Predicting the relationship between consumer buying behavior (CBB) and consumption metaphor (CM) through machine learning (ML)

There are many studies associating tattoos and other body arts with CBB. Sanders (1985) examined the personalization of body parts as a component of consumer behavior. Watson (1998) and Horne et al. (2007) consider purchasing tattoos and other body arts as a form of ownership. Watson (1998) also suggested that the tattoo consumer makes at least two decisions regarding the choice of symbol and determining which part of the body to place it on. According to Kowalski (1999), young consumers are highly influenced by the effects of peer pressure when deciding to get a tattoo. Older consumers, on the other hand, are influenced by their children or friends when deciding on a tattoo (Hebden et al., 2011).

When consumers consider buying tattoos, they are influenced by their desire to enhance and support their self-concept in the eyes of others (Allison et al., 2000). Research has revealed that women are the group who buy the most tattoos and other body arts. Women think that a tattoo makes them more attractive (Harris Interactive, 2008). Burnkrant and Cousineau (1975) claimed that consumers buy tattoos in the belief that harmony can only be achieved when their behavior is seen by others. Park and Lessig (1977) also supported this view. According to Fishbein and Ajzen (2010), general attributes of a tattoo include price, future purchase intention, location on the body, and design. On the other hand, the global attitude toward tattoos also has an impact on the purchase of tattoos and other body arts. It seems that there are many other factors, especially demographic factors, that affect the consumer's decision when buying a tattoo.

It is important to discuss in detail the terminological differences and the interactions between these concepts in order to better understand the relationship between CBB and the CM. Accordingly, consumer behavior encompasses a broader scope compared to CBB. It includes processes such as purchasing, consuming, and disposing of products and services (Belch & Belch, 1998; Engel et al., 1995; Kotler et al., 2005; Solomon, 2007). CBB refers specifically to purchasing products or services for personal or household use (Pride & Ferrell, 2000). CBB can be influenced by various factors, including packaging, branding, and demographic characteristics. For instance, some studies have found that packaging and its components influence consumers’ buying decisions (Hsee et al., 2009). Branding is another factor that significantly affects CBB. Additionally, studies on brand image and advertisements have demonstrated a strong relationship with CBB (Malik et al., 2013; Raheem et al., 2014). The effect of demographic characteristics on CBB is also undeniable (Fishbein & Ajzen, 2010; Harris Interactive, 2008; Hebden et al., 2011).

A metaphor is a figure of speech that explains something by analogy or by describing it in terms of something else (Lakoff, 2008). Various studies have explored the relationship between metaphor and CBB, known as the CM (Forceville, 1994; Leiss et al., 1990; McQuarrie & Mick, 1999; Stern, 1988). Zaltman and Coulter (1995) suggested that preferences, including the behavior of buying a pack of margarine, can be explained by metaphors they define as balance, transformation, journey, container, connection, control, and resource.

Just like events and facts, actions are given more meaning through metaphors. This is the main reason why metaphors are used in CBB (Lakoff, 2008). In consumer behavior research, analyzing data within the context of cultural metaphors offers a more accurate approach to minimize data loss (Moisander & Valtonen, 2006). Levy (1981) stated that consumers use metaphors in product usage and brand interpretation. Analyses based on metaphors are frequently applied in advertising and other marketing communication efforts. Through metaphors, the deep meanings formed by individuals’ experiences are revealed (Stern, 1995). Techniques like ZMET (Zaltman Metaphor Elicitation Technique) further facilitate uncovering consumers’ underlying thoughts (Coulter et al., 2001).

By analyzing the metaphors used by different consumer groups, valuable insights can be gained, providing various opportunities for market strategy (Hirschman, 2007). For example, a water company used religious connotations after some consumers indicated that bottled water represented religious purification. Ritzer (2003) explained modern shopping centers using the metaphor of “islands of the living dead,” highlighting how consumers spend significant time in such places. Holt (1995) discussed CMs across four dimensions: experience, integration, play, and classification. From this perspective, Chelminski and Ekin (2007) viewed CMs as tools rather than outcomes. In one study, consumers seeking tattoos chose metaphors such as gaining experience, joining a community, socialization, and social differentiation. The hypothesis H1 is formulated as follows:

H1: There is a significant relationship between CBB and the CM (p < 0.05).

