Deep neural collaborative filtering model for personalized travel recommendation

Abstract

Developing personalized travel recommendation systems is significant for better user experiences. This work introduces an advanced model using Neural Collaborative Filtering (NCF) to address diverse user preferences in travel planning. Traditional collaborative filtering struggles with data sparsity, cold start, and less common items, leading to suboptimal predictions. Our novel NCF model, learned by a neural network, overcomes these limitations and captures complex user-travel relationships. Employing a multi-layer perceptron, it refines predictions based on past interactions and real-time behavior updates. Experiments on real travel data show NCF significantly outperforms traditional methods in accuracy and user satisfaction, advancing personalized travel recommendation by effectively handling data sparsity and preference diversity.

Introduction

In the context of an increasingly global economy and improving living conditions, the tourism industry is undergoing a considerable evolution towards a customization/user-centric model. In recent years, the main way of travel has been changing from traditional group tourism to independent experiences such as self-drive tours, rural leisure, and experiential travel. This move represents the evolving expectations of customization by travelers which would improve overall quality and user experience¹. Today, the tourism industry has also changed with travel being taken online and many different websites for travel planning. Such platforms give travelers more variety to base their decisions on before they can come up with well-informed travel plans. Yet, with a rapid expansion in available volumes of travel data, every user ends up encountering difficulty processing this information under a cognitive overload burden, where it is never clear which recommendations would appeal specifically to them². As the basic search functionalities provided by these platforms often fall short of capturing user demand effectively, it is essential to design intelligent travel recommendation models that can cater accurately at each step, tailoring different needs with the help of advanced technology. Tailored travel recommendations help users discover personalized destinations, accommodations, and activities by understanding their unique preferences for how they think about, plan for, or experience travel, including solo adventures or group trips. This not only provides a great list of travel options to choose from but also improves user engagement and satisfaction on online travel platforms. Additionally, personalized recommendation systems drive higher return rates and help transform the business of a digital media company. Personalization and recommender systems have been well-studied areas, with several applications across sectors such as e-commerce (e.g., Amazon) and streaming services (e.g., Netflix), to deliver tailored content that improves user engagement through widely accepted personalized models³. Considerable development of recommendation systems in the travel domain has been witnessed through diversified approaches, such as data mining-based ones, location-based social networks (LBSN), or hybrid recommendation systems for suggesting tours, packages, and attractions to users in accordance with their preferences⁴. The current phenomenon using mobile internet has resulted in a flood of travel information; hence, the only source which can curb the circumstance of information overload is recommendation systems^5,6. One commonly applied recommendation system technique is collaborative filtering. It analyzes user behavior and historical data to monitor the likes and dislikes of users and suggest relevant travel information. Nevertheless, classical collaborative filtering approaches become overwhelmed to destination sites and typical attributes such as budget constraints, travel time, etc. are overlooked. Above limitations will overcome the proposed study putting forth the personalized travel recommendation model based on NCF. With the recent growth of neural and intelligent Industry systems, it has been shown that deep neural networks can significantly out-perform traditional methods of collaborative filtering, especially when dealing with large, sparse datasets, by modelling very complex and non-linear interaction between users and items for large-scale recommendation problems⁷. Since NCF uses deep learning approaches it should cover more complex user-item interactions that should be implemented for personalization in recommendations. The system will adapt dynamically according to the input from the users with explicit ratings such as thumbs up or implicit feedbacks. Another was that the NCF-based and user-preference modeling overcame the challenge of sparse data and commonly missing rating items in travelling recommendations so that the users received more accurate travel recommendations with greater satisfaction. The main objectives of this study are outlined as follows:

1. To design a neural collaborative filtering framework that effectively models user preferences for tourist attraction recommendations.

2. To enhance similarity estimation in scenarios with limited overlapping user ratings by leveraging both local and global preference information.

3. To reduce the impact of data sparsity in tourism datasets and improve recommendation reliability.

4. To validate the effectiveness of the proposed approach through comparative performance analysis with traditional collaborative filtering methods.

The remainder of this paper is divided into sections:

Overview Summarize and discuss the previous work related to collaborative filtering algorithm, modeling of user preference, and the gaps in existing methodologies.

Neural Collaborative Filtering Approach outlines the proposed NCF model, detailing its architecture and the techniques employed to extract user preferences;

Experiments presents the experimental design, datasets used, and performance evaluation of the NCF model compared to traditional methods;

Discussion and results summarizes the findings, discusses implications for the tourism industry, and outlines potential directions for future research.

