Introduction

Social media play a significant role in people’s lives for several purposes, such as communication, education, employment, and entertainment. A broad global survey conducted in 2023 of internet users showed that the number of active social media users reached 4.8 billion, which is equivalent to more than half of the world’s total population1. People of different ages, occupations, education levels, backgrounds, geographical locations, and cultures use these platforms, providing a valuable opportunity for businesses.

The fact that personalized ads on social media are beneficial for the customer is not ignored; 80% of customers reported that when marketers personalized the experience, they were more likely to purchase2. However, social media ads lead to privacy issues, including privacy fatigue. Privacy fatigue is a relatively novel psychological phenomenon that refers to users’ boredom with online privacy issues and requirements, which leads to a lack of privacy-protective behavior3,4,5. The dimensions of privacy fatigue are cynicism and emotional exhaustion5. Studies showed that privacy fatigue affects users’ behavior more than privacy concerns5,6.

Several factors could affect privacy fatigue, such as users’ Information Privacy Awareness (IPA) and personality traits6,7,8. In this study, IPA refers to users’ awareness of the personal data that is collected, used, and shared for social media ads and the impact of this practice in the future. According to personality psychologists, there are five basic dimensions of personality, known as “the big five”: extraversion, neuroticism, openness, conscientiousness, and agreeableness9.

The use of Machine Learning (ML) and social media data to predict human behavior and emotions, such as cyberbullying, anxiety, depression, and concerns is achieving tremendous attention among researchers10,11,12,13. It is mainly used to prevent, avoid, or reduce the impact of these issues. Previous research has focused on privacy concerns regarding social media ads13,14,15,16. Therefore, with only a few studies on the phenomenon of privacy fatigue5,6,17, this study extends the concept into the social media context. It aims to predict users who will be affected by high privacy fatigue, based on their IPA and personality traits, so that they can avoid the lack of privacy-protective behavior and the consequences of failing to take appropriate precautions.

Therefore, the objectives of this study are as follows: (a) find whether the privacy fatigue psychological state exists regarding social media ads; (b) examine the relationship between privacy fatigue and users’ IPA level regarding the personal data collected, used, and shared for social media ads; (c) discover whether users’ personality traits have an impact on privacy fatigue; and (d) develop a model using a machine learning classification algorithm to predict users with privacy fatigue due to social media ads, based on their IPA and personality traits.

Related works

Privacy fatigue

With the evolution of modern communication technologies, information privacy has been under increasing threat and has become an area of interest for researchers. A relatively novel phenomenon regarding online privacy called privacy fatigue is currently under discussion by researchers5,6,17. This type of fatigue is psychological. The concept of psychological fatigue was introduced first in the medical field to reflect unpleasant feelings and tiredness resulting from high expectations and unmet demands18,19. Psychological fatigue has been explored in several research fields, such as clinical research20, occupational research21, and recently online research due to technological advances5,6,7.

Online privacy fatigue is a psychological state that reflects an individual’s sense of weariness toward online privacy issues; they believe that there is no effective means of managing their personal information and maintaining privacy on the Internet3,4,5. The two key dimensions of online privacy fatigue are cynicism and emotional exhaustion5. Cynicism is defined as “an attitude toward an object accompanied by frustration, hopelessness, and disillusionment”5. Emotional exhaustion indicates a “chronic state of emotional and physical depletion”5.

Privacy fatigue has various causes and effects. The causes of this state are the complexity of measures to manage individuals’ own personal data, feeling a loss of control over these data, and frequent exposure to data breaches5. The effect is that users with privacy fatigue tend not to engage in privacy-protective behavior17.

Few studies, as illustrated in Table 1, have explored privacy fatigue, its antecedents and/or consequences, despite its strong effect on behavior. Most of the existing privacy fatigue studies discuss contexts that contain massive amounts of sensitive personal information, such as mobile apps6, mHealth22, e-government17, and the Internet of Things (IoT)8. Additionally, several antecedents that induce privacy fatigue have been proposed, such as personality traits6, privacy policy effectiveness22, users’ knowledge8, and privacy information transparency17.

