This study was conducted in Benin (Figure 1), a West African country characterized by geographical and sociocultural diversity, across a range of representative sites covering the country’s main tourist areas. It first covers the southern coastal and urban hub, centered on Cotonou, the country’s economic capital, and its immediate surroundings, notably Ouidah, a coastal city with significant heritage value marked by the history of the slave trade, as well as Grand-Popo, a seaside

Figure 1. Geographic distribution and spatial representation of the study sites in Benin.

resort renowned for its beaches and biodiversity. The southeastern region is represented by Porto-Novo, the administrative capital with a rich cultural heritage, and by Ganvié, an iconic lakeside village for tourism. The study also extends to the north of the country, covering Parakou, the main urban center of the Borgou department, as well as the commune of Nikki, the cultural hub of the northeast.

The data used comes from a survey of 430 tourists visiting several major sites in Benin between May and September 2025. The tourist sites surveyed include coastal areas such as Fidjrossè and Grand-Popo beaches, the lakeside town of Ganvié, and the royal palaces of Nikki and Parakou. It also includes major heritage and historical sites such as the Gate of No Return, the Slave Route, the open-air museum, Nonvitcha Square, and the Villa Karo Cultural Center in Grand-Popo. Sites of religious and traditional significance, such as the Zangbéto Temple, the Temple of the Pythons, the Abessan Honto Vodun Temple, and the Sacred Forest of Kpassè, have also been included.

To examine the relative probability of choosing a destination compared to a reference destination, we consider a multinomial logistic regression model. For k = 1 , ⋯ , K − 1 , where K denotes the number of different types of destinations. The model is written as:

where X i = ( 1 , x i , 1 , ⋯ , x i , p ) ⊤ et β k = ( β 0 , k , ⋯ , β p , k ) ⊤

s the parameter vector associated with destination k , and π i , k is the probability of choosing a given destination.

The predicted probabilities are given by:

The full parameter vector is β = { β 1 , ⋯ , β K − 1 } and is estimated using the maximum likelihood method. The likelihood function is given by:

and the corresponding log-likelihood:

Due to the nonlinearity of the model, the parameters are obtained through numerical optimization using the Newton-Raphson algorithm, which is based on iteration:

where h denotes the iteration.

Since the perceived attractiveness of the destinations is ordered, we adopt an ordinal logistic model with proportional odds.

For j = 1 , ⋯ , K − 1 , the model is written as:

We obtain the cumulative probabilities:

where α j is the intercept associated with category j and β is the vector of coefficients common to all categories.

The central assumption of the model is that of proportional odds: the vector β is identical for all categories of the variable of interest; only the intercepts α j vary.

This model allows us to identify the factors influencing the cumulative probability of belonging to an attractiveness level at most equal to j , while taking into account the natural order of the categories.

The full vector of parameters is estimated using the maximum likelihood method. The likelihood function is written as:

and the corresponding log-likelihood:

Random Forests are an ensemble model based on the aggregation of decision trees constructed independently using the bagging principle. Introduced by [10], this approach aims primarily to reduce the variance of a single tree while maintaining low correlation among the individual models.

Let D = { ( x i , y i ) } i = 1 n be a training sample. For each tree b = 1 , ⋯ , B , a bootstrap sample D ( b ) is drawn with replacement from D . At each node, a random selection of m variables from among p predictors ( m < p ) is considered to determine the best partition.

The aggregated predictor is given by:

In regression, where f b denotes the b -th tree.

In classification, the prediction is based on a majority vote:

Each tree is very deep, which results in low bias but high variance; aggregation then helps stabilize the predictions. Random selection of observations (bootstrapping) and explanatory variables ensures that the trees differ from one another, which improves the overall accuracy of the predictions.

The total variance of a prediction f ^ R F ( x ) can be formalized as follows:

where ρ is the average correlation coefficient between two trees, σ 2 is the variance of a single tree, and N is the total number of trees in the forest. The method also allows for estimating the importance of explanatory variables using measures such as the mean decrease in impurity or the mean decrease in accuracy, and for evaluating the model’s performance without explicit cross-validation using Out-of-Bag (OOB) observations.