Literature review

Current research on personalized travel recommendation system

Now days the advancement of Artificial intelligence (AI) is sky-rocketing and AI has turned out to be popular for tourism which improves customers experience via advance technology^8,9. The Big Data now gives a lot of information available to the user but hardly can filter it out on relevant insights. In response to this, stakeholders have created personalized travel recommendation technology to recommend destinations according to user preferences¹⁰. These systems provide personalized decision-making support by providing customized alternatives that help in increasing the efficiency and accuracy of matching user preferences in the travel industry¹¹. Karthiyayini and Anandhi, R. J. investigates how different machine learning algorithms can effectively generate personalized travel recommendations using large-scale, heterogeneous data. The study evaluates traditional, hybrid, and deep learning models using rigorous experimental methods and performance metrics. They have discussed the insights, limitations, ethical considerations, and future research directions for improving personalization in travel recommendation systems¹². As a general outset a personalized travel recommendations include accommodations, transport, sightseeing, activities, and shopping. Recommendations can be classified as single travel product, travel package, or itinerary-focused¹¹. The rise of digital platforms has transformed travel, giving users an abundance of information for planning. However, traditional recommendation systems like collaborative filtering struggle to provide personalized suggestions, especially with travel data. Collaborative filtering faces limitations such as cold-start problems and data sparsity¹³. While content-based methods help by analysing item features, they often lead to repetitive suggestions and lack adaptability to evolving user preferences. This has spurred research into advanced techniques, such as Neural Collaborative Filtering (NCF). The NCF model replaces the traditional inner product in collaborative filtering with neural networks. Compared to others, the user-item interactions are more complexly captured. Zhang et al. exemplified studies on NCF’s effect in sparse data context¹⁴; Although NCF improves personalization, further development is necessary for synchronization with users’ actual time preferences¹⁵. Additionally, Zhao et al. emphasized sentiment analysis, which helps travel recommendation systems refine suggestions based on user reviews, making recommendations more contextually relevant¹⁶. Acharya and Mohbey did a time-aware cross-domain point-of-interest recommendation model that uses social network information to improve personalization, showing the importance of temporal and contextual signals in travel-related decision making¹⁷. Zheng et al. proposed synthetic data generation to maintain recommendation relevance in areas with sparse data¹⁸. Takashita et al. used Flickr photos tagged with location-specific information to recommend travel attractions based on user selections¹⁹, while Yang et al. applied GPS-tagged photos for similar recommendations²⁰. Alhamid et al. introduced a context-ranking method for personalized recommendations using photos from Panoramio, clustering landmarks by popularity and user-preferred locations²¹. Such route recommendation system suggests personalized travel itineraries based on user-defined routes, utilizing systems like STIMS to rank landmarks¹⁶. Traditional personalized recommendation systems usually apply methods, such as: content-based, collaborative filtering, knowledge-based recommendations, or some combination known as hybrid recommendations. Xu et al. concluded for tourism recommendation systems that both contextual information and domain knowledge were required for high relevance, however, Chen et al. suggested that knowledge- and session-based approaches would work well for tourism because they rely less on rich histories of user activity^22,23. Chen et al. developed the approach to keywords-enhanced contrastive learning using semantic travel attributes that improve representation quality. This shows the effectiveness of advanced neural learning strategies in capturing nuanced travel preferences²⁴. A personalized travel recommendation system typically contains user modelling, contextual awareness, databases of travel information, recommendation algorithms, and the way results are presented. Explicit methods include surveys, and implicit methods include browsing histories. Contextual data is sourced from GPS information and up-to-date external sources, for example, meteorological updates. A recommendation engine produces recommendations for user locations using collaborative filtering and content-based filtering algorithms, filtered through and presented over a usable interface that can be explored easily on both web and mobile platforms.

Current research on collaborative filtering recommendation algorithm

Modern online real-time recommender systems adopt techniques that fall into three classes: association rule-based recommendation, content-based recommendation, and collaborative filtering. Of these, CF is claimed to be the most applied approach. For instance, today’s NetEase Cloud Music song list, “Guess Your Favourite” in Taobao Mall, and Jingdong Mall’s product recommendations. In essence, a recommendation system analyses users’ browsing activities to forecast their preferences with the ultimate aim of recommending relevant information meeting requirements of the users.