Table 1 Privacy fatigue studies.
Full size table

Prediction of psychology using ML

The prediction of human behavior, personality, and emotions in general, and using data from social media platforms, in particular, is receiving tremendous attention among researchers23. Some of these studies are summarized in Table 2. For instance, ML is used to predict aggressive behaviors on these platforms, such as cyberbullying. A review of cyberbullying prediction studies found that the main data collection strategy used data extracted from social media, either by using keywords, such as hashtags, or by using user profiles10. The study also stated that the ML algorithm most often applied in cyberbullying prediction is Support Vector Machine (SVM), followed by Naïve Bayes (NB).

Table 2 ML algorithms used for behavior and psychology prediction based on social media data.
Full size table

Social media users’ posts, comments, and likes are used to study and predict users’ personality traits. One study predicted personality traits based on color preference and posted photos and found a relationship between them9. The study used tree-based algorithms, such as Random Forest (RF), for the proposed model. The result showed that all the models performed similarly and accurately, with no large errors or outliers in the prediction. This study used personality traits as response variables, whereas other studies used them as explanatory variables. For example, a study24 considered users’ personality traits and smartphone usage for happiness recognition.

ML is also used to predict human emotions, such as anxiety and depression. A study was conducted in Saudi Arabia to predict anxiety levels during the COVID-19 pandemic for the purpose of preventing potential public mental health crises11. Data were collected using an online questionnaire and the model was trained by using these data, employing SVM and Decision Tree (DT) algorithms. The results showed that the proposed ML models demonstrated promising potential for early prediction of anxiety. Additionally, a paper25 reviewed 20 studies that used Artificial Intelligence (AI) and ML techniques to detect, predict, and analyze depression from social media platforms to prevent consequences such as suicide.

One of the reviewed studies developed a prediction model for depression using several classifiers, such as SVM, DT, and K-Nearest Neighbors (KNN)12. The outcome showed that SVM performed better compared to other classifiers, whereas regarding precision of features, recall, and F-measure, DT performed better.

The current gap

Privacy issues regarding social media ads seem attractive to scientific researchers. The previous studies in this field mainly focused on privacy concerns13,15,16,26, privacy calculus14,27, and the personalization-privacy paradox28,29. Researchers stated that users’ negative emotions affect their privacy decisions and behaviors6, including both privacy concerns and fatigue. However, studies showed that privacy fatigue affects privacy behavior substantially more than privacy concerns5,6.

Privacy fatigue has spread widely; however, there are still few works existing on this phenomenon (Table 1)5,6,17. Thus, this research contributes to discovering the privacy fatigue phenomenon in the social media context, a gap that has not yet been explored. This is done by investigating whether the frequent collection and use of personal data for social media ads influences users’ privacy fatigue.

This study also contributes to the field of knowledge by examining whether privacy fatigue will continue, even if users have a high IPA level and if personality traits have an impact. According to the literature, an additional gap was found where these two privacy fatigue antecedents were discussed by only a few studies and in different contexts, which can be illustrated by two main points. First, the effect of users’ privacy awareness and knowledge of privacy fatigue was examined only in the context of the IoT8. Second, although personality traits were explored in both privacy fatigue and social media fatigue studies7, the privacy fatigue study examined personality traits only in the context of mobile apps in general6.

Additionally, to the reader’s knowledge, ML has been used to predict several human psychological issues, as discussed in “Prediction of psychology using ML" section, but not privacy fatigue, which highlights the last gap. Therefore, this research contributes to predicting users with privacy fatigue using ML, where users’ IPA level and personality traits play a significant role as predictors.

Research model and hypothesis development

The model of this study is shown in Fig. 1. Details regarding the study hypothesis are illustrated next.