Chaotic systems such as human behavior, weather conditions, and road traffic do not follow well-defined logical rules. The behavior of chaotic systems is highly sensitive to numerous parameters, making linear methods ineffective for predicting future states. Nonlinear ML methods have emerged as an alternative field of study, enabling researchers to uncover patterns that recur within disorder. In recent years, research on predicting human behavior across various domains using ML methods has gained momentum, especially due to its potential for commercial and political applications (Janssen et al., 2022; Van Noordt & Misuraca, 2022). This study provides an overview of research on CBB and its analysis through ML.

Chen et al. (2021) emphasize that tourism purchases can be accurately predicted using ML algorithms. Their study introduces a purchase forecasting model for online tourism and evaluates the impact of various behavioral variables on online tourism purchases. Shebl et al. (2021) analyze customer characteristics to examine the influence of economic and social factors on consumer behavior. Using data from the fast-food restaurant sector, they apply statistical methods and conclude that CBB is more strongly influenced by economic factors.

Chaudhuri et al. (2021) investigate CBB on an e-commerce platform to assist platform designers in planning user engagements. The study evaluates multiple ML algorithms and a deep learning method on a large multidimensional dataset. Alshehri et al. (2021) focus on predicting revenue for massive open online courses by analyzing students’ buying behavior. Their research compares the predictive accuracy of several ML algorithms in determining course purchasability.

Huiru et al. (2018) explore consumer behavior in wine consumption by applying two ML methods, providing valuable insights for wine production and marketing managers to make informed decisions. Zuo et al. (2016) propose a method for analyzing CBB by utilizing two representative ML methods to process RFID data collected from supermarkets. Chiu (2002) presents a case-based reasoning system enhanced by a genetic algorithm (GA-CBR) to predict CBB. Developed using real-world data from a global insurance company, the system incorporates an optimization mechanism to identify the most and least likely customers to purchase insurance.

Based on these reviewed studies, ML algorithms in this research are employed to predict CBB by utilizing variables collected from a tattoo survey as problem features. The study further validates the H1 hypothesis before applying ML methods to analyze the data.

This part of the study was designed in the general survey model based on quantitative data and the relational survey model. Therefore, the study was conducted on 378 university students who have or intend to have a tattoo (n = 378), and printed surveys were applied to students from Selcuk and Konya Technical Universities. In the study, firstly, the effect of the CM on CBB was examined. Therefore, as a data collection technique, a survey was applied to 378 university students. The obtained data were analyzed with JAMOVI and SPSS 26.0 programs. The random sampling method was preferred in determining the students to be included in the sampling. Questions about the CM were adapted from Rammstedt and John (2007). A Likert scale (5) was used. The questions were replied to in the range of 1 = I strongly disagree, 5 = I strongly agree. Confirmatory factor analysis was performed to test the validity of the scales. Non-parametric analyses were performed on the data. In the analysis of the data, Cronbach’s Alpha reliability test and correlation analysis were performed. A cross-tab Chi-square analysis was conducted to determine the relationship between consumers’ CM and CBB. The significance level of the study was taken as (p < 0.05). The results showed that the CM is effective on CBB. In this context, the research hypothesis and research model are designed as follows. Thus, it will be evaluated whether the results obtained in the study are statistically significant, and whether the hypothesis is confirmed or not will be tested.

H1: There is a significant relationship between CBB and CM as shown in Figure 1 (p < 0.05) (Coulter et al., 2001; Hirschman, 2007; Lakoff, 2008; Stern, 1995).

Figure 1

Relationship between CM and CBB.

The second scenario of the research consists of an ML application. The ML application is also divided into two sub-stages. In the first stage of the ML application, the estimation of CBB was carried out. In the second stage, the effect of the CM on CBB was estimated using the ML technique. The result obtained through the statistical analysis of marketing techniques was confirmed in this process. The estimation process was carried out in two steps: feature selection and model creation. NCFS was used in the first step to identify important predictors. Afterward, three popular learning algorithms kNN, SVM, and ELM were used to predict nonlinear human behavior.

Feature selection is the process of investigating the effectiveness of features in data mining. Determining the subset involving the most significant features assists decrease overfitting, reduce training time, and improve accuracy of a learning model (Ayyıldız & Tuncer, 2020). More comprehensive reviews on feature selection methodologies can be referred to (Guyon and Elisseeff, 2003; Guyon et al., 2008; Saeys et al., 2007). In this study, the NCA which is a feature weighting approach was employed as a feature selection (Goldberger et al., 2004). NCFS attempts to state the weights of features by optimizing an average leave-one-out classification accuracy objective function.