Generally, a recommendation system has two components: online and offline. The online portion is the interface for the user, and the offline aspect holds the backend logging system as well as the recommendation algorithm itself. From the offline analysis of profiles and behavioural logs, a predictive model is produced for recommendations in new contexts and then presented in the interface of the online component²⁵. Pragmatically, CF predicts how target users might rate items they have not rated yet by learning from items they have rated before. Collaborative filtering extends the same by suggesting items liked by users with similar preferences. Matching special user interests with personalized recommendations is the main objective of tourism recommendation systems. However, collaborative filtering models have typically been affected by the data sparsity problem or failed to precisely capture user diversity. This paper proposes a novel algorithm of collaborative filtering with improved user preferences. The algorithm combines user preferences with Jeffries-Matusita distance metric and provides recommendations with improved precision that include personalized suggestions delivered for improving evaluation quality. This methodology improves precision by incorporating the variability of user preferences along with the individual rating behaviours of users. The Jeffries-Matusita distance additionally facilitates a more refined method for distinguishing user preferences, surpassing the capabilities of conventional metrics. This work seeks to develop a collaborative tourism recommendation algorithm that overcomes limitations of traditional CF methods by integrating specific user preferences. Collaborative filtering has several advantages: (1) real-time recommendation accuracy, (2) no significant technical barriers, and (3) reduced sensitivity to data sparsity²⁶.

These two versions of collaborative filtering are: Memory-Based CF and Model-Based CF. The former typically employs the nearest neighbour method whereby distances are determined in relation to the prior history of other users’ preferences and scores given by similar users are applied in order to predict preferences for certain items. It encompasses User-based CF²⁷ and Item-based CF¹⁵. It predicts ratings based on similarity among users, whereas the item-based CF predicts ratings based on similarity between items. Since these methods are differentiated based on their respective principles, each one performs differently across scenarios. User-based recommendations show community popularity, while the item-based recommendation is heavily personalized. Common similarity measures in memory-based CF include Pearson correlation, cosine similarity, Jaccard coefficient, and Euclidean distance²⁸, and no single best method exists; adaptability is the way to go. Sun et al., demonstrates that the retrospective revaluation concept of how neural systems can adjust associations even in the absence of direct input, highlighting the ability to learn from indirect signals. This principle aligns with deep neural networks’ capacity to capture complex, non-linear relationships in sparse data. Similarly, Neural Collaborative Filtering leverages such learning to model user preferences accurately, improving recommendation performance in tourism systems²⁹.

In contrast, model-based CF simulates users rating behaviour with machine learning and data mining, for instance using clustering models³⁰, Bayesian models^31,32, matrix decomposition to predict ratings over unobserved items^33,34. These are, however, only a few types of recommendation techniques used by websites including others such as association rule-based techniques or content techniques or collaborative filtering or hybrid recommendations. Collaborative Filtering is a powerful recommendation approach in the below aspects:

1. General assessment range; it pulls recommendations from user similarity instead of the content characteristics of things, so well suited for heterogeneous item types in large amount.

2. Recommendation creation; CF is independent of the content descriptor but only depends on its context compared to other products that enables both exploration and exploitation.

3. High rating sparsity; CF has less requirement on sufficient number of ratings for good accuracy (> 95%), and thus limit input space, and.

4. Low technical complexity; user preference can be easily collected through direct survey operates, while they fall into natural form as longitude data vector simple algorithm.

Collaborative filtering has proven successful, but issues such as (1) sparsity, where limited ratings provide unreliable data, and (2) scalability, as the expanding rating matrix can become too complex for real-time processing, present challenges to its continued development.

User preference based collaborative filtering algorithm

The selection of customized recommendations based on neural networks that are conducting analysis of interactions between users and items utilizes the methodology of collaborative filtering specifically in NCF. Indeed, NCF represents a critical move from traditional approaches that easily identify the non-linear relation between users and items, which is actually a critical factor due to the sparse rating data in real scenarios. This methodology is very suitable for the project proposal in the context of suggesting travel destinations and activities regarding user preferences and historical data. In real-world scenarios, user ratings for items such as travel destinations, hotels, and activities are often sparse, presenting challenges in identifying similarity among users. The lack of overlapping rated items can hinder the effectiveness of conventional collaborative filtering techniques. NCF addresses these challenges by embedding users and items in a continuous latent space, allowing for a more nuanced understanding of their interactions. This capability is crucial for the project, where recommendations need to be tailored to individual user preferences.