Figure 1
figure 1

The theoretical model.

Full size image

Information privacy awareness

Users with privacy fatigue suffer from a lack of privacy-protective behavior5. A previous study17 explored the literature to find factors that motivate such behavior to examine its effect on privacy fatigue and found that information transparency motivates privacy behavior. The study’s results confirmed lower privacy information transparency would increase cynicism and emotional exhaustion. From the literature, IPA also motivates privacy-protective behavior30, which is needed for users who suffer from privacy fatigue. A study that explored the privacy fatigue phenomenon in the IoT environment found that users’ security awareness and knowledge about IoT affected their privacy fatigue8, such that a user’s low level of IoT security knowledge is likely to increase privacy fatigue. Based on that, the first hypothesis of this study is:

H1

IPA level has a significant relationship with privacy fatigue.

Personality trait

Personality traits have been associated with privacy fatigue. A study of privacy fatigue antecedents and consequences suggested that future research must incorporate personality traits to understand better which social personality traits of media users make them highly susceptible to experiencing fatigue31. A study7 examined only the moderating role of neuroticism and extraversion traits. They found from the literature that these two traits may influence social media fatigue the most7. Their results showed that extroverted users are less likely to be concerned regarding privacy. On the other hand, neurotic users feel more insecure and disturbed and have intense feelings of invasion of privacy. However, they reported the limitation of examining only two traits.

Additionally, a study6 found that all five personality traits were significantly related to privacy fatigue, with neuroticism being positively related, and the other traits negatively related. Therefore, it is hypothesized that:

H2

Neuroticism has a significant relationship with privacy fatigue.

H3

Extraversion has a significant relationship with privacy fatigue.

H4

Openness to experience has a significant relationship with privacy fatigue.

H5

Conscientiousness has a significant relationship with privacy fatigue.

H6

Agreeableness has a significant relationship with privacy fatigue.

Materials and methods

Procedure

The four main stages followed by this study are Collect, Assess and Clean, Analyze, and Model (Fig. 2). The Collect stage starts by finding the appropriate instrument to collect the required data. The selected instrument was a questionnaire adopted from the literature. Minor modifications were made to the adopted measurements to match the study purpose, along with English-to-Arabic translation. The evaluation method of back-translation was used to ensure translation quality5,7. Before the main questionnaire distribution, a pilot questionnaire was carried out and final modifications were made to address challenges and ensure clarity and reasonable response time. It was conducted using a convenience sampling method. The data collected from the pilot study was not included in the main study to avoid bias issues associated with convenience sampling32. Thereafter, the main questionnaire was distributed.

Figure 2
figure 2

Methodology brief.

Full size image

The Assess and Clean stage started after receiving a number of responses and closing the questionnaire. The main purpose of this stage is to facilitate and improve the Analyze stage to get valid and reliable results. Therefore, data assessment was carried out to address issues related to data content and structure, i.e., completeness, validity, accuracy, and consistency. In the Clean stage, the following has been done:

  • Delete unnecessary columns.

  • Rename columns to short, meaningful names.

  • Delete some records based on exclusion criteria (e.g., untargeted users and identical responses).

  • Remove Arabic text.

  • Transform answers to a numerical scale.

  • Fill in missing values with the most frequent answer.

  • Correct some "Other" answers if the answer is already in the options above.

  • Reverse the scale of some personality trait questions.

  • Split some columns.

The Analyze stage included descriptive statistics and inferential statistics. Descriptive statistics include finding Mean, Median, Standard Deviation (SD), maximum (Max), and Minimum (Min), while inferential statistics include discovering p-values and coefficients. These statistics helped to find the relationships between variables and validate the hypothesis.

The previous stages built a solid basis for the Model stage, in which an intelligent ML model was built to predict users who were likely to suffer from privacy fatigue. This involved splitting the data using five-fold cross-validation, training the model using several classification algorithms based on the literature, i.e., SVM, NB, KNN, DT, and RF, evaluating the classifier accuracy, recall, precision, and F1, and finally selecting the most accurate algorithm based on the evaluation results.

