International Journal of Advances in Applied Sciences (IJAAS) Vol.
No.
September 2025, pp.
ISSN: 2252-8814.
DOI: 10.
11591/ijaas.
A method classifying the domestic tourist destination base similarity measuring Nguyen Thi Hoi.
Tran Thi Nhung.
Bui Quang Truong.
Nguyen Quang Trung Faculty of Economic Information System and E-commerce.
Thuongmai University.
Hanoi.
Vietnam
Article Info
ABSTRACT
Article history:
The classification problem is crucial in business, providing an effective method for supporting search activities in areas such as e-commerce, education, and marketing.
This has become especially important in the wake of the COVID-19 pandemic, which has increased the need to promote and stimulate domestic tourism.
This research focuses on recommending tourist destinations based on historical search data related to domestic tourism.
The study uses techniques like term frequency-inverse document frequency (TFIDF) weight vector analysis and similarity measures to calculate recommendation scores.
Data was collected from various tourism websites, covering destinations across all 63 provinces and cities in Vietnam.
Experiments were conducted using three approaches: cosine similarity, the brute force algorithm, and long short-term memory (LSTM) for long-text The results indicate that similarity-based methods produce recommendations that closely match user preferences.
For full-sentence queries, the brute force algorithm delivers more accurate results, while LSTM provides faster processing times.
These findings offer businesses multiple strategies for improving recommender systems in practical Received Oct 6, 2024 Revised May 10, 2025 Accepted Jun 8, 2025 Keywords:
Domestic tourist Historical search data Machine learning Recommendation system Similarity measures This is an open access article under the CC BY-SA license.
Corresponding Author:
Nguyen Thi Hoi Faculty of Economic Information System and E-commerce.
Thuongmai University No.
Ho Tung Mau Street.
Cau Giay District.
Hanoi.
Vietnam Email: hoint@tmu.
INTRODUCTION
With the rapid development of information and communications technology (ICT) and the widespread availability of the internet, users are increasingly relying on digital platforms, such as search engines and online entertainment services, to access a wide range of information and utilities.
This digital transformation has significantly reshaped consumer behavior, facilitating easier access to travel information, the ability to compare destinations, and the convenience of booking travel services with just a few clicks .
Ae.
Tourism remains a vital engine of global economic growth.
According to the World Travel & Tourism Council (WTTC) .
, the travel and tourism industry contributed 9.
1% to global gross domestic product (GDP) in 2023.
International tourism expenditures increased by 33.
1%, reaching USD 1.
63 trillion, while domestic expenditures rose by 18%, surpassing USD 5 trillion .
, .
Leisure-related spending saw a growth of 21.
2%, and business travel spending rose by 22.
4%, reflecting a strong post-pandemic recovery .
, .
In terms of direct economic impact, the travel and tourism sector added USD 3,059.
3 billion to global GDP .
8%), primarily driven by industries such as hospitality, travel agencies, and passenger transportation services .
, .
From 2024 to 2034, the sectorAos contribution is projected to grow at an annual rate of 3.
7%, reaching USD 4,865.
7 billion .
5% of GDP) .
Ae.
Journal homepage: http://ijaas.
Int J Adv Appl Sci
ISSN: 2252-8814
In Vietnam, the tourism industry has demonstrated impressive post-pandemic recovery, especially in the domestic sector.
According to the General Statistics Office of Vietnam .
, international tourist arrivals 6 million in 2023, approximately 3.
5 times higher than in 2022, recovering to 70% of prepandemic levels .
, .
This surpassed AsiaAos average recovery rate of 65%.
Domestic tourism was even more robust, recording 110 million domestic trips in 2023 .
Infrastructure improvements, enhanced service quality, and strategic investment in cultural preservation have contributed to this growth .
Ae.
This growth has generated significant economic spillover effects across various industries, including hospitality, food and beverage services, transportation, and retail.
In addition, cultural festivals and sightseeing activities contribute to the development of the creative and entertainment industries while supporting local businesses and traditional craft villages .
Ae.
Tourists now benefit from real-time information access, easy price comparisons, and instant booking features.