Methodology

Neural collaborative filtering model

The NCF model employs a multi-layer perceptron (MLP) architecture to learn the interaction between user and item embeddings. This architecture allows the model to learn complex interactions between users and items by passing the concatenated user and item embeddings through multiple layers of neurons. The recommendation process can be expressed mathematically as in (1).

1

Where:

is the predicted rating for user u on item i.

are the user and item embeddings respectively.

are weight matrices of the first and second layers of the MLP.

are biases for the respective layers.

σ is the activation function (e.g., sigmoid).

⊕ denotes the concatenation operation.

Given the user embedding U(u) and place embedding P(p), the embeddings are concatenated as in (2):

2

This concatenated vector is then processed through deep layers as in (3) (4) (5) and (6):

1. Dense Layer (256 units):

   3

2. Dropout Layer: A 50% dropout rate helps prevent overfitting.

3. Dense Layer (128 units):

   4
4. Dense Layer (64 units):

   5
5. Output Layer: Produces a single rating prediction for the user-item pair.

   6

By training the model on historical rating data, the MLP learns to capture the preferences of users and the characteristics of items, enabling it to generate accurate predictions of user ratings for items they have not yet interacted with. This aspect is particularly valuable in recommending lesser-known travel destinations that align with a user’s preferences.

Model compilation and training

The NCF model is compiled using Mean Squared Error (MSE) as the loss function, with Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) as evaluation metrics as in (7), (8) and (9).

7

8

9

The model is trained over 100 epochs with a batch size of 20, achieving optimal convergence and minimizing error rates.

Evaluation and performance metrics

The model was evaluated using a test dataset to validate its predictive accuracy. Key metrics, including MAE and RMSE, demonstrated improved performance over traditional CF models:

Closeness measurement

To enhance the recommendations further, the concept of “closeness” can be integrated using Hamming closeness, which assesses the similarity between user-item pairs. This metric allows for more refined recommendations by identifying users with similar tastes. The distance between two fuzzy sets can be calculated using the Eq. (10):

10

This formulation quantifies the similarity between user preferences based on their interactions with items, aiding in the identification of users who may have overlapping interests. Incorporating this measure into the recommendation system can help highlight travel experiences that a user may not have considered but would likely enjoy.

User rating preference

In the context of user preferences, the ratings can be represented as a vector where the ratings reflect the likelihood of a user enjoying specific items. Understanding how to calculate the closeness between users based on their ratings is critical for effective recommendations. The calculation of closeness can be defined as in (11):

11

This equation reflects the similarity between user and user based on their common ratings, facilitating the identification of potential matches for new travel recommendations. By analyzing the preferences of similar users, the system can suggest destinations or activities that have been positively rated by those with aligned tastes.

Global rating context

To address the sparse data problem, global ratings can be utilized by expanding the analysis to include all ratings rather than limiting the focus to common items. This broader perspective helps in generating more accurate recommendations. The probability distribution of ratings can be modeled as in (12):

12

Incorporating this global perspective enables the recommendation system to generate suggestions based on a wider range of data, thus enhancing its accuracy and relevance to users’ preferences.

Distance metric

The Jeffries-Matusita (J-M) distance can be used to quantify the similarity between user preferences, formalized as in 13.

13

This distance reflects how similar two users are based on their global ratings. Utilizing this metric can improve the recommendation algorithm’s ability to identify users with similar travel preferences, allowing the system to suggest experiences that other similar users have enjoyed.

Overall recommendation framework

The final NCF model for collaborative filtering recommendations as in Fig. 1 includes the principles and formulations discussed. Specifically, such user-item interaction modelling based distance metric can provide the personalized recommendation to effectively enrich customer experience and satisfaction. By taking a holistic perspective, the model ensures whether the users are recommended for what really appeal to them and these ultimately result into higher engagement satisfaction within their travel experiences. Recent studies produced neural optimization and representation learning which suggests multi-layer architectures based on non-linear functions that gives better representation of complex model structures for user-item relationships, supporting MLP-based NCF framework³⁵.

Fig. 1.

High-level architecture of travel recommendation system.

In summary, the NCF framework was defined and designed in order to refine personalization when providing members with travel recommendations for newly possible travel experiences.