Measurement development

Measurements of all the study variables were derived from related works5,33,34,35. Some of the questions were slightly modified to match the study’s purpose and scope (Table 3). All questions were measured using a five-point Likert scale ranging from “Strongly disagree” to “Strongly agree”.

Table 3 Measurement of the variables.
Full size table

Fatigue, in general, is characterized by emotional exhaustion and cynicism, as discussed in the literature5. Therefore, the adopted scale measuring individuals’ privacy fatigue is based on these two core dimensions, each of which has 3 items to measure it5. This scale was selected because it was the first scale developed to measure privacy fatigue; moreover, it has been adopted by all the subsequent studies and its validity and reliability have been reported17,22.

To measure IPA, two models were combined. Several studies have measured IPA by measuring privacy concerns, assuming that individuals’ concerns about future risks could be associated with their awareness of potential risks36,37. However, to measure IPA directly, a model called Information Privacy Awareness (IPA) indicated that researchers should consider three aspects that make up privacy awareness34. These aspects cover the awareness of:

  1. (1)

    The element related to information privacy.

  2. (2)

    The element’s existence in the current environment.

  3. (3)

    The element’s impact, where the element could be technology, regulation, or practice.

In this study, the element is practice: The users’ awareness level of the practice of collecting, using, and sharing personal data for social media ads was measured.

Considering these three aspects, a scale called the Information Privacy Concerns (IPC) model, was adopted33. The model is comprehensive and has been employed by several studies to measure internet privacy concerns, where awareness researchers chose the appropriate dimensions and items of the model based on the purpose of their research22,30. The model includes several dimensions of internet privacy concerns, such as personal information collection, unauthorized secondary usage, and improper access. Each dimension includes a set of valid and reliable items. Therefore, appropriate items were chosen from the IPC model33, with respect to the IPA aspects noted above34. Items were modified to address awareness instead of concerns, so as to measure awareness directly, as suggested by Correia and Compeau (2017).

To measure participants’ personality traits, the Big Five Inventory (BFI-10) scale was employed35. Considering participants’ limited time, there are two scales that offer only 10 items of the BFI: the Ten-Item Personality Inventory (TIPI) and BFI-1035,38. Both measure extraversion, neuroticism, openness, conscientiousness, and agreeableness traits using two items for each. Both have proved their simplicity, reliability, and validity. However, BFI-10 is the most recent and was clearer when translated into Arabic. Additionally, a study mentioned that the BFI-10 could be an option if the author is not interested in inferring individual differences39. This study does not aim to investigate individual differences but to capture the overall effect of personality traits on privacy fatigue.

Sample

The demographics of the respondents are summarized in Table 4. The sampling technique used for the main study is probabilistic sampling, particularly simple random sampling32. This technique was used to ensure an unbiased random selection and a representative sample where each member of the population has an equal chance of being selected, which results in more accurate and generalizable results40.

Table 4 Demographics of the sample.
Full size table

All methods were carried out in accordance with relevant guidelines and regulations as this research includes human participants. The experimental protocols and procedures were approved by the research committee of the Information Systems department at King Abdulaziz University, as it follows common and predefined regulations, and does not expose any personal information nor include any dangerous or harmful activities. However, the following guidelines were applied based on the committee’s recommendations to maintain confidentiality and anonymity, increase the response rate, ensure the quality of the data, and gain the respondents’ trust:

  • The purpose of the study was declared at the beginning of the questionnaire.

  • It was clearly stated that the data will be collected and used for study purposes only.

  • It was made clear that respondents had the right to skip any question or to stop at any stage if they did not want to continue.

  • Participation consent was required: therefore, respondents had the opportunity to agree or decline to participate.

  • Questions that identified respondents were not included.