However, the volume of available information has become overwhelming, often hindering travelers from efficiently identifying destinations that align with their preferences .
, .
, .
Ae.
To address this, tourism recommendation systems (TRS.
have become increasingly important.
These systems apply machine learning techniques to analyze user behavior, historical data, and preferences to offer personalized travel suggestions.
By utilizing both static and dynamic .
eal-time reviews, social media, and rating.
TRSs can deliver context-aware recommendations that improve user satisfaction .
Ae.
, .
, .
Ae.
, .
, .
Such systems are already in use across various industries including e-commerce (Amazo.
, streaming (Netflix and YouTub.
, recruitment, and tourism .
Ae.
Despite the economic potential of tourism, participation among local communities in tourism-related activities remains limited .
, .
Ae.
, .
Most Vietnamese travel websites provide basic lists of attractions, but lack personalized recommendations tailored to individual behaviors or previous interactions.
This leads to decision fatigue and inefficient travel planning .
, .
Ae.
, .
Ae.
To overcome these limitations, this study proposes the development of a personalized TRS tailored to the domestic travel market in Vietnam.
By leveraging users' browsing behavior and search patterns, the system will suggest lesser-known destinations that align with individual preferences.
This not only enhances the user experience but also promotes equitable tourism development across regions.
The proposed system integrates advanced computational techniques, including cosine similarity, brute force matching, and long short-term memory (LSTM) models .
, .
, .
, .
, .
, .
Ae.
These techniques are implemented on a prototype website to assess recommendation accuracy and user relevance.
By integrating these algorithms, the system aims to offer more accurate, real-time, and interestbased travel suggestions, while contributing to the visibility and economic uplift of underexplored destinations across Vietnam.
The article includes the following sections: problem statement, related works, research model, research results, and some implications for localities with many tourist destinations in Vietnam.
RELATED WORK
Recommender systems for domestic travel destinations have emerged as a vital tool in the digital transformation of the tourism industry, offering personalized travel suggestions based on individual user These systems aim to generate tailored lists of tourist sites that align with user expectations, thereby improving decision-making and user satisfaction.
As travelers increasingly demand convenient and customized planning solutions, particularly in the post-pandemic context, tourism recommender systems have become widely adopted not only in Vietnam but globally .
Ae.
, .
, .
, .
, .
, .
, .
, .
Ae.
Current research in travel recommendation systems is typically structured around two main methodological approaches.
The first approach involves clustering tourist destinations based on shared attributes such as natural landscapes .
each and mountain.
, cultural and historical value, or urban experiences .
, .
, .
, .
, .
, .
, .
, .
, .
Ae.
These categorizations are achieved using unsupervised learning techniques, allowing the system to classify destinations into meaningful groups that reflect user preferences.
For example, .
, .
, .
, .
, .
demonstrated the application of K-means or classification-based association (CBA)-fuzzy algorithms for destination clustering, yielding promising results in automated recommendations.
Similarly, .
, .
, .
applied support vector machines (SVM) and Bayesian algorithms to analyze large-scale datasets derived from user behavior and survey feedback, enabling improved segmentation of user travel intentions.
Significantly, .
introduced an innovative deep learning framework based on EfficientNetlite, optimized for mobile deployment, achieving over 80% user satisfaction with more than 50% of recommendations meeting expectations.
This model underscores the potential of deep learning in enhancing recommendation systems, particularly in real-time and user-centered mobile applications.
The second approach focuses on supervised and semi-supervised learning methods for generating recommendations .
Ae.
, .
, .
Ae.
, .
, .
Ae.
, .
, .
These models are trained on extensive labeled datasets to optimize the relevance and accuracy of destination suggestions.
For instance, .
evaluated six machine A method classifying the domestic tourist destination base similarity measuring (Nguyen Thi Ho.
A ISSN: 2252-8814 learning algorithms nayve Bayes (NB).
SVM, logistic regression, neural networks, decision trees, and random forests on tourism data, with SVM, logistic regression, and neural networks demonstrating superior performance in predicting user preferences.