Theoretical contributions

Using NCF, the model improves on existing benchmarks for personalized travel recommendations based on several key contributions. It takes an enhanced NCF algorithm, wherein combining different local and global rating data with user preferences is adaptive. Using the Jeffries-Matusita distance to represent similarity, the model captured slight variations in his different preference states which led to much more accurate predictions on where and what to travel. It addresses cold start issue by utilizing the preference data of corresponding user together with rating behaviours (aggregate) from each individual. In other words, even in the wake of the dearth of aggressive data, providing a pretty strong inference based on that which is known is still something whereby the system delivers. In addition, this paper introduces a theoretical framework to represent external data intakes during its pre-processing processes. By allowing this kind of integration, data becomes more colorful and as such provides new insights that otherwise would have not revealed itself with data collecting.

Managerial implications

The NCF-based recommendation algorithm proposed is a robust framework for future advancements of the travel industry, and the above recommendation algorithm improves the comfort and interest of users with more accurate and personalized recommendations for hotel points of interest related to tourist attractions. On the other hand, it will increase revenues and bring positive evaluations from customers for the entities involved in the travel and tourism industry. Additional external sources in pre-processing: Such an aspect adds auxiliary external sources in a travel organization to deepen the understanding of user preferences and behaviour. Information of this nature makes business decisions strategically effective and leads to better marketing targeting that can help companies express what they actually offer in a manner directly related to consumer needs. Furthermore, the inherent properties of NCF as a general algorithm, as illustrated in Fig. 2, make it an alternative to many other recommendation systems that apply deep learning in different domains, especially e-commerce and music or film recommendations. These examples of flexibility come with a wide scope of applicability in designing personalized recommendation systems that can provide tailor-made and high-quality suggestions in lots of fields.

Fig. 2.

Flow chart of neural collaborative filtering algorithm.

Results and discussions

A personalized travel recommendation system using Neural Collaborative Filtering (NCF) was developed on a real-world dataset with ratings and destination attributes like food quality, cultural sites, accommodations, and weather. The model’s performance was evaluated using key metrics such as loss, mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE). The model showed strong improvements early in training. Training loss dropped rapidly in the first 20 epochs, moving towards zero, indicating that the model effectively fit the data patterns. Validation loss stabilized around 0.4, suggesting good generalization with minimal overfitting. MAE showed similar improvements, with validation MAE stabilizing near 0.55, further confirming the model’s accuracy. These metrics indicate that predictions closely matched actual user ratings, with MAE confirming the model’s precision in personalizing travel recommendations. The dataset included various features describing aspects like cultural and shopping environments, weather, and tourist attractions. The model embedded these features into a lower-dimensional space, capturing complex relationships between user preferences and destination attributes. Key model parameters, such as 3 layers and an embedding size of 64, were optimized empirically. Comparing NCF with traditional collaborative filtering (CF) and content-based filtering (CBF) models revealed that NCF outperformed both. CF had a higher RMSE of 0.85, while CBF had 0.78, but NCF achieved a superior RMSE of 0.6688. This made it particularly effective in sparse data or cold-start scenarios, thanks to its ability to capture complex user-item interactions. The training process showed a clear progression from high error in early epochs to stable, low error by later epochs, with convergence around epoch 50, indicating that the model learned key patterns early on and refined them with further training. The final model reached near-optimal performance after epoch 100. Top-K accuracy, which measures how often the true item appears in the top K recommendations, was used to evaluate personalized recommendation strength. The model achieved a top-5 accuracy of 84%, confirming its ability to recommend relevant travel destinations aligned with individual user preferences. It also demonstrated diversity and novelty, offering users a broad array of options. Overall, the NCF model successfully personalized travel recommendations, outperforming traditional algorithms. Its robustness in capturing diverse preferences and generalizing well to new data makes it a promising solution for dynamic and personalized travel recommendation systems.

Key findings of the model

Table 2 shows an analysis of model performance metrics that are derived from both training and validation phases. This reveals how effectively Neural Collaborative Filtering (NCF) can be used for personalized travel recommendations. Figure 3 shows the training loss over epochs. As observed in Table 1, the training loss decreases significantly over time, while the validation loss also follows a similar trend, indicating that the model is learning and generalizing well. The corresponding Mean Absolute Error (MAE) and Mean Squared Error (MSE) for both training and validation show improvements, confirming the model’s increasing effectiveness across epochs.