The main questionnaire was distributed from October 10, 2022, to October 30, 2022, through social media platforms such as Instagram, Snapchat, WhatsApp, and Telegram. The connections on the authors’ social media are mainly from the Arabian Gulf which indicates a specific targeted culture sample. The questionnaire applies to individuals proficient in Arabic and English as it was distributed in both languages. After completing the data cleaning, the number of responses obtained was 508 out of 538.

Female respondents formed the majority of the sample, at 399 individuals (79%). 260 participants fell into the 20–34 age group (51%), and 333 participants had a Bachelor’s degree (66%). The occupations showed close percentages, with 116 students (23%), 116 private sector employees (23%), 112 unemployed participants (22%), and 108 employees of public/government organizations (21%). Regarding the sample’s social media usage, 502 participants (99%) used social media daily and 225 respondents (45%) reported spending 3 h more on average on social media daily.

Data analysis justification

The variables of this study, i.e., privacy fatigue, IPA, and personality traits are latent variables41,42. This implies that each variable is measured using multiple questions. Various methods can be used to analyze latent variables, such as structural equation modeling (SEM) or taking the sum/mean of all the questions’ scores for a particular variable41,42. The advantage of SEM is that it considers the reliability and validity of the questions during the analysis. However, with a small sample size, SEM will do poorly in discovering the actual effect42. Therefore, considering the sample size of this study and assuming that all questions are reliable and valid based on the literature, the mean of the answers of each variable was taken. There are 2 questions for each personality trait, 8 for IPA and 6 for privacy fatigue. The answers to these questions ranged from 1 to 5, as the options were a 5-point Likert scale. The average of each variable’s answers was calculated and used for analysis.

However, a classification for privacy fatigue was required in this study (0 or 1). Therefore, after taking the average, people who scored 3 or more were classified as having a medium-to-high level of privacy fatigue and labeled 1. Those scoring lower than 3 were classified as having a low level of privacy fatigue and labeled 0.

Model development

Supervised machine learning was used to develop the prediction model, as it used a labeled dataset. Classification algorithms were applied to classify the users into two classes, i.e., low, or medium-to-high level, of privacy fatigue. The model predicts privacy fatigue based on 5 personality traits and IPA all holding values of 0 or 1. The dataset size is 508. As the dataset cleaning was conducted before the analysis, no further data preprocessing was required during the training.

For model development, several steps were taken. The classification algorithms that were used in this study are SVM, NB, KNN, DT, and RF. The selection is based on the algorithms used in the literature (Table 2). Other algorithms on the table were not selected: Maximum Entropy, GBT, and ET, because Maximum Entropy is a text classifier, which is not needed, while GBT and ET perform similarly to RF which is used in this study9.

K-fold cross-validation is considered the most reliable and profound method that helps avoid overfitted models, one of the most common issues in ML43. To get the best learning result and high accuracy, it is necessary to maximize the size of the testing and training dataset, which is what cross-validation does because it uses all the data for training as well as for testing43. In short, to perform cross-validation, the data is divided into a number of equal-sized subsets called folds. This is done in iterations, where each fold is removed once, and the rest of the folds are used as a training set. Then, the removed fold is used as a test. The accuracy of each iteration allows them to be compared, with highly divergent accuracy values indicating errors. The overall accuracy is calculated by taking the average accuracy of all the iterations. The average of F1, recall, and precision is also considered. Cross-validation using five or ten-folds is preferred based on empirical evidence44. In the previously discussed studies, three used ten-fold cross-validation9,11,12. However, five-fold cross-validation is used in this study as the dataset is not large.

Ethical approval

The Information Systems Department research committee at King Abdulaziz University approved the experimental protocols and procedures. There were several measures advised by the committee presented in "Materials and methods" section. A written informed consent was obtained from the participants for the study before they conducted it. All methods were carried out in accordance with relevant guidelines and regulations.