Boosting is an ensemble method that uses sequential aggregation to reduce the bias of a basic model with low bias and high variance. Each model uses information from previous models and seeks to improve the prediction.

[11] proposed XGBoost, a regularized and optimized version of gradient boosting. The model is additive and is written as:

where each f k is a decision tree belonging to the space ℱ .

The estimation is based on minimizing a penalized objective function:

where ℓ denotes the loss function (mean squared error in regression or log-loss in classification) and introduces regularization on the complexity of the trees.

where T is the number of leaves and w j are the weights associated with the leaves.

The optimization is performed sequentially using a second-order Taylor approximation of the loss function, utilizing the gradient and the Hessian. The trees are constructed by maximizing an analytical gain associated with the partitions, and the predictions are updated according to:

where η is the learning rate and q ( x i ) denotes the leaf to which the observation x i is associated.

In this section, several variables were examined: the destination, the destination’s perceived appeal among tourists, and the tourists’ country of origin.

The breakdown of tourist destination types (Figure 2) reveals a marked predominance of cultural and historical destinations, which account for 44.2% of the sample. This high proportion suggests that a significant share of tourists are motivated by a desire to explore local heritage, history, and traditions. This result likely reflects the promotion of the national cultural offering and the particular appeal of most historical sites in the study area. Spiritual destinations (26.7%) reflect a growing interest in spiritual well-being. Community destinations (14.4%) and seaside destinations (14.7%) have similar shares, indicating a diversification of motivations between authentic encounters and coastal leisure activities.

The distribution of perceived destination attractiveness (Figure 3) shows that the “strong” category clearly dominates over the others. This result indicates that most destinations in the sample are considered particularly attractive, reflecting either a high quality of tourism offerings or a good match between tourists’

Figure 2. Breakdown of tourist destinations.

Figure 3. Distribution of the perceived attractiveness of destinations.

expectations and the destinations’ characteristics. The low frequency of the “low” rating indicates an overall positive perception, while also identifying room for improvement for a small number of less attractive sites.

An analysis of tourist origins shows that Benin attracts visitors from several continents (Figure 4). The significant presence of tourists from Nigeria, Senegal, Togo, Gabon, and other countries underscores the importance of regional tourism, which is facilitated by easy access and cross-border exchanges. Benin thus emerges as a major regional tourism hub. The presence of visitors from Europe (France, Belgium, Germany) and the Americas (Canada, United States) attests to the international appeal of Benin’s tourism offerings. This positive trend can be attributed to several government initiatives, notably: the modernization and

Figure 4. Breakdown of tourists by country of origin.

restoration of certain tourist sites, the implementation of the e-Visa, the promotion of intangible cultural heritage, and the enhancement of site security.

To deepen our analysis, we compared several variables using a graphical representation.

Figure 5 reveals marked differences in destination choices between men and women. Spiritual destinations attract 58.3% of female visitors, while other destinations attract more than 60% of men. We therefore conclude that women express a greater need for transformative experiences and personal development. This disparity suggests that certain tourism offerings specifically target a female audience, possibly for reasons of safety or thematic suitability.

The following box plot reveals significant differences in the distances traveled depending on the type of destination (Figure 6). The spiritual destination has the highest medians, indicating long-distance appeal. In contrast, community destinations primarily attract local or regional visitors. The variation in distances for cultural and historical sites contrasts with the consistency seen among beach destinations.

Figure 5. Visitor numbers by gender.

Figure 6. Distance traveled by destination.

4.2.1. Multinomial Logistic Regression

Backward selection based on the Akaike Information Criterion (AIC) reduced the initial model of 29 variables to a parsimony-based model of 19 explanatory variables, with a notable decrease in the AIC (from 845.02 to 796.64).

The multinomial logistic regression analysis shows that several variables are statistically significant, indicating that they influence the choice of tourist destination relative to the community reference. Among these variables are: age, gender, marital status, education level, monthly income, distance, reason for visit, average expenditure, cost of living at the destination, transportation costs, attractiveness, as well as certain satisfaction indicators such as facilities and accommodations, safety, and weather conditions.

