBREP (Figure 7D) show that as education levels increase, consumers tend to prioritise the broader shopping experience, such as satisfaction with delivery and customer services, rather than product—or price-related issues. This underlines the need for e-tailers to provide more comprehensive and personalised services to accommodate the expectations of more educated consumer groups.

4.2. Multiple Correspondence Analysis With Hierarchical Clustering on PCs

For the MCA, mainly using significant variables, three potential clusters may be identified along the axes (Figure 4C). To maintain objectivity in the identification of user profiles, we rely on Hierarchical Clustering on Principal Components (HCPC). The HCPC method uses Euclidian distance for clustering of points, while Ward’s linkage is used for cluster selection. The complete analysis is performed in R studio FactoMineR (v2.11).

The dendrogram (Figure 8A) suggests that three clusters are selected, while the interpretation can also be suitable for selecting up to six clusters based on the inertia gain. However, for the simplicity of interpretation, three clusters are selected for the analysis (Figure 8B). The data behind clusters’ are provided in Appendix C. To understand the interpretation of tables (e.g., see Table C1), let us go through the features “Cla/Mod”, “Mod/Cla”, “Global”, “p-value”, and “v-test”. For instance, Cluster 1 (black) shows that Age = Early Adulthood (“Cla/Mod” = 92.913, v-test = 24.465) suggest that 92.9% belongs to Cluster 1, while 78.67% of Age proportion (“Mod/Cla”) in Cluster 1 is Early Adulthood. The “Global” shows the overall proportion of a particular class in a complete dataset. Note that only classes (with associated categories) that are statistically significant (v-test = ±1.96, p-value < 0.05) are represented.

The Cluster 1 (black), therefore, suggests that respondents are mainly in Age = Early Adulthood, Students (with High School diploma), WORKSTAT = Unemployed, from RESI = Rural areas, and Gender = Females (and persons that do not want to disclose gender identity) being over-represented. A particular group of respondents show an association with PURFRE = Once a month, MIPB1T = To have positive customer reviews, RFQ = Because of negative reviews, and RFQ = I want to see the product live. This cluster suggests that women are dominant e-consumers, mainly driven by positive reviews for repeating the purchase (and vice versa, quitting the because of negative reviews), prefer seeing the product live with purchasing via mobile apps. The purchase frequency is characterised by moderate purchase frequency, suggesting at least once a month.

Cluster 2 (red) mainly describes WORKSTAT = Retiree (Cla/Mod = 100%) and Age = Older Adulthood (Cla/Mod = 100%), with minimal purchase frequency of once every 12 months, that is mainly driven by attractive new offers when wanting to repeat the purchase but are also quitting if they cannot see the product in person. Ultimately, although with limited respondents (Global = 1.214%), this suggests that these e-consumers are retirees and infrequent cautious shoppers, and the key to repeating the purchase of these respondents is customer service and trust.

Cluster 3 (green) is comprised of diverse demographics. Namely, the cluster suggests that e-customers are mainly WORKSTAT = Employed (Cla/Mod = 95.82%), AGE = Early Midlife – Midlife – Late Midlife (99.11%, 94.74%, and 95.16% respectively), EDUC = BSc - MSc - PhD (74,64%, 91.52%, and 97.1% respectively), and mostly dominated by males (73.1%) with township residence (67.54%). The user preferences of these e-consumers, being frequent consumers (several times a week to several times a month), show that fast and accurate delivery is essential when buying the first time, while inappropriate and hidden information and long delivery are the main reasons for quitting the purchase. Overall, this cluster resembles early to late midlife, highly educated respondents, mostly male frequent shoppers who prioritise convenience and service efficiency.

4.3. Validity of Findings from Classifiers and Feature Importance

The ML classifiers offer several critical inights about the effectiveness in distinguishing different e-customer profiles based on demographics and user preferences. Namely, the SVM, GBM and RF algorithms demonstrate highest accuracy, achivieving robust performance across proposeed metrics. The SVM achieved highest overall accuracy, suggesting high efficiency in handling multi-class problem with imbalanced dataset. This can be attributed to the SVM capacity to capture class separability by optimised hyperplanes. The Hinge loss optimisation enables SVM to minimise missclassifications near decision boundaries, which is particularly useful in the case of MCA-HCPC user profile labelling. Similarly, ensemble methods, i.e., GBM and RF algorithms, performed consisteny well, with GBM outperforming in F1-score. This can be attirubted to GBM’s iterative boosting process that prioritises correcting previous falsely classified labels performing fine-tuning. On the other hand, TF’s robust performance is seen from the use of multiple decision trees, ultimately reducing the risk of overfitting and enhnacing generasibility.

Although performing adequately, the GNB classifier offered less competitive outcomes in multi-class accuracy and AUROC than the SVM and ensemble methods. Such results can likely be due to GNB’s feature independence assumption, which results in oversimplified data relations in complex user profiles. Lastly, the KNN classifier showed lower performance, presumably due to sensitivity to feature scaling and variations across classes, which were prevalent in this categorical dataset.

After obtaining important features affecting ML performance, the interpretation of e-customer profiles suggests the following. Age is consistently ranked as the highest and most influential, one can assume that younger adults are typically of most interest to e-tailers. Hence, we can assume a correlation between Age and technology adoption and comfort with online shopping. Younger adults typically display more purchases and are drawn by user-friendly and interactive platforms. On the other hand, older adults often favour quality assurance and transparency. Ultimately, age-based preferences can reflect generational differences in digital engagement, where young people are more amenable to risk and immediacy, while the elderly prioritise security and familiarity.