Table 2.

Test performance summary.

METRIC	VALUE
Test Loss (MSE)	0.4472
Test MAE	0.5585
Test MSE	0.4472
Root mean squared error (RMSE)	0.6688
Mean absolute error (MAE)	0.5585
Correlation between sentiment and rating	0.9416

Fig. 3.

Training loss over epochs.

Table 1.

Epoch summary for model performance.

Epoch	Training loss	Training MAE	Training MSE	Validation loss	Validation MAE	Validation MSE
1	5.118	1.733	5.118	0.3522	0.5041	0.3522
5	0.210	0.361	0.210	0.4117	0.5413	0.4117
10	0.053	0.181	0.053	0.3738	0.5195	0.3738
15	0.027	0.131	0.027	0.3578	0.5074	0.3578
30	0.014	0.096	0.014	0.3540	0.5074	0.3540
50	0.008	0.065	0.008	0.3915	0.5249	0.3915
75	0.009	0.060	0.009	0.3965	0.5276	0.3965
100	0.007	0.052	0.007	0.4181	0.5353	0.4181

Model loss & accuracy trend

Mean Absolute Error (MAE) and Mean Squared Error (MSE) across epochs as in Fig. 4 show that as epochs progress, the error metrics reduce, signalling improved predictive power. MAE decreases and stabilizes around 0.5585, while MSE reaches ~ 0.4472 by the last epoch. These values suggest the model’s accuracy in making precise recommendations with minimal deviations from user preferences.

Fig. 4.

MAE per epoch: training mean absolute error (MAE) and validation mean absolute error (MAE).

Convergence and generalization

The steady decline in both training and validation loss over multiple epochs shows smooth convergence, critical for effective performance in real-world applications. The similar trends in training and validation loss suggest the model generalizes well to new data, enhancing its use in practical recommendation scenarios.

Correlation with external factors

A correlation plot as in Fig. 5 shows a high positive correlation (0.9416) between user sentiment and ratings. This indicates that sentiment analysis is an essential predictor of user preferences, allowing the model to personalize recommendations based on real-time user sentiment effectively. Figure 6 shows the comparison of the evaluation metrics for test data.

Fig. 5.

Correlation between user sentiment and rating.

Fig. 6.

Test set evaluation metrics.

Sentiment and rating analysis

Sentiment polarity scores across travel reviews as in Fig. 7 show a bimodal distribution, with peaks around neutral and positive sentiments. This suggests that users tend to express neutral or positive views more frequently, which is valuable for refining positive and personalized recommendations. A histogram of ratings as in Fig. 8 reveals a right-skewed distribution, with most places receiving high ratings in the 4.5–5.0 range. This indicates a general tendency for positive reviews, which can guide the recommendation system toward highly-rated destinations.

Fig. 7.

Frequency of sentiment polarity categories.

Fig. 8.

Popularity distribution of places based on ratings.

Correlation and feature insights

The correlation heatmap as in Fig. 9 indicates a strong positive correlation between “Rating” and “Sentiment,” implying that places receiving high ratings tend to correlate with positive sentiment. Most other correlations are weak, indicating feature independence. This insight is valuable for feature selection and optimizing the model.

Fig. 9.

Correlation heatmap of numerical features.

The histogram of ratings as in Fig. 10 reveals a right-skewed distribution, with a majority of places receiving high ratings in the range of 4.5 to 5.0. This trend, illustrated in the rating distribution graph, indicates a strong preference for highly-rated places among users, which can guide the recommendation system to emphasize destinations with consistently positive feedback.

Fig. 10.

Rating distribution analysis.

Identification of top-rated places and users

Figures 11 and 12 explore the rating concentration patterns across destinations and users, respectively. These provide insight into item popularity and user rating behavior. Figure 11 indicates that some of the destinations receive consistent ratings near the upper bound, highlighting the strong alignment between user preferences and key travel attributes such as cultural significance, accessibility, and quality of experience. These highly-rated destinations serve as anchor items within the recommendation space, enabling the model to propagate preference signals more effectively, especially under sparse interaction conditions.

Fig. 11.

Ranking of places based on user ratings.

Fig. 12.

Ranking of users based on average ratings.