The primary goal is to identify a weighting vector w on the labeled training set T = {(x _i , y _i ), i = 1, 2, …, N} where x i {x}_{i} denotes the feature vector and y i {y}_{i} is the class label. The weighted distance between two samples x i {x}_{i} and x j {x}_{j} is defined in equation (1). (1) D w ( x i , x j ) = ∑ m = 1 r w m 2 | x im − x jm | , {D}_{{\rm{w}}}({x}_{i},{x}_{j})=\mathop{\sum }\limits_{m=1}^{r}{w}_{m}^{2}|{x}_{{im}}-{x}_{{jm}}|, where w m {w}_{m} denotes the weight of the mth feature. A kernel function defines a relation between the probability P ij {P}_{{ij}} and the weighted distance D w {D}_{{\rm{w}}} . The kernel function k k returns large values for small D w {D}_{{\rm{w}}} . P ij {P}_{{ij}} can be defined as shown in equation (2). (2) p ij = k ( D w ( X i , X j ) ) ∑ j = 1 , j ≠ i n k ( D w ( X i , X j ) ) . {p}_{{ij}}=\frac{k({D}_{{\rm{w}}}({X}_{i},{X}_{j}))}{\mathop{\sum }\limits_{j=1,j\ne i}^{n}k({D}_{{\rm{w}}}({X}_{i},{X}_{j}))}.

If i = j , i=j, then P ij = 0 {P}_{{ij}}=0 . Kernel function is calculated as k ( z ) = exp ( − z / σ ) k\left(z)\left=\exp (-z\left/\sigma ) where parameter σ \sigma denotes kernel width and it affects the probability of each point being selected as the reference point. The probability of correct classification of x i {x}_{i} is given in equation (3). (3) p i = ∑ j y ij P ij , {p}_{i}=\sum _{j}{y}_{{ij}}{P}_{{ij}}, where y ij {y}_{{ij}} is an indicator to show if a true class is predicted. If y i = y j {y}_{i}={y}_{j} , it returns 1 otherwise 0. The pseudocode of the algorithm is given in Table 1 (Yang et al., 2012).

Table 1

1	Procedure NCFS ( T , α , σ , λ , η ) {\rm{NCFS}}(T,\alpha ,\sigma ,\lambda \left,\eta ) : T: training set, α \alpha : initial step length, σ \sigma : kernel width, λ \lambda : regularization parameter, η \eta : small positive constant;
2	Initialization: w ( 0 ) = ( 1 , 1 , … , 1 ) , ϵ ( 0 ) = − ∞ , t = 0 {w}^{\left(0)}=(1,1,\ldots ,\hspace{1em}1),{\epsilon }^{\left(0)}=-{\rm{\infty }},t=0
3	repeat
4	for I = 1, …, N do
5	Compute p ij {p}_{{ij}} and p i {p}_{i} using w ( t ) {w}^{\left(t)} according to (2) and (3)
6	for l = 1, …, d do
7	∆ l = 2 1 σ ∑ i p i ∑ j ≠ i p ij | x il − x jl | − ∑ j y ij p ij | x il − x jl | − λ w l ( t ) {\triangle }_{l}=2\left(\phantom{\rule[-0.75em]{}{0ex}},\frac{1}{\sigma }{\sum }_{i}\left(\phantom{\rule[-0.75em]{}{0ex}}{p}_{i}{\sum }_{j\ne i}{p}_{{ij}}|{x}_{{il}}-{x}_{{jl}}|-{\sum }_{j}{y}_{{ij}}{p}_{{ij}}|{x}_{{il}}-{x}_{{jl}}|\right)-\lambda \right){w}_{l}^{(t)}
8	t = t + 1 t=t+1
9	w l ( t ) = w l ( t − 1 ) {w}_{l}^{\left(t)}={w}_{l}^{(t-1)}
10	ϵ ( t ) = ε ( w ( t − 1 ) ) {\epsilon }^{\left(t)}=\varepsilon ({w}^{(t\left-1)})
11	if ϵ ( t ) > ϵ ( t − 1 ) {\epsilon }^{\left(t)}\gt {\epsilon }^{(t\left-1)} then
12	α = 1.01 α \alpha =1.01\alpha
13	else
14	α = 0.4 α \alpha =0.4\alpha
15	until | ϵ ( t ) − ϵ ( t − 1 ) | < η |{\epsilon }^{\left(t)}-{\epsilon }^{(t\left-1)}|\lt \eta
16	w = w ( t ) w={w}^{\left(t)}
17	return w w

Source: http://www.jcomputers.us/vol7/jcp0701-19.pdf.