Furthermore, .
compared Doc2Vec embeddings with traditional techniques such as item-k-nearest neighbors (KNN) and neural collaborative filtering, finding that Doc2Vec significantly outperformed conventional models, particularly in handling sparse datasets.
Building on these foundations, our study proposes a personalized travel recommender system designed specifically for the Vietnamese domestic tourism context.
This system analyzes user input, such as search queries and browsing patterns, and compares them with structured destination profiles to generate accurate and relevant recommendations.
To facilitate this, we utilize N-gram analysis and a Wikipedia-based dictionary to extract semantic features from user input and destination metadata.
To further enhance recommendation accuracy, we apply cosine similarity and Pearson correlation to measure alignment between user interests and destination characteristics.
Expanding the dataset is another key objective to improve system robustness and accuracy.
including more diverse user profiles and regional destination data, the systemAos adaptability can be enhanced.
Beyond tourism, the proposed architecture holds potential for cross-domain applications in education, entertainment, and healthcare, where personalized information delivery is equally valuable.
PROPOSED RESEARCH MODEL
Proposed model The problem of suggesting domestic tourist destinations can be briefly described as based on the user's request, the system provides suggestions as a list of tourist destinations that are considered closest to the user's requests.
From this problem, the research proposed a model as follows:
Oe Input: question or query with a list of keywords that describe the user's wishes.
Oe Output: a list of tourist destinations close to the user's wishes.
Oe Perform: .
step 1: analyze the question or query.
step 2: compare based on similarity between the question and database by weight or based on training.
step 3: perform evaluation and sorting based on similarity results.
and i.
step 4: provide a list of locations.
Similarity measures and correlation To operationalize the proposed research model, a comprehensive, multi-stage methodology was This approach integrates descriptive data analysis with advanced text normalization techniques, tailored for both structured and unstructured data within tourism systems.
The primary objective was to construct a robust recommendation architecture capable of accurately interpreting user-generated queries and effectively matching them to relevant destination metadata.
The initial phase involved rigorous preprocessing of textual data collected from various sources, including destination descriptions, user reviews, and behavioral logs.
The preprocessing steps encompassed tokenization, lowercasing, stop-word removal, and lemmatization, aiming to normalize linguistic features across datasets.
Subsequently, the cleaned textual data was transformed into numerical form using SVM, primarily through term frequency-inverse document frequency (TF-IDF) weighting schemes.
TF-IDF has been shown to be effective in reducing noise and emphasizing semantically significant terms in sparse textual datasets .
, .
, .
, .
, .
Ae.
Calculate weight is calculated based on keyword TF, the IDF, and the TFyIDF is calculated as .
ycycnyc = { ycNya O yayaya = .
ceycnyc )) O ycoycuyci ( ycA 1 yccyceycn ) ycnyce ycyceycnyc Ou 1 0 ycnyce ycyceycnyc = 0 Where ycycnyc is the weight of ycycn in text yccyc , yceycnyc is the frequency of appearance of ycycnyc in text yccyc .
This transformation enables the application of various computational techniques such as cosine similarity.
SVM, and K-means clustering, which operate effectively in high-dimensional feature spaces .
Ae.
, .
, .
, .
Ae.
, .
To identify semantic alignment between user queries and destination profiles, the cosine similarity metric was employed.
This measure evaluates the cosine of the angle between two non-zero vectors in a multi-dimensional space, making it highly effective in text mining and recommendation systems.
Such cross-domain adaptability is made possible by the systemAos reliance on generalized deep learning and information retrieval techniques, rather than domain-specific rules or hardcoded logic.
Int J Adv Appl Sci.
Vol.
No.
September 2025: 740-750
Int J Adv Appl Sci
ISSN: 2252-8814
EE) = ya y yaA ycaycuycycnycuyce .
aE, yaA .
yaA| A .
e y yaA EE is the dot product of two vectors.
|A|, |B| is the magnitude of the two vectors.
In which: ya Cosine similarity is particularly advantageous for high-dimensional TF-IDF vectors, where traditional distance metrics like Euclidean distance are less reliable due to the curse of dimensionality .