Conversely, certain variables prove to be insignificant, such as length of stay, as well as certain aspects of marital status, educational level, reason for the visit, and perceived attractiveness. This suggests that, statistically speaking, these factors have no demonstrable influence on tourists’ choice of seaside destination.

SeasideDestination

Table 1 reveals that the probability of choosing a seaside destination over a

Table 1. Multinomial logistic regression (seaside vs community).

community-based destination is influenced by a range of socioeconomic, cultural, and contextual variables. The results show that, with a 5% margin of error, age, education level, marital status, cost of living and transportation, participation in cultural events, quality of facilities and accommodations, as well as weather conditions have a significant influence on the choice of this type of destination compared to tourists aged 26 to 50. In particular, tourists over the age of 50 are much more likely to choose a seaside destination. Conversely, those who are divorced are significantly less attracted to this type of destination.

The results also show that tourists who perceive the cost of living to be moderate and transportation costs to be affordable are more likely to prefer beach destinations over community-oriented ones. Furthermore, participation in cultural events increases the likelihood of choosing a beach destination by a factor of 9.78, reflecting the importance of cultural and social factors in visitors’ motivations. Conversely, satisfaction with accommodations significantly reduces the likelihood of choosing this type of destination, which can be explained by a hospitality offering that is sometimes standardized and less focused on the authenticity sought by certain types of tourists. Similarly, tourists who visited the sites under unfavorable weather conditions notably characterized by cloudy skies without precipitation, frequent rain, or variable weather are significantly less likely to choose a seaside destination, highlighting this tourist destination’s high sensitivity to weather fluctuations.

These findings suggest that the development of coastal tourism in Benin should be based on a strategy that combines improvements in the quality of hospitality infrastructure, the diversification of recreational and cultural activities, and the sustainable management of the coastline. By investing in the modernization of tourism facilities, the promotion of enriching experiences, and responsible environmental management, it would be possible to enhance the competitiveness and appeal of coastal areas.

Cultural & Historical Destination

Table 2 reveals that the probability of choosing a cultural and historical destination over a community-based destination is influenced by several socioeconomic, cultural, and contextual variables. The results show that at the 5% significance level, age, marital status, monthly income, average spending, cost of living and transportation, distance, perceived attractiveness, participation in cultural events, quality of infrastructure, safety, weather conditions, and the reason for the visit have a significant influence on the choice of this type of destination. In particular, tourists over the age of 50 are much more likely to choose a cultural and historical destination (OR = 19.18), reflecting a preference for educational and heritage-focused travel. Conversely, divorced individuals or those with a monthly income below 500,000 CFA francs are significantly less likely to choose this type of destination, suggesting that social status and purchasing power influence tourism choices.

The results also indicate that tourists who perceive the cost of living at the destination to be moderate (OR = 18.66) and who spend less than 200,000 CFA francs (OR = 7.02) are more likely to prefer a cultural and historical destination, unlike those who consider transportation costs to be too low or too high. Furthermore, distance and perceived attractiveness play a decisive role: the more the destination is perceived as distant and attractive, the higher the probability of being drawn to it (OR = 11.88 and 8.41, respectively). Participation in cultural

Table 2. Multinomial logistic regression (cultural_historical vs community)

events is also associated with an increase in this probability (6.58 times higher odds), as is satisfaction with tourism infrastructure (OR = 11.22), confirming the importance of the cultural setting, facilities, and local atmosphere in the appeal of historic sites. Conversely, satisfaction with accommodations and safety significantly reduces the likelihood of choosing such a destination, which can be explained by tourists’ higher expectations regarding comfort and safety.

Finally, travel motivations related to adventure, culture, relaxation, and social interaction have a very significant influence, increasing the likelihood of choosing a cultural and historical destination by factors ranging from 20 to over 350. Conversely, unfavorable weather conditions (cloudy skies without precipitation, frequent rain, and changeable weather) significantly reduce this probability, highlighting this tourism segment’s dependence on favorable weather conditions.

These results suggest that the development of cultural and historical tourism in Benin should be based on the promotion of tangible and intangible heritage, the enhancement of cultural infrastructure, and improvements in the comfort and safety of sites. By investing in the creation of regular cultural events and the promotion of immersive experiences, it would be possible to enhance the appeal of this segment while contributing to the preservation of the national heritage and the sustainable diversification of Benin’s tourism offerings.

