The expansion of cities has created a demand for advancedsystemsthatcanaidresidentsandtouristsinnavigating urban areas. This paper introduces a recommendation system that aims to improve the city exploration experience by utilizinga review-based approach. By analyzing user-generated reviews, sentiment analysis, and location/place data, the system suggests points of interest, enhancing decision-making for users [3] [4]. Thesystemusesmachinelearningmethods,suchascollaborative filteringandnaturallanguageprocessing,toanalyzeinformation. By considering both location-specific details and the sentiment conveyed in reviews, the approach guarantees that recommenda- tionsaresuitedtothespecificneedsandpreferencesoftheuser.
Introduction
Overview
In the age of digital travel, existing platforms like TripAdvisor and Google Maps often provide generic recommendations. This project develops a smart, personalized place recommendation system that goes beyond popularity by incorporating user-specific filters (e.g., wheelchair access, pet-friendliness) and sentiment analysis of real user reviews to deliver tailored city exploration experiences.
Key Features
Highly Personalized Filters: 20–30 filters including accessibility, facilities, and user interests.
Sentiment-Based Ranking: Uses natural language processing to assess user sentiment and prioritize places accordingly.
Interactive Map Interface: Leaflet.js integration for visualizing recommendations.
Impact
Helps tourists maximize limited time in a city and locals discover new places.
Provides a more intuitive, context-aware alternative to static maps or generic travel apps.
Incorporates collaborative, content-based, and hybrid recommendation techniques for precision and adaptability.
Methodology
Data Collection
Sourced via Google Maps API and web scraping, storing place names, coordinates, user reviews, ratings, and accessibility metadata.
Data Preprocessing
Cleaning: Removed noisy, incomplete, or duplicate data.
Filtering: Only included places with ≥100 reviews to ensure sentiment reliability.
Text Preprocessing: Tokenization, stopword removal, and Word2Vec embedding of review text and place subtypes.
Sentiment Analysis & Fake Review Detection
Used TextBlob to score reviews as positive, neutral, or negative.
Heuristics filtered out fake or redundant reviews based on exaggerated sentiment, duplication, and length extremes.
Machine Learning Model
Random Forest Classifier trained on Word2Vec vectors to classify place types.
Achieved expected accuracy of 80–90% based on data diversity and embedding quality.
Feature Extraction
Parsed nested JSON fields (e.g., about) to extract boolean filters like has_parking or wheelchair_accessible.
Implementation
Tech Stack
Backend: Python, Flask, Pandas, MySQL
ML & NLP: Scikit-learn, Gensim (Word2Vec), TextBlob, NLTK
Frontend: HTML/CSS/JS, Bootstrap, Leaflet.js
Authentication: Google OAuth
Steps
Load and clean Excel and API-collected datasets.
Preprocess reviews and generate vector embeddings.
Train and evaluate the classification model.
Detect and remove fake reviews to ensure recommendation integrity.
Deploy Flask web app with user login, map-based place visualization, and recommendation display.
Results & Discussion
Accuracy: Data cleaning and sentiment analysis significantly enhanced the reliability of user feedback and place classification.
Effectiveness:
Combining live API data with structured files produced a rich and updated dataset.
Sentiment analysis made recommendations more emotionally aligned with user preferences.
Conclusion
In this implementation, we developed a comprehensive travel recommendation system that leverages advanced data preprocessing,machinelearningalgorithms,anduserfeedback analysis to provide personalized and relevant recommenda- tions. The system architecture includes data collection from reliable sources, accurate data preprocessing, an effective recommendationengine,andauser-friendlywebapplication.
A. Summary of Implementation
1) DataCollectionandPreprocessing::
• Data Sources: Data was collected from the Google Places API and pre-collected Excel files, ensuring a comprehensive and up-to-date dataset.
• Preprocessing Steps: The dataset underwent rigorous preprocessing, including handling null values, removing unnecessarycolumns,sentimentanalysisusingTextBlob, and fake review detection.
• Text Vectorization: Text data was converted into nu- merical vectors using Word2Vec, making it suitable for machine learning algorithms.
2) RecommendationEngine::
• Machine Learning Model: A Random Forest classifier was trained to classify places based on their type, ensur- ing accurate classification.
• Collaborative Filtering: Collaborative filtering