, .
, .
, .
, .
A brute force algorithm was also implemented to serve as a baseline method.
This approach exhaustively compares each user profile vector with all available destination vectors.
While computationally intensive, brute force ensures maximum matching precision, making it a useful benchmark for evaluating more scalable algorithms .
, .
, .
, .
In smaller datasets or controlled experimental scenarios, brute force methods are highly interpretable and often outperform more complex models in terms of raw accuracy.
To capture the temporal and contextual dynamics of user behavior, an LSTM network, a variant of recurrent neural networks (RNN.
, was integrated into the recommendation engine.
LSTM networks are particularly well-suited for sequence modeling due to their ability to maintain long-term dependencies via gated mechanisms .
, .
, .
Ae.
, .
The LSTM model used in this study comprises three key gates: .
forget gate: filters irrelevant information from the previous cell state.
input gate: integrates new input into the current state.
output gate: determines which information contributes to the output.
This architecture enables the system to distinguish between short-term and long-term user preferences .
, temporary interest in beach resorts vs.
consistent preference for historical landmark.
, .
, .
, .
, .
, .
, .
, .
Ae.
To overcome the sparsity and high dimensionality problems inherent in TF-IDF matrices, enhancements to the cosine similarity measure were adopted.
These included adaptive term weighting, soft cosine similarity, and local similarity thresholds, which together increased both recall and precision in sparse datasets .
, .
, .
, .
, .
Ae.
, .
Such enhancements allowed the system to better handle semantically similar but lexically different queries, especially in multilingual or domain-specific contexts.
The integration of vector-based similarity, brute force baselines, and sequence-aware LSTM models results in a hybrid methodology that is not only robust but also scalable across domains.
Although initially implemented for Vietnamese domestic tourism, the underlying architecture is transferable to adjacent sectors such as education, entertainment, and healthcare .
, .
, .
, .
, .
, .
EXPERIMENT RESULTS AND DISCUSSION
Data collection In this study, a comprehensive dataset of tourist destinations across all 63 provinces and centrally governed cities in Vietnam was constructed to support the training, evaluation, and validation of the proposed TRS.
The data acquisition process was conducted through a semi-automated pipeline leveraging publicly available information retrieved from multiple prominent travel service platforms, including Traveloka.
Vinpearl.
DiscoveryTravel.
VietTravel.
Vntrip.
vn, and VietnamBooking.
These platforms are widely used in tourism data analytics due to their rich, diverse, and regularly updated repositories of destination metadata .
Ae.
, .
, .
, .
An initial corpus of 8,940 raw descriptive records was collected from these platforms, encompassing various types of tourist locations, metadata categories, and user-generated textual content.
To ensure data quality and eliminate redundancies, a comprehensive preprocessing and cleaning phase was applied.
This process included: .
removal of duplicate entries.
normalization of naming conventions across sources.
filtering out entries with missing, ambiguous, or non-informative content.
Following this refinement, a curated dataset of 2,100 high-quality and unique destination profiles was obtained, deemed suitable for experimental modeling and machine learning-based training purposes.
The finalized dataset captures critical attributes for recommendation modeling, such as: .
location name and i.
geographic region.
, nature, heritage, entertainmen.
descriptive summary.
user-generated tags and text reviews.
These features were designed to support both text vectorization and semantic similarity computation in downstream recommendation algorithms.
The statistical distribution and structural summary of the dataset are presented in Table 1 and serve as the basis for clustering, classification, and deep learning models applied in subsequent phases .
, .
, .
, .
Ae.
, .
, .
, .
, .
Ae.
, .
Ae.
, .
, .
To further enhance the datasetAos textual quality, a rigorous preprocessing pipeline was implemented.
This phase aligned with state-of-the-art methodologies proposed in .
, .
, .
, .
Ae.
, .
, .
, .
, .
Ae.
, .
Ae.
, .
, .
, recognizing that real-world travel descriptions often exhibit: .
lexical inconsistencies .
, spelling error.
redundancy and repetition.
non-standardized terminology and abbreviations.