Figure 12 emphasizes a subset of users whose ratings across different destinations tend to be high, reflecting different user segments of optimistic or experience-driven evaluation behavior. Such users are important to identify because their interaction histories contribute more to preference generalization and similarity estimation in the NCF framework. These analyses jointly show how the model can tell the global popularity apart from user-specific preference patterns and enable the weighting of user influences adaptively to improve the reliability of recommendations for new or sparsely represented users.

Confusion matrix for rating predictions

The confusion matrix as in Fig. 13 evaluates the DNCF model’s prediction accuracy across rating categories. High accuracy is observed for ratings of 3 and 4, while the model struggles with ratings of 1, 2, and 5. This analysis highlights the model’s strengths and limitations in predicting different rating levels, guiding areas for further model refinement.

Fig. 13.

Confusion-matrix for rating classification.

User embedding analysis

The PCA-based 2D visualization as in Fig. 14 reveals potential clusters of users with similar preferences. This clustering insight can support targeted recommendation strategies by identifying distinct user groups with shared interests.

Fig. 14.

2D visualization of user embeddings.

Temporal trends in ratings for popular places

The line chart as in Fig. 15 shows fluctuations in average ratings for popular places (e.g., Unknown Place, Jama Masjid, and Birla Mandir) over time. This trend analysis helps identify seasonal or event-driven variations in user preferences, allowing the recommendation system to adjust suggestions based on temporal factors.

Fig. 15.

Dynamic rating analysis for popular places.

Additional evaluation insights

The cold start analysis graph as in Fig. 16 shows, how the model demonstrates a strong ability to recommend destinations to new users with limited interaction data. Despite the inherent challenges of cold start scenarios, the model performs effectively, showing its capability to deliver relevant recommendations even with minimal user history. The user engagement trends graph as in Fig. 17 illustrates the frequency of user interactions with the recommendation system over time, with noticeable peaks during certain periods, likely due to seasonal travel patterns or events. By analyzing these engagement trends, the system can optimize its responsiveness and enhance recommendation quality during high-demand periods. The feedback incorporation graph as in Fig. 18 tracks the model’s adaptation to user feedback, displaying steady improvements in recommendation accuracy as more feedback is incorporated. This adaptive capability enhances the model’s relevance, allowing it to continuously learn and improve its recommendations based on evolving user preferences. The performance comparison between Neural Collaborative Filtering (NCF) and Deep Neural Collaborative Filtering (DNCF), as depicted in Fig. 19, highlights that DNCF consistently achieves lower error metrics. This underscores DNCF’s effectiveness in handling complex recommendation scenarios, making it a superior choice for a personalized travel recommendation system.

Fig. 16.

Histogram of predicted ratings.

Fig. 17.

User-place interaction heatmap.

Fig. 18.

Place popularity bubble plot.

Fig. 19.

Residual error scatter plot.

Strengths and limitations

The proposed NCF-based personalized travel recommendation model has the following strengths and limitations that impact its effectiveness.

Strengths

High accuracy and strong sentiment correlation: The correlation coefficient of 0.9416 between sentiment and ratings demonstrates the model’s ability to adjust recommendations to align with users’ emotional responses, which is vital for personalized systems where user satisfaction impacts engagement and retention rates.

Effective convergence and generalization: The steady reduction in errors across training and validation reflects a balanced learning process, which minimizes overfitting. This stability ensures the model performs reliably on new data, a critical quality for production-level recommendation systems that handle diverse and changing data.

Scalability for large datasets: NCF’s combination of neural networks with collaborative filtering is effective for vast datasets with complex user-item interactions. This scalability is essential for recommendation engines operating on platforms with large user bases and a variety of content.

Limitations

Cold-start problem: Like other collaborative filtering models, NCF struggles with the cold-start issue, where recommending for new users or items with minimal interaction data is challenging. This can reduce effectiveness for new users or items, impacting user experience.

High data requirements: Deep learning models like NCF require substantial data to optimize and avoid overfitting. Performance may decline with sparse datasets, as capturing meaningful patterns becomes difficult.

Limited interpretability: Deep learning models are often “black boxes,” making it hard to explain specific recommendations, which may affect transparency and trustworthiness.

Explanation of findings

The model’s performance metrics highlight the potential and advantages of using NCF for personalized recommendations, especially in travel.

Effectiveness of NCF in capturing user preferences

Neural Collaborative Filtering combines collaborative filtering with deep learning, capturing intricate patterns in user behavior. This is reflected in the model’s low error metrics, as it accurately predicts user preferences by considering both explicit ratings and implicit sentiment feedback. This accuracy is vital for delivering recommendations that resonate with users’ travel preferences.