SpiritualDestination

Table 3 shows that the likelihood of choosing a spiritual destination over a community destination is influenced by several variables, such as age, gender, monthly income, average spending, cost of living and transportation, distance,

Table 3. Multinomial logistic regression (spiritual vs community).

perceived attractiveness, participation in cultural events, quality of tourism infrastructure, facilities and accommodations, as well as weather conditions. In particular, tourists over the age of 50 are significantly more likely to choose a spiritual destination (OR = 26.79), reflecting a heightened interest in introspective and symbolic experiences. In contrast, men and individuals with a monthly income below 500,000 CFA francs are significantly less likely to choose this type of destination, suggesting that spiritual tourism appeals more to women and individuals with a higher income level.

The results also show that tourists with low average spending (less than 200,000 CFA francs) who perceive the cost of living to be moderate are more likely to choose a spiritual destination, reflecting a preference for accessible environments that offer a setting conducive to tranquility and meditation. Furthermore, participation in cultural events (OR = 15.40) and a positive perception of tourism infrastructure (OR = 9.46) significantly increase this likelihood, reflecting the close link between spirituality, culture, and regional ties. Conversely, satisfaction with accommodations significantly reduces the likelihood of choosing this type of destination (OR = 0.017), likely because spiritual experiences are often had in simple, unadorned settings. The positive coefficients associated with distance (OR = 18.77) and perceived attractiveness (OR = 17.54) also suggest that tourists are willing to travel longer distances to experience spiritual encounters deemed unique and meaningful.

Finally, adverse weather conditions such as frequent rain (OR = 0.019), changeable weather (OR = 0.002), or cloudy skies significantly reduce the likelihood of choosing a spiritual destination, highlighting the vulnerability of spiritual tourism to weather fluctuations.

These results therefore suggest that the development of spiritual tourism in Benin should be based on a strategy aimed at promoting sacred sites and places of worship with strong symbolic significance, while improving access, basic infrastructure, the integration of cultural activities, and the preservation of the natural environment, it would be possible to enhance the attractiveness and sustainability of this rapidly emerging tourism segment.

4.2.2. Ordinal Regression

To analyze the perceived attractiveness of destinations, an ordinal logistic regression was estimated. The top-down selection procedure based on the AIC criterion led to the selection of a parsimony-based ordered logistic model, reducing the number of explanatory variables from sixteen (16) to nine (9), while significantly improving the model fit, as indicated by the decrease in the AIC from 630.12 to 617.7965.

Seaside Destination

Analysis of the ordered logistic regression model (Table 4) reveals that several variables have a significant impact on the perceived attractiveness of seaside destinations in Benin.

The distance variable, which is significant at the 5% threshold, indicates that the more accessible or closer a destination is, the more attractive it is perceived to be by tourists. A cost of living perceived as low or moderate has a negative effect, suggesting that tourists prefer destinations where the relative cost is reasonable but not excessively low. The quality of the environment, which received a level of satisfaction deemed positive by tourists, and political and social stability enhance attractiveness, indicating that tourists value safety and the environmental setting.

Table 4. Ordinal regression for seaside destination.

Furthermore, the reasons for visiting whether for culture, relaxation, or social interaction are highly significant and have a marked positive effect, reflecting the diversity of tourist expectations and the importance of tailoring the tourism offering to these different motivations. Conversely, the variables hospitality and information and cultural diversity are not significant, suggesting that they have a limited influence on the immediate perception of the attractiveness of seaside destinations.

The low|medium and medium|high thresholds indicate that there are perceptible differences between levels of attractiveness, although these differences are not uniform across all categories of variables.

Overall, it is primarily factors related to distance, cost of living, environmental quality, socio-political stability, and reasons for visiting that determine how attractive seaside destinations are perceived to be. This shows that tourists do not judge destinations solely on material or logistical criteria, but also on qualitative and contextual criteria that influence their perceived experience.

CommunityDestination

As shown in Table 5, several variables significantly influence the perceived attractiveness of EU destinations, while others have no notable effect.