To mitigate these issues, the study adopted a Wikipedia-based lexical dictionary as a semantic reference source.
This dictionary was employed to normalize key terms, identify alternative A method classifying the domestic tourist destination base similarity measuring (Nguyen Thi Ho.
A ISSN: 2252-8814 expressions, and enrich sparse content.
The use of Wikipedia for semantic expansion is consistent with prior research that highlights its capacity to extract domain-agnostic, semantically rich representations from noisy text data .
, .
, .
, .
In parallel, a comprehensive Vietnamese stop-word list was compiled from both WikipediaAos linguistic tools and the Vietnamese Language Research Project, as discussed in .
, .
, .
Ae.
, .
These resources were used to identify and eliminate non-informative terms .
, "", "rut", "mot ni") that commonly appear in user reviews but offer limited semantic contribution to destination profiling.
Once cleaned, the textual data was subjected to N-gram analysis, enabling the extraction of co-occurrence patterns that reflect contextual and syntactic relationships.
Experiments were conducted using N values ranging from 1 to 4, with bigrams (N=.
yielding the most coherent results for Vietnamese sentence structures, where compound concepts often span two or more syllables .
Trigrams (N=.
also provided promising performance, especially for complex location descriptors.
Consequently, the final feature extraction focused on bigrams and trigrams, which were then transformed into TF-IDF vectors.
This transformation assigns each term a weight that reflects its importance within a specific document relative to the entire corpus, ensuring that both common and distinctive terms are appropriately captured.
Table 1.
Sample list of websites Name Traveloka .
ttps://w.
com/vi-v.
Vinpearl .
ttps://vinpearl.
com/v.
Du lUch khym phy .
ttps://dulichkhampha.
Du lUch ViNt .
ttps://dulichviet.
Vntrip.
ttps://w.
vn/e.
VietNam Booking .
ttps://w.
Total (Source: authors' statistics from resources in Vietna.
Volume 1,890 1,760 1,575 1,260 1,510 8,940 Sample 2,100 Experimental data sample To implement the algorithms and measurements, the study built a sample data set consisting of: no.
name of destination, city, description, and multimedia.
After processing the data, the researchers selected a set of 2,100 locations.
They include beach tourism, mountain tourism, ecotourism, historical relics, churches, pagodas, and entertainment areas.
Table 2 shows the structure of a sample of data about a tourist destination.
Table 2.
A sample of data for the experiment No.
Name Khu du lUch Tam cBych aong City Ninh Bynh Description Khu du lUch Tam cAeBych aong cc vy nh "vUnh H Long cn" vui nhiAu cnh cp nh Tam c, cAn Thyi Vi, chya Bych aong, cong Tiyn, hang But, thung Nng, thung Nham, vyn chim.
Nii ting vui danh xng AoNam thiyn cN nhU congAo.
Tam c Bych aong s hu cnh sc lyng quy yyn bynh cyng hN thcng hang cong nyi cy vyi un tng.
Ly mot phn trong Qun thE danh thng Tryng An.
Tam c Bych aong ly ciEm cn hoyn ho dynh cho nhng ai mucn khym phy trsn vn v cp non syng.
aUa chO: Ninh Hi.
Hoa L.
Ninh Bynh .
eng bng syng Heng, miAn Bc ViNt Na.
Phyn loi du lUch: du lUch sinh thyi, du lUch vEn hya, du lUch nghO dng vy du lUch khym phy (Source: from data sample set built by author.
Multimedia https://staticimages.
net/files/pu blish/2023/7/8/ninh-binhanh-so-du-lich-cc-1021864.
Experimental scenario Research carried out by scenario as follows: .
step 1: build a set of user queries.
step 2: build a word set for each query, preprocessing.
N-gram separation, stop-word removal.
step 3: calculate the weight vector by TF.
IDF.
Calculate the correlation of each query with the locations.
step 4: conduct manual matching from volunteers.
step 5: conduct search using models and evaluate the search performance of the model.