The kNN algorithm is an instance-based classification algorithm proposed by Fix and Hodges (1951). The method is based on predicting the class of the new sample by investigating the distance of the features of the new instance to the features of known instances in the training set as follows. (4) d i = ‖ X u − X i j ‖ , {d}_{i}=\Vert {X}^{u}-{X}_{i}^{j}\Vert , where d i {d}_{i} is the distance, X u {X}^{u} is the feature vector of the instance whose class is unknown, and X i j {X}_{i}^{j} is the feature vectors of all instances in the training set. One of the most commonly used metrics for calculating the linear distance between instances is the Euclidean distance. According to the nearest neighbor algorithm approach, the class of the instance is considered as the class of its nearest neighbor. The k-nearest neighbor approach can be preferred to reduce the classification error. The class of the instance is determined by making a majority calculation among the kNNs. The steps of the algorithm are explained in Table 2.

Table 2

1	Determine the value of k k ,
2	Calculate the distance between the new instance ( x ) \left(x) and the entire training set d ( x , y ) d(x\left,y) ,
3	Determine the k k nearest neighbors by sorting the distances from smallest to largest,
4	Determine the class of the new instance by majority voting.

Source: Authors’ own research results.

SVM algorithm is based on the idea of finding a decision boundary called hyperplane that classifies the data points in high-dimensional space (Vapnik, 2013). The objective of the algorithm is to determine a plane that has the maximum margin among many possible hyperplanes. Any hyperplane can be written as equation (5). (5) w T x − b = 0 , {w}^{T}x-b=0, where w is the normal vector to the hyperplane, x x is data points, b b is the bias term of the hyperplane equation. If ‖ w ‖ \Vert w\Vert is minimized, the best hyperplane that can separate the classes can be determined. If the data points denoted by D D are separable linearly, a kind of optimization problem is solved. Data points are shown as equation (6). (6) D = { ( x i , y i ) | x i ∈ R N , y i ∈ { − 1 , 1 } } i = 1 n , y i ( w T x i − b ) ≥ 1 . D={\{({x}_{i},{y}_{i}){\rm{|}}{x}_{i}\in {R}^{N},{y}_{i}\in \{-1,1\}\}}_{i=1}^{n}\left,\hspace{1em}{y}_{i}({w}^{T}{x}_{i}\left-b)\ge 1.

As it is known, in real-world problems, data are often noisy and linearly inseparable. Defining hing loss function is shown in 7. SVM was able to solve such problems as well. Thus, the optimization problem takes the following form:

max ( 0 , 1 − y i ( w T x i − b ) ) \max (0,1-{y}_{i}({w}^{T}{x}_{i}-b)) (7) 1 n ∑ i = 1 n max ( 0,1 − y i ( w T x i − b ) ) + λ ‖ w ‖ 2 , \left[\frac{1}{n}\mathop{\sum }\limits_{i=1}^{n}\max (\mathrm{0,1}-{y}_{i}({w}^{T}{x}_{i}-b))\right]+\lambda {\Vert w\Vert }^{2}, where λ \lambda allows splitting the x i {x}_{i} to the correct side of the boundary by maximizing the margin.