Another critical factor is work status, which affects purchasing power and shopping references. Employed individuals tend to have more disposable income and thus exhibit preferences for efficient and value-driven experiences, suggesting that fast and reliable delivery services play a more important role than price. Conversely, students and unemployed e-customers exhibit patterns in shopping behaviour that reflect more price-sensitive decisions that place greater emphasis on reviews and promotional offers.

Unlike the work status, which can be considered an indirect indicator of budgetary flexibility and consumption priorities, the importance of the Education feature may suggest two things: (a) technology savviness and (b) service quality expectations. What do we mean by this? From the perspective of the sample demographics, it seems that higher education is associated with comfort in navigating through digital platforms, leading to higher purchases and trust in e-commerce. Similarly, educated individuals also exhibit higher standards for customer service, preferring e-tailers with strong reputations consistent with reliability, customer support, and transparency. The feature offers additional insights into the relationship between cognitive engagement and customer loyalty factors, as educated e-customers might scrutinise product information and reviews more closely.

Income also acts as one of the main determinants of purchasing power and often correlates with shopping frequency. The findings also suggest that higher-income respondents prioritise convenience, service quality and delivery over price, making them more inclined to purchase “premium” products or services. E-tailers may utilise such implications in providing platforms that offer enhanced customer experiences. On the other hand, low-income e-consumers typically emphasise affordability, promotional deals and discounts. Finally, while not the most dominant feature, the residence also differentiates user profiles, particularly for urban and rural e-customers. Specifically, rural users may face logistical constraints regarding reliability and transparency of delivery, which makes them more inclined to quiet if logistical support appears uncertain.

Findings from ML classifiers’ feature importance show major implications for tailored e-cummerce strategies. Age, as the most influential feature, stress the generational divide in digital engagement, which distinguishes younger and older adults on users who seek convenience and interaction and users who prioritise transparency and security, respectively. Simultaneously, work status and income further provides a contrast in purchasing power and preferences, where employed and high-income users prioritise efficient service over cost, while students and lower-income customers remain price-sensitive often guided by reviews and promotional offers. As a cognitive component, Education plays a role in service standards and critical engagement about product detials. Overall, the proposed ML models’ findings offer meaningful insights and strategies for e-tailers in shaping user profiles and developing marketing strategies.

5. Conclusions

5.1. Concluding Remarks

The study performs threfold analysis. Firstly, the study determines statistically significant variables by investigating the association between customer demographics and user behavioural purchase preferences in the Republic of Serbia on a survey performed on a sample of n = 906 respondents. The findings show an 8/24 significant association, with extreme evidence considering association between Age and Purchase Frequency, Income Status and Purchase Frequency, and Work Status and Purchase Frequency. Interestingly, there is no significant association reported between demographics and items of “Most important property when purchasing for the first time” (p < 0.05). Still, all reported tests suggest a small effect size, reported per Cramer’s V statistic, suggesting that many more variables explain the factors of user preferences outside of used demographics.

Secondly, given that statistical tests do not expose particular user profiles, we performed MCA-HCPC (Multiple Correspondence Analysis Hierarchical Clustering on Principal Components) identifying three main clusters. The first cluster mainly comprises early adult students (unemployed), primarily females from rural areas. These profiles are characterised by moderate purchase frequency (at least once a month) driven by positive customer reviews and wanting to see the product in person. The second cluster describes retirees who exhibit infrequent (once in 12 months) purchases driven by attractive offers but hesitant if they cannot inspect the product. The third cluster includes a more diverse group but primarily describes males in early to late midlife stages who are employed who tend to prioritise fast and accurate delivery. The unsupervised labels, i.e., clusters obtained from an MCA-HCPC analysis, are subjected to ML classifiers.

Finally, merged dataset with C cluster labels are subjected to the following ML algorithms: Support Vector Machine, Gradient Boosting Machine, Decision Trees, Random Forest, k-Nearest Neighbors, and Gaussian Naïve Bayes. The results suggest >90% classification accuracy, where GNB, RF and SVM offer highest classification metrics considering Accuracy, Precision, Recall, F1-score and AUROC. Additionally, from the obtained classifiers we extract most important features by permutations feature importance that relies on mean dropout loss. The evidence suggests that age, work status, education, income and residence are the main determinants needed for shaping user profiles and developing e-tailers marketing strategies.

5.2. Limitations of the Study

Regarding the demographics, there is a low representation of the elderly and residents from rural areas and specifics of the local market, such as a high percentage of e-customers who did not want to disclose their income. This introduced a slight increase of heterogeneity in the sample, i.e., reduced confidence in the relationship between income status and user preferences, which may also impact the application of the research results.

The Multiple Correspondence Analysis suggests low inertia explained by the first three components. This may affect the interpretation of association, both in individuals and among class categories. This certainly does not downplay our findings, given that most of the captured (inertia) variance corresponds well to investigated test statistics. However, a higher percentage of captured inertia would undoubtedly increase the confidence in understanding user profiles explained through the Hierarchical Clustering of Principal Components.

5.3. Implications

Regarding e-customer preferences, it seems that website design plays a minimal role for most respondents, showing that the user interface of online stores plays less of a barrier than logistical and trust-related issues. Thus, improved logistics (e.g., fast and delivery), mainly for township areas, are a significant concern, likely due to infrastructure issues. Also, product transparency (e.g., hidden or inappropriate information) plays a significant role in quitting the purchase, in addition to negative reviews. Thus, platforms could benefit from more proactive review management, encouraging satisfied customers to leave positive reviews and addressing negative feedback promptly.

In future research, we want to focus more on the behavior of specific subgroups, which could provide a better understanding of e-customer preferences and consumption patterns. We also want to expand the research to other regions and countries, mainly on the Balkan peninsula, so we can compare and identify differences in user behaviour on a more global scale.