Results

Descriptive statistics

The descriptive statistics of the study variables are summarized in Table 5. Personality traits and IPA values ranged from 1 to 5. The min value for most of these variables is 1, except for agreeableness and conscientiousness. The conscientiousness level started with 1.5, implying that no one in the sample considers themselves completely impulsive, careless, and disorganized. The min agreeableness score is 2; thus, the sample does not include people who 100% see themselves as critical, uncooperative, and suspicious. For all the variables discussed, the max value recorded is 5. Privacy fatigue, however, has only two values: the min value is 0, meaning that there are participants who have a low level of privacy fatigue, and the max value is 1, where participants have a moderate-to-high level of privacy fatigue.

Table 5 Descriptive statistics.
Full size table

The mean of the IPA and personality trait variables ranged from 2.6929 to 4. The mean values for neuroticism and openness traits were 2.6929 and 2.8967, implying that the participants’ tendency toward unstable emotions and active imagination is moderate on average. Also, on average, participants have neutral to high levels of extraversion, conscientiousness, and IPA, with mean values of 3.4390, 3.8051, and 3.8216, respectively. Participants with the agreeableness trait are helpful, trusting, and empathetic, and these form the majority of the sample with an average value of 4. Additionally, on average, participants tend to have a medium to high level of privacy fatigue, with a mean value equal to 0.7776.

Inferential statistics

For inferential statistics, regression analysis was conducted. This analysis is used to see how well the measures of personality traits and IPA level in the sample can predict the measure of privacy fatigue. The analysis is done by assessing several values, which are explained next.

The values that were explored are the coefficient of determination (R2), path coefficient, and path significance (p-value). R2 reflects the effect of all the independent variables combined on the dependent variable. Another way to view R2 is that it is a measure of the model’s predictive accuracy. This effect ranges from 0 to 1 where 0.75, 0.50, and 0.25 represent essential, moderate, and weak levels of predictive accuracy, respectively45. Path coefficients represent the relationships by values ranging from − 1 to + 1. Coefficients closer to + 1 indicate strong positive relationships, and coefficients closer to − 1 indicate strong negative relationships45.

In order to accept or reject the null hypothesis, the level of significance should be defined. In related studies, three significance levels are used, i.e., 0.001, 0.01, and 0.055,6,22. Therefore, these three levels are also used in this study. P-values smaller or equal to these significance levels indicate a significant relationship, meaning that the null hypothesis is rejected.

As shown in Fig. 3, 2 relationships were significant, while 4 were not. The values on the arrows represent the path coefficient while the p-value is represented by one, two or three asterisks depending on the significance level, or by N.S if not significant.

Figure 3
figure 3

Theoretical model results (*p < 0.05,**p < 0.01,***p < 0.001, N.S., Not Significant).

Full size image

The R2 result showed that around 5% of the variance in privacy fatigue was explained by IPA and personality traits, indicating a weak level of predictive accuracy.

IPA level was found to have a significant positive relationship with privacy fatigue (coefficient = 0.1301, p = 1.44*10−6). This indicates that an increase in IPA level results in an increase in privacy fatigue level.

Four personality traits out of five did not significantly affect privacy fatigue. Agreeableness had the most insignificant effect among other traits, with a coefficient equal to 0.0023 and a p-value equal to 0.9422. This was followed by neuroticism and openness, with similar p-values equal to 0.7327 and 0.7243, and coefficients in opposite directions equal to 0.0087 and − 0.0080 respectively. The last trait that had an insignificant effect on privacy fatigue was extraversion (coefficient = 0.0219, p-value = 0.3208).

Conscientiousness is the only personality trait found to have a significant negative effect on privacy fatigue (coefficient = − 0.0630, p-value = 0.0272). This means that an increase in the conscientiousness trait, i.e., people who are competent and dependable, resulted in a decrease in the level of privacy fatigue.