Significant variables include the cost of living at the destination, the socio-political situation, reasons for visiting, tourist infrastructure, safety, hospitality and information, and cultural diversity. The positive effect of low and moderate cost of living at the destination suggests that tourists prefer EU destinations where the relative cost is affordable but perceived as offering good value for money.

Table 5. Ordinal regression for community destination.

A stable political and social situation and safety significantly enhance attractiveness, underscoring the importance visitors place on a safe and stable environment. The reasons for visiting, whether related to culture, relaxation, or spirituality, are also significant and positive, reflecting the diversity of tourist expectations and the influence of personal motivation on the perception of attractiveness. Conversely, distance has a negative and significant effect, while certain reasons for visiting, such as social gatherings and the quality of the environment, have less influence on perceived attractiveness.

HistoricandCultural Destination

Analysis of the ordered logistic regression model (Table 6) reveals that the perceived attractiveness of cultural-historical destinations is influenced by a combination of logistic, contextual, and motivational factors.

Distance has a marked negative effect ( β = − 0.50 ), indicating that the distance from a cultural-historical destination reduces the likelihood of it being perceived as attractive. Cost of living has a positive effect at low and moderate levels, suggesting that tourists value affordable cultural and historical sites. A satisfactory environmental quality and a stable socio-political situation enhance perceived attractiveness, underscoring the importance of a safe and pleasant setting.

The reasons for visiting show contrasting effects: culture, relaxation, and social interaction significantly increase attractiveness, while spirituality reduces the likelihood that the destination will be perceived as attractive, reflecting the diversity of tourists’ expectations. Furthermore, variables such as tourism infrastructure,

Table 6. Ordinal regression for cultural_historical destination.

safety, hospitality and information, and cultural diversity are all significant and contribute positively, reflecting the importance of the quality of service and the experience offered to visitors.

Spiritual Destination

Table 7 shows that several variables significantly influence perceptions of the attractiveness of spiritual destinations. The positive effects of low or moderate cost of living, political and social stability, safety, cultural diversity, as well as motives for visiting focused on culture, relaxation, and social interaction suggest that spiritual attractiveness is not limited to the religious or symbolic dimension of the place. It also depends on a favorable material and social environment. In other words, a spiritual space perceived as safe, culturally diverse, and economically accessible attracts more visitors. Conversely, the quality of the environment, tourist infrastructure, hospitality, and the motive of spirituality itself do not prove to be decisive factors, indicating that the spiritual quest in Benin seems to be guided more by conviviality and the social atmosphere than by the physical quality of the setting.

The negative effect of distance indicates that the farther away the destination is, the less attractive it is perceived to be, confirming that spiritual tourists are sensitive to accessibility. This finding reflects the local roots of spiritual tourism, which is often linked to community, family, or cultural practices close to home. Furthermore, the motives of relaxation and social interaction reinforce the idea that a spiritual visit is perceived not only as an experience of contemplation but also as an opportunity for socializing and cultural exchange. Furthermore, socio-political

Table 7. Ordinal regression for spiritual destination.

stability and cultural diversity appear to be essential conditions for the expression of the sense of trust and harmony that visitors seek in this type of destination. These results thus reveal a form of hybridization between spirituality and well-being, where the religious dimension coexists with a quest for security, peace, and an enriching human experience.

To prepare for model training, the dataset was first divided into training and test sets (80% versus 20%), stratified by the response variable “destination” to ensure that the distributions remained consistent between the two samples.

4.3.1. Random Forest Model

The random forest model incorporates a wider range of variables, including factors related to tourists’ characteristics and satisfaction levels. In particular, the model takes into account the variable that determines why a tourist chooses a particular destination.

Figure 7 presents the evolution of the Out-Of-Bag (OOB) error according to the number of trees in the random forest model. The error decreases rapidly during the first iterations before stabilizing around 300 trees, indicating convergence of the model. The final OOB error is approximately 0.35, corresponding to an overall accuracy close to 65%. Differences between class-specific error curves reveal that some tourist destination categories are more difficult to predict than others, highlighting heterogeneity in tourist profiles and destination characteristics.