The detailed steps of the scenario are as follows: to rigorously evaluate the performance of the proposed personalized TRS, the research team developed a structured dataset of 100 user queries, each simulating a real-world search intent in the tourism domain.
These queries varied in word length, lexical Int J Adv Appl Sci.
Vol.
No.
September 2025: 740-750
Int J Adv Appl Sci
ISSN: 2252-8814
complexity, and thematic focus, covering interests such as ecotourism, cultural heritage, coastal destinations, and urban exploration.
Each query was stored in an Excel spreadsheet with two primary columns: .
serial number .
uery ID) and i.
query text.
This structure ensured traceability and reproducibility throughout the experimental workflow.
An additional column was later added to log the expected number of matching destinations, determined through human annotation.
The query evaluation process followed a multi-stage architecture, consisting of the following steps:
Stage 1: query design and ground truth annotation: each query was constructed to reflect natural language expressions of travel interest.
A group of human evaluators was recruited to manually associate each query with a relevant set of destinations from the curated dataset of 2,100 tourist locations, previously collected across all 63 provinces and municipalities in Vietnam.
These associations served as the gold standard reference set for evaluating algorithmic outputs .
, .
Ae.
, .
, .
, .
Stage 2: text preprocessing and normalization: each query was tokenized and normalized using standard natural language processing (NLP) techniques: tokenization and stop-word removal based on linguistic resources from Wikipedia and the Vietnamese Language Research Project .
, .
, .
, .
N-gram decomposition with values of N=1 to 4, to capture the frequent occurrence of compound and multisyllabic expressions in Vietnamese .
, .
, .
, .
Lowercasing and morphological normalization to ensure uniform representation across datasets.
Stage 3: query vectorization via TF-IDF: following preprocessing, both the user queries and the textual metadata of each tourist destination were transformed into high-dimensional vector representations using the TF-IDF model.
This weighting scheme effectively quantifies term significance within the corpus, enabling accurate semantic comparison .
, .
, .
, .
Stage 4: algorithmic evaluation and similarity computation: three algorithms were employed to assess semantic similarity between queries and destination descriptions: cosine similarity: computes the cosine of the angle between TF-IDF vectors, facilitating a scalable and interpretable matching mechanism .
, .
, .
, .
, .
brute force matching: performs exhaustive pairwise comparisons between query vectors and all destination vectors .
, .
, .
, .
Ae.
Although computationally intensive, it guarantees maximum recall and serves as a baseline evaluation method .
, .
, .
Ae.
, .
LSTM neural network: utilizes a gated RNN architecture to capture temporal and contextual dependencies in text sequences.
This method is particularly well-suited for interpreting nuanced patterns in user intent .
Ae.
, .
, .
, .
, .
, .
Experiment setup: all implementations were carried out in Python, using Visual Studio Code .
as the integrated development Experiments were executed on a Windows 11 machine powered by an Intel(R) Core i7-1185G7 @ 3.
00 GHz CPU and 32 GB RAM, ensuring stable and reproducible computations.
To establish an empirical performance benchmark, the research team conducted a manual evaluation process involving a group of trained volunteers.
Each participant was instructed to select a set of relevant destinations from the master dataset for each of the 100 queries.
These manually selected results were consolidated to form the ground-truth relevance set for each query.
For each algorithm, the system-generated recommendations were compared to the ground truth using a binary scoring mechanism: let N denote the number of correct locations .
rom human annotatio.
, let M represent the number of locations retrieved by the algorithm.
for each match between the retrieved and reference sets, the accuracy score was incremented by one unit.
This per-query matching accuracy was then aggregated across all 100 queries to assess overall system effectiveness .
, .
, .
In addition to binary accuracy scoring, the system's predictive correctness was evaluated using the mean squared error (MSE) MSE measures the average squared deviation between the number of expected and retrieved results per query, offering a quantitative indicator of over- or under-prediction .
, .
, .
, .
, .
, .
, .
ycAycIya = Ocycuycn=1.
cyycn Oe ycycn )2 .
ycu With each ycyceycycyceycycycn , ycn = 1 .
100, ycoycuycaycaycycnycuycuyc , yc = 1 .