In summary, the null hypotheses of H1 and H5 are rejected. For H2, H3, H4, and H6, the findings failed to reject the null hypothesis. All results are summarized in Table 6.

Table 6 Hypothesis testing results (*p < 0.05,**p < 0.01,***p < 0.001).
Full size table

Comparison of the models

The accuracy, F1 Score, recall, precision metrics were calculated to compare the models (Table 7). The SVM and NB models had the highest accuracy (78%), followed by RF (75%) and KNN (73%). The lowest accuracy was for the DT model, with 65%. These results imply that SVM and NB had the highest proportion of correct predictions out of the total number of predictions. Likewise, based on the F1 score, SVM and NB had the best F1 score (87%), followed by RF (85%), KNN (84%), and DT (77%), which means that SVM and NB models had better precision and recall compared to other models.

Table 7 Comparison of models.
Full size table

The recall percentage of the algorithms ranged from 76 to 100%. The SVM algorithm had the highest recall percentage, which tells us that 100% of the actual users with medium-to-high privacy fatigue were correctly identified. This result could be because of the robustness of SVM to the noise in the data, as the decision boundary (called hyperplane) is determined by the closest data points to the boundary (called support vectors)46. NB, RF, and KNN also had high recall percentages of 98%, 93%, and 91%, respectively. In contrast, the DT algorithm, with the lowest percentage, only correctly predicted 76% of users with privacy fatigue.

Most of the algorithms had 78% precision. This implies that, among all the users that were predicted to have medium-to-high privacy fatigue, using the SVM, KNN, and RF, 78% truly belonged to this class, and 79% using NB.

In summary, as shown in Fig. 4, the SVM and NB classifiers performed slightly better than the others. Both performed similarly based on the four metrics. This suggests that SVM and NB had greater accuracies in predicting users with privacy fatigue in relation to social media personalized ads based on personality traits and IPA.

Figure 4
figure 4

Comparison of models.

Full size image

Discussion

The results of this study could be used in several real-world applications. The ML model may help communities to detect users who are more exposed to privacy breaches as a result of privacy fatigue. For example, schools could use this model to find students with a specific personality trait and IPA level to provide courses about online privacy. Based on the results, a student who is careless, impulsive, and aware of information privacy issues is more eligible for this course. They can learn how to protect themselves from privacy breaches and feel less cynical and emotionally exhausted regarding their privacy. On the other hand, the model could also help marketers to promote these kinds of courses to the right audience.

Some weaknesses and strengths of this study compared to previous studies on privacy fatigue (Table 1) should be discussed. A strength of this study is the sample size of 508. 4 out of 5 studies have a lower sample size; only one study has more (620 participants)17. Additionally, this study is the only one incorporating ML. Using ML shows how the results could be useful in real life by predicting users likely to suffer from privacy fatigue. On the other hand, most of the related studies used the SEM method5,6,17,22, unlike this study. This method ensures the validity and reliability of the measurement model before testing the structural model. This weakness is illustrated next in “Limitations and future work” section.

Theoretical implications

To the best of our knowledge, there are few studies on the privacy fatigue phenomenon. Therefore, this research provides theoretical support for a better understanding of online privacy behaviors in the social media context. This study also has implications in highlighting that, even if users have a high level of IPA, they will still have privacy fatigue. It also focuses attention on the weak effect of personality traits on privacy fatigue.

Based on the literature, IPA motivates privacy-protective behavior needed for users who suffer from privacy fatigue30. In addition, a study8 stated that a low level of security and privacy knowledge is likely to increase privacy fatigue, showing a positive relationship. However, our findings showed that IPA negatively correlated with privacy fatigue. This result may corroborate the fact that several studies have stated that individuals’ privacy concerns are associated with their awareness of potential privacy risks34. These studies measured awareness by measuring privacy concerns, assuming that a high level of privacy concerns implies a high level of awareness. Therefore, the same could be interpreted with privacy fatigue based on this study's results. The more people are aware of privacy issues, the more frustrated they become.