ELM proposed by Huang et al. (2006) are random weight networks in the type of single-layer feedforward neural networks (SLFN). ELM uses the Moore-Penrose generalized inverse to adjust neural weights instead of gradient-based backpropagation. Due to the weights stated analytically, the method tends to provide good generalization performance at an extremely fast learning speed. SLFNs with random hidden nodes are mathematically modeled as equation (8). (8) ∑ i = 1 N ˆ β i g i ( x j ) = ∑ i = 1 N ˆ β i g i ( w j · x j + b i ) = o j , j = 1 , … , N , \mathop{\sum }\limits_{i=1}^{\hat{N}}{\beta }_{i}{g}_{i}({x}_{j})=\mathop{\sum }\limits_{i=1}^{\hat{N}}{\beta }_{i}{g}_{i}({w}_{j}{\rm{\cdot }}{x}_{j}+{b}_{i})={o}_{j},j=1,\ldots ,N, where N ˆ \hat{N} is the number of hidden units, β \beta is the weight vector between i i th hidden layer and output, g g is an activation function, x x is an input vector, w w is a weight vector between input and hidden layer, b is a bias vector, and N N is the number of training samples. The equation can be shortened as: (9) H β = T , H\beta =T, where, (10) H = g ( w 1 · x 1 + b 1 ) ⋯ g ( w N ˆ · x 1 + b N ˆ ) ⋮ ⋱ ⋮ g ( w 1 · x N + b 1 ) ⋯ g ( w N ˆ · x N + b N ˆ ) N × N ˆ , β = β 1 T ⋮ β N ˆ T N ˆ × m , T = t 1 T ⋮ t N T N × m , H={\left[\begin{array}{ccc}g({w}_{1}{\rm{\cdot }}{x}_{1}\left+{b}_{1})& \cdots & g({w}_{\hat{N}}{\rm{\cdot }}{x}_{1}\left+{b}_{\hat{N}})\\ \vdots & \ddots & \vdots \\ g({w}_{1}{\rm{\cdot }}{x}_{N}\left+{b}_{1})& \cdots & g({w}_{\hat{N}}{\rm{\cdot }}{x}_{N}\left+{b}_{\hat{N}})\end{array}\right]}_{N\times \hat{N}},\beta ={\left[\begin{array}{c}{\beta }_{1}^{T}\\ \vdots \\ {\beta }_{\hat{N}}^{T}\end{array}\right]}_{\hat{N}\times m},T={\left[\begin{array}{c}{t}_{1}^{T}\\ \vdots \\ {t}_{N}^{T}\end{array}\right]}_{N\times m}, where H H is the hidden layer output matrix, T T is the target matrix of training data, m m is the number of outputs. If N ˆ = N \hat{N}=N , matrix H H is invertible but in most cases the number of hidden nodes is much less than the number of training samples, N ˆ ≪ N \hspace{1em}\hat{N}\ll N ; thus, the inverse of H H is calculated via the Moore–Penrose generalized inverse shown as equation (11). (11) β ˆ = H † · T . \hat{\beta }={H}^{\dagger }{\rm{\cdot }}T.

As a result, the neural network model is illustrated in Figure 2, and the algorithm is summarized in Table 3.

Figure 2

Neural network model.

Table 3

1	Assign weights w i {w}_{i} and biases b i {b}_{i} randomly
2	Calculate hidden layer output, H
3	Calculate output weight matrix , β ˆ = H † · T \hat{\beta }={H}^{\dagger }{\rm{\cdot }}T
4	Use T = H β ˆ T=H\hat{\beta } to predict the classes of testing data

Source: Authors’ own research results based on ELM.

This is the part in which the CBB and the CM are associated. First, the frequency distribution table according to the demographic characteristics of the participants is given in Tables 4 and 5.

When Table 4, regarding the characteristics of the sample, is examined, 47.6% of the participants are female, and 52.4% are male. The 19–20 age group constitutes 46.8% of the age distribution. Additionally, 62.7% of body types are classified as normal. Furthermore, 79.1% of the education levels consist of an associate's degree. The field of science for the participants is predominantly social sciences, with a rate of 85.7%. In terms of income, 59.3% of participants earn 1,000 TL or less.

When Table 5, regarding the socio-demographic characteristics of the sample, is examined, 65.1% of the participants reported not being in a relationship. Places of birth are predominantly provinces and districts, accounting for 52.4%, while living places are primarily big cities, with a rate of 64.3%. The mother’s education level is primary and secondary education for 82.6% of participants, and the father’s education level is primary and secondary education for 79.1%. Additionally, “no response” constitutes 87.3% of the state of illness responses.

The CM survey questions were adapted from Rammstedt and John (2007). The CM scale consists of four questions and is designed in a five-point Likert scale format, ranging from 1 = Strongly Disagree to 5 = Strongly Agree.