Figure 8 presents the most influential variables driving tourists’ destination choices. The results reveal that cost_of_living, country_of_origin, and average_expense are among the most important determinants of destination selection. These variables capture key economic and socio-demographic dimensions of tourist behavior, highlighting the central role of financial considerations and travelers’ backgrounds in shaping destination preferences.

Figure 7. Error OOB.

Figure 8. Relevance of variables—Random forest.

In addition, distance, mode_of_transportation, monthly_income, and weather_ conditions exert heterogeneous effects on decision-making across individuals and contexts. This observed heterogeneity reflects the diversity of tourist profiles and suggests that destination choice is driven by the joint interaction of economic, geographic, and environmental determinants.

Furthermore, the algorithm provides an estimated overall error rate of 40% for the resulting forest. This indicates that, on average, 40% of destination predictions are incorrect. With an accuracy of 60% and an average F1 score of 57.3%, the model demonstrates moderate predictive capability, where the gap between these two metrics reveals significant inter-class performance disparities. The following graph shows the confusion matrix.

Analysis of the confusion matrix Figure 9 indicates that the cultural-historical category is predicted with 29 correctly classified observations. This performance highlights that the model is able to effectively identify tourist profiles motivated by cultural and historical exploration, reflecting the relevance of explanatory variables such as participation in cultural events, level of education, and length of stay.

Figure 9. Confusion matrix—Random forest.

In contrast, the spiritual, community, and seaside destinations exhibit much lower classification rates, with 17, 4, and 1 correctly predicted observations, respectively, highlighting the model’s difficulty in distinguishing certain classes whose socioeconomic or behavioral characteristics overlap.

These results show that, despite its ability to capture certain implicit structures in the data, the model has limitations when it comes to making subtle distinctions between destinations with similar profiles.

4.3.2. Extreme Gradient Boosting (XGBoost) Model

Figure 10 shows the relative importance of the explanatory variables in the XGBoost model. An examination of the figure reveals that the cost of living variable emerges by a wide margin as the most influential predictive factor, with an importance level far exceeding that of all other variables. This dominance indicates that the cost of living constitutes the primary discriminating factor in the classification of destinations. Next, weather conditions, country_of_origin, and length of stay exhibit relatively high importance, reflecting tourists’ sensitivity to climatic and logistical constraints associated with travel.

Figure 10.Significance of the variables—XGBoost.

A comparison between the Random Forest and XGBoost models highlights differences in variable importance patterns. Random Forest spreads importance across several economic, socio-demographic, and travel-related factors, indicating a multidimensional structure of destination choice determinants. In contrast, XGBoost concentrates importance on a limited set of key predictors, with cost_of_living as the dominant variable, followed by weather_conditions, country_of_origin, and length_of_stay, indicating a more selective and hierarchical structure.

Furthermore, the algorithm estimates the overall error rate of the XGBoost model at 37.65%, indicating that, on average, 37.65% of destination predictions are incorrect. With an accuracy of 62.35% and an average F1-score of 58.41%, the model exhibits moderate predictive performance. The discrepancy between these two metrics, however, suggests marked differences in performance across classes, revealing that some destinations are predicted better than others.

Analysis of the confusion matrix Figure 11 indicates that the cultural_historical and spiritual categories correctly predicted 29 and 17 observations, respectively. This performance highlights that the model is able to effectively identify tourist profiles motivated by heritage and historical exploration, as well as spiritual experiences.

Figure 11. Confusion matrix—XGBoost.

In contrast, the community and seaside destinations show much lower classification rates, with 4 and 1 correctly predicted observations, respectively, highlighting the model’s difficulty in distinguishing certain classes whose socioeconomic or behavioral characteristics overlap.

The XGBoost model performs slightly better than the Random Forest model, with an error rate of 37.65%, an accuracy of 62.35%, and an F1-score of 58.41%. It classifies well the cultural_historical and spiritual categories, but struggles with community and seaside destinations.

In comparison, the Random Forest model shows slightly lower performance, with an error rate of 40%, an accuracy of 60%, and an F1-score of 57.3%. It is also effective for cultural_historical destinations but faces similar difficulties for the other categories.

Overall, XGBoost is more suitable for this task because it achieves a better balance between accuracy and F1-score and better captures key patterns in the data.