In various studies, personality traits showed a significant relationship to psychological fatigue6,7, contrasting with this study. This could be due to several reasons. One study6 concluded that, in different contexts, there is an apparent conflict between the effect of personality traits on privacy fatigue and privacy concerns. The other study7, used the long version of BFI (BFI-44) to measure personality traits. In contrast, in this study, the short version (BFI-10) was used here, which could affect the validity and reliability of the results. Therefore, the insignificant effect of personality traits on privacy fatigue could be due to the context of this study or the selected measurements.

Managerial implications

For managerial implications, social media operators should be aware of the existence and potential effect of privacy fatigue on future services. This psychological state could lead to users’ dissatisfaction with using social media5. Policymakers also need to recognize the phenomenon and the extensive use of users’ data. Although governments have set regulations on information privacy, these should be enforced to meet an acceptable level of privacy protection.

Limitations and future work

This study has some limitations that could be improved in future works. The limitations include the self-report questionnaire, questionnaire validity and reliability, the short version of personality traits’ measurement, and privacy fatigue classification.

The self-report questionnaire is susceptible to response biases because it relies on respondents reporting about themselves47. Participants may give answers they consider more socially acceptable, rather than being honest, or may fail to assess themselves accurately. In this study, most participants stated that they are aware of social media practices and their impact, which may not be the case; there is huge uncertainty about social media practices, even among social media owners and developers48. Future work may consider collecting data using a different method, such as observation.

The validity and reliability of the study questionnaire were not considered in this study. The SEM method could be used instead of taking the average of the latent variable scores, because it considers the validity and reliability of the questionnaire. SEM was not selected in this study because of the sample size42. Additionally, the short version of personality traits was used, with only two questions for each trait. If one of the questions were not valid or reliable, it could not be deleted because SEM calculates validity and reliability based on several questions49. Although the validity and reliability of the BFI-10 are affected by factors such as age, culture, and language39,50, this study used this scale to capture the overall effect of personality traits on privacy fatigue and not to compare individuals’ differences. However, the long version of personality traits and the SEM method are suggested for use in future research to ensure higher validity and reliability and to capture the full complexity of an individual's personality.

In the ML stage, privacy fatigue was classified into 0 and 1 based on the Likert scale. A study on predicting cyberbullying on social media stated that a comprehensive investigation is required to define and categorize the severity of cyberbullying from social and psychological perceptions10. Similarly, efforts from several disciplines are required in the future to identify the levels of severity of privacy fatigue.

Finally, a future study may be conducted based on real privacy-fatigued users. This should involve collaborating with medical professionals/institutes to provide real, comprehensive, and rich samples for more accurate performance.

Conclusion

This study explored privacy fatigue regarding social media ads. The data were collected using an online questionnaire. The relationships between personality traits and IPA (independent variables) and privacy fatigue (dependent variable) were discovered using regression analysis. ML models were also developed to predict privacy fatigue, using five classification algorithms: SVM, KNN, DT, RF, and NB. The models were evaluated using accuracy, recall, precision, and F1 metrics. The aim was to uncover the impact of social media practices and help target users who may be susceptible to privacy fatigue to motivate their privacy-protective behavior.

The results showed that privacy fatigue exists regarding social media ads. 395 out of 508 participants had a moderate-to-high level of privacy fatigue, which answers the extent of users’ emotional exhaustion and cynicism regarding personalized ads on social media. IPA, i.e., awareness of the collection and use of data for social media ads and its impact had a significant positive relationship with privacy fatigue. Among the five personality traits, i.e., extraversion, neuroticism, openness, conscientiousness, and agreeableness, only the conscientiousness trait had a significant negative association with privacy fatigue. The models with the highest accuracy in predicting privacy fatigue based on IPA and personality traits were SVM and NB, with 78% accuracy and 87% F1.