Model stability across epochs

The stability in training and validation loss trends after 60 epochs indicates a reliable balance in learning. The absence of large oscillations in loss means the model generalizes well, essential for applications that need consistent performance across varied interactions.

Implications for real-world applications

The high sentiment-rating correlation means the NCF model could adapt recommendations dynamically based on user feedback. For instance, if a user prefers beach locations, the system can prioritize such recommendations. Addressing cold-start challenges through a hybrid approach with content-based filtering could further enhance usability for new users with minimal interaction history.

Enhanced personalization for diverse user needs

The model’s architecture suits applications requiring high personalization. By capturing latent factors in user-item interactions, NCF can recommend destinations, hotels, and activities tailored to individual preferences, valuable in travel recommendations where targeted suggestions improve engagement and satisfaction.

Conclusions

This paper proposed a novel collaborative filtering scheme that relies on NCF as an underlying model to represent user preferences. Collaborative filtering is the most widely used approach in the area of intelligent recommendation systems. This method also caters for the cases when users have too few common rating items, which eventually influences the similarity assessment. This approach introduces proximity between global and local ratings whereby the tourist attraction interested to a user is used in calculating the preferences of the latter. Metrics in this category increase the similarity and reduce the data sparsity impacts of similarity metrics. The experimental results on tourism datasets validate the fact that an NCF-based recommendation method that combines user preferences provides superior recommendation performance.

Such promising results based on the suggested NCF model for tourist attraction recommendations may help improve the recommendation accuracy within the tourism industries. The current results provide a starting point for follow-up research tracks. Hence, scalability and robustness of the model should be evaluated further over larger, more heterogeneous datasets while considering various sectors of tourism and different demographics of users. It can also be very rich by integrating user data from social media, location, and even search behavior, which potentially allows for better accuracy in recommendations. Alternative machine learning approaches, such as reinforcement learning or hybrid models or advanced neural architectures, could potentially provide better performance gains on the recommendation system. Pre-processing is another very important phase in machine learning projects like recommendation systems. This step is aimed at cleaning, normalization, feature selection, and dimensionality reduction before analyzing raw data. Good quality, as well as new emergent patterns, can be achieved with high-quality pre-processed data. Further, linking the data of recommendation systems with other external sources-mostly on social media, travel reviews, or location services will result in increasing novelty and relevance of the system. By incorporating the additional data streams, this model helps understand user preferences with a better recommendation accuracy and system coverage. The NCF algorithm for travel recommendations presents several strengths and challenges.

Identified strengths

1. Incorporation of User Preferences: Factoring user preferences into the NCF algorithm enhances recommendation accuracy, leading to greater user satisfaction.

2. Mitigation of Sparse Data Challenges: This method effectively addresses sparse data, improving recommendation accuracy even with limited common rating items.

3. Improved Performance: Experimental results indicate that the NCF-based method surpasses traditional collaborative filtering algorithms, providing higher accuracy in sparse tourism datasets. 

Challenges faced

1. Increased Complexity: The complexity of the NCF algorithm may present implementation challenges and demand higher computational resources due to the need to process both global and local ratings.

2. Deep Data Pre-processing: Preprocessing is often deep, time-consuming, and only possible by a trained data analyst.

3. Limited Scope: The NCF framework may face challenges in suggesting lesser-known or recently established tourist destinations that have limited rating data.

This methodological framework based on NCF has promise in addressing the sparsity of data and inducting user preference effectively. Its complexity, however, may be a restriction on actual usability in certain scenarios. Subsequent studies should test the model based on NCF in real-life applications, such as online reservation systems for tourism and mobile applications for use with tourism services, to assess practical applicability within the tourism industry and user experience and involvement.

Author contributions

Aarif K and Sanjana P were responsible for the software development and implementation of the recommendation model. Ashwin Kumar handled the collection and preprocessing of the travel data. Ashwin Kumar and Aarif conducted the formal analysis of the experimental results and wrote the results section of the manuscript. Deepika J conceived the research direction, supervised the project, and contributed to the introduction and discussion sections. All authors reviewed and approved the final manuscript.

Funding

Open access funding provided by Vellore Institute of Technology.

Data availability

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Code availability

On request, access will be given to the GitHub repo.

Declarations

Competing interests

The authors declare no competing interests.