2100, total as .
= 210000 pairs, and n=210000.
The MSE formula states that accuracy increases as the value of MSE gets closer to zero, and vice versa.
Acquired experimental findings as in .
ycu ycAycIya = Ocycuycn=1.
cyycn Oe ycycn )2 = Oc210000 .
cyycn Oe ycycn )2 = 0,1251 ycn=1 .
Corresponding to the precision of the experiment with cosine is .
yaycI = .
Ae ycAycIy.
y 100% = .
Oe 0,1.
y 100% = 87,49% .
A method classifying the domestic tourist destination base similarity measuring (Nguyen Thi Ho.
A ISSN: 2252-8814 This evaluation framework ensured that both the precision of matching and volume of relevant results were systematically assessed across all models and queries.
Discussion and limitation The following is the query search results table of the models: cosine similarity, brute force algorithm, and LSTM model in Step 3.
Accuracy in modes is illustrated in Table 3.
Following the completion of experimental trials, the study identified that the cosine similarity algorithm outperformed other models in terms of recommendation accuracy, achieving an accuracy rate of 87.
49%, followed by the brute force matching method at 84.
55%, and the LSTM model at 82.
As detailed in Table 3, cosine similarity not only yielded the highest average accuracy but also demonstrated the lowest standard deviation, indicating strong stability and consistency across various query types.
Conversely, the brute force model exhibited the highest variability, suggesting sensitivity to input query structure and vector sparsity .
Ae.
, .
, .
, .
, .
, .
Ae.
Table 3.
The results of the three modes No.
Brute force (%) Cosine similarity (%) LSTM (%) The observed superiority of cosine similarity is primarily attributed to its computational efficiency in measuring angular distance between TF-IDF vectors derived from user queries and destination profiles .
, .
, .
This approach facilitates the extraction of semantically aligned results without relying on historical user interaction data.
The transformation of unstructured text into high-dimensional structured feature vectors enabled the model to match user intent with destination metadata in a purely content-based framework .
, .
Ae.
, .
In operational terms, the most relevant destinations, those with the highest cosine similarity scores, were automatically ranked at the top, providing users with suggestions that closely mirrored their intent.
The brute force matching model, though computationally expensive, was shown to be particularly effective in handling full-sentence queries where semantic nuance and context were essential.
This model conducted exhaustive pairwise comparisons between every query vector and the destination corpus, thus maximizing coverage .
, .
, .
, .
Meanwhile, the LSTM model, despite being slightly less accurate in top-1 recommendation ranking, demonstrated notable strengths in processing speed and scalability, especially when deployed in environments requiring real-time responses or handling large query volumes .
, .
, .
, .
, .
, .
, .
To evaluate the practical applicability of the proposed models, a web-based TRS was developed using the FastAPI framework, deployed in a Python environment, and executed via Visual Studio Code.
The system interface enables users to input free-text queries and receive an interactive output interface that presents: .
the number of matched results.
a ranked list of destination suggestions per algorithm.
real-time visual comparison of outputs from cosine similarity, brute force, and LSTM methods.
This feature allowed researchers and test users to qualitatively assess algorithm outputs and validate model behavior under different linguistic scenarios.
Several key insights derived from system experimentation are that the brute force matching achieved marginally higher accuracy for long-form queries, aligning well with user-selected destination sets in manual validation.
LSTM, while slightly trailing in accuracy, offered faster runtime and better resource optimization, suggesting potential for real-time or mobile deployment.
Cosine similarity, especially when fine-tuned with optimized TF-IDF parameters and bigram modeling, consistently delivered semantically rich and accurate results across all query types.
These findings reaffirm the strength of textual similarity-based models, particularly when combined with robust preprocessing and vector space transformation techniques, in capturing user preferences with high fidelity.
When benchmarked against leading recommender models from recent literature, the proposed hybrid framework demonstrated several notable advantages: in contrast to .
, which primarily relies on user reviews and historical behavior, the proposed system mitigates popularity bias by prioritizing semantic relevance between query content and destination profiles.
Compared to .