The CBB scale consists of 11 questions. The internal consistency coefficient of the CBB scale was calculated as α = 0.845, while the CM scale was calculated as α = 0.961. If these values are above 0.70, it indicates that the reliability of the scale is high. To analyze the data, Cronbach’s Alpha test, descriptive statistics, and correlation analysis were used. The significance level of the study was set at p < 0.05. The findings were interpreted by transforming them into tables aligned with the research questions.

According to the Cronbach Alpha values in Table 6, both scales are above 0.70. Accordingly, it is understood that the scales used for CM and CBB are reliable.

The KMO value should be above 0.50 (Bartlett, 1950). Table 7 indicates that the KMO test clearly meets this requirement, with a value of 0.835.

Items with factor loadings below 0.30 and a difference of no greater than 0.100 between loadings in different factors (i.e., any item loading on more than one factor with a difference of less than 0.1) were excluded from the analysis (SD4, SD6, SD8, SD9). The remaining items were analyzed using principal components analysis and the Varimax orthogonal rotation technique. The total variance explained was found to be 58.812%.

According to Table 8, the structures of the scales tested through the exploratory factor analysis and internal consistency findings were confirmed.

Confirmatory factor analysis was performed to test the validity of the scales used in the study. As can be seen in Tables 4 and 5, the structures of the scales tested according to confirmatory factor analysis and internal consistency findings were confirmed.

When the model fit criteria, goodness of fit index reference ranges in Table 9, model results, and Table 10 are examined, the goodness of fit for the CM and CBB scales is high. The Kolmogorov–Smirnov test was used to assess the distribution of the data. According to the test results, the CM measurement data (test statistic: 0.112, p = 0.00) and the CBB measurement data (test statistic: 0.102, p = 0.00) are not normally distributed, as indicated by the rejection of the null hypothesis.

To avoid Type I and Type II errors, non-parametric analyses were conducted. The data were analyzed using the Cronbach’s Alpha reliability test and descriptive statistics, while correlation analysis was used to compare the two variables. A cross-tabulation Chi-Square analysis was also conducted to examine the relationship between CM and CBB. The significance level for the study was set at p < 0.05. The findings obtained from the analyses were interpreted and presented in tables aligned with the survey questions.

According to the results of the simple correlation analysis of the relationship between CM and CBB measurement data in Table 11, a positive, moderate, and significant correlation is observed (r = +0.315, p < 0.01).

After identifying this significant relationship, a chi-square test was performed to analyze the association between CBB categories and the CM. The analysis results are presented in Table 12. When the results of the Chi-Square analysis test among the categories of CBB are examined in Table 12, it is observed that 96.4% of the participants expressed a desire to get a new tattoo, while 3.6% stated that they did not plan to have one. Additionally, 46.0% of those who wanted a tattoo responded, “I did not have it” when asked if they would like to have a tattoo. Furthermore, 70.7% of the study participants answered, “I did not have it” to the question, “Would you like to have a tattoo?”

Overall, 54.3% of the participants indicated, “I would like” to have a tattoo, while 45.7% stated, “I would not.” The Chi-Square analysis results for these findings were statistically significant (X ² = 188.920, p = 0.000). Moreover, the two-way significance value (p) was less than 0.05 within the 95% confidence interval. Consequently, the H1 hypothesis was accepted (H1: There is a significant relationship between CBB and CM, p < 0.05).

This study focused on a specific group of university students, which inherently limits the generalizability of the results to a broader population. Additionally, logistical and economic challenges associated with data collection prevented the use of larger, more diverse samples. As a result, the study was intentionally conducted on a homogeneous and relatively small group to establish a foundation for future research. To address generalizability, we utilized algorithms that are current in the literature and known for their high generalization capabilities.

Future studies could build on this work by testing the generalizability of the results using larger samples with diverse socio-demographic characteristics. Expanding the dataset in this manner would enhance the robustness and applicability of the findings.

A review of the ML literature reveals the vast range of available algorithms, including those designed for clustering and prediction tasks. Examples include artificial neural networks (ANN), logistic regression, and decision trees (Mohri et al., 2012). The absence of additional learning algorithms in this study represents one of its limitations (Canepa, 2016). Future research could explore the use of a wider variety of ML techniques to compare their effectiveness.

Further, incorporating additional questions related to the CM scenarios could improve the accuracy of predictions in ML applications. Given that predicting CBB is a chaotic and non-linear problem, future studies could also focus on developing new deep learning models to uncover higher-level patterns in the CBB of individuals who express interest in tattooing.