, which emphasizes image-based and activity recommendations, our model offers stronger textual processing capabilities but could be extended with image metadata and caption embedding for multimodal integration.
Unlike .
, which activates recommendations only after the user selects a destination, our system enables pre-decision exploration, supporting discovery of non-mainstream or under-visited locations.
While .
, .
, and .
Int J Adv Appl Sci.
Vol.
No.
September 2025: 740-750
Int J Adv Appl Sci
ISSN: 2252-8814
present a hybrid recommendation framework, the accuracy metrics achieved through our TF-IDF cosine similarity LSTM ensemble demonstrated superior performance and lower error rates.
Based on empirical findings, several strategic enhancements are recommended to elevate the systemAos accuracy and applicability:
Dataset expansion: increasing the number of curated destination entries will improve model generalization, reduce vector sparsity, and enhance recommendation diversity .
, .
, .
, .
, .
Multimodal data integration: enriching the dataset with user reviews.
GPS traces, and visual data .
mages, video.
would improve context awareness, enabling emotionally intelligent and situationally relevant suggestions .
, .
, .
, .
Cross-domain transferability: the current frameworkAos modular design enables easy transfer to domains such as education .
ourse recommendatio.
, entertainment .
ovie or event suggestion.
, and healthcare .
ellness service.
, where personalized content delivery can drive user engagement .
, .
, .
, .
By incorporating these enhancements, the proposed recommendation engine is expected to evolve into a scalable, domain-independent, and highly adaptive system capable of delivering personalized and contextually meaningful recommendations in both tourism and broader knowledge-based application areas.
CONCLUSION
In conclusion, this study presents a practical and effective solution for aligning user preferences, expressed in natural language, with structured destination metadata.
By applying content-based filtering grounded in semantic similarity, the system delivers personalized domestic tourism recommendations.
Core techniques such as N-gram decomposition.
TF-IDF vectorization, and the use of a domain-specific dictionary allow for deeper interpretation of user intent.
The integration of similarity metrics like Cosine Similarity and Pearson Correlation ensures accurate query matching.
Experimental results demonstrate the systemAos effectiveness in capturing user expectations through relevant suggestions.
The findings highlight the importance of dataset scale and semantic diversity in optimizing recommendation performance.
This approach also proves to be flexible and scalable, with the potential for extension to other application Additionally, the framework supports multilingual capabilities with minimal adaptation.
Ultimately, this research contributes a robust foundation for intelligent, user-centric recommendation systems in the field of digital tourism.
ACKNOWLEDGMENTS
We would like to express our sincere appreciation to the student group from Thuongmai University for their valuable support in data collection and preprocessing.
FUNDING INFORMATION
This research is funded by Thuongmai University.
Hanoi.
Vietnam.
AUTHOR CONTRIBUTIONS STATEMENT
This journal uses the Contributor Roles Taxonomy (CRediT) to recognize individual author contributions, reduce authorship disputes, and facilitate collaboration.
Name of Author Nguyen Thi Hoi Tran Thi Nhung Bui Quang Truong Nguyen Quang Trung C : Conceptualization M : Methodology So : Software Va : Validation Fo : Formal analysis ue I : Investigation R : Resources D : Data Curation O : Writing - Original Draft E : Writing - Review & Editing ue Vi : Visualization Su : Supervision P : Project administration Fu : Funding acquisition A method classifying the domestic tourist destination base similarity measuring (Nguyen Thi Ho.
A ISSN: 2252-8814
CONFLICT OF INTEREST STATEMENT
The authors declare that there are no conflicts of interest related to this study.
INFORMED CONSENT
We have obtained informed consent from all individuals included in this study.
ETHICAL APPROVAL
The research related to human use has been compiled with all the relevant national regulations and institutional policies in accordance with the tenets of the Helsinki Declaration and has been approved by the authors' institutional review board or equivalent committee.
DATA AVAILABILITY
All data utilized for the testing phase were systematically collected from reputable travel websites featuring domestic destinations across Vietnam.
A comprehensive list of these data sources is presented in Table 1 of this paper.
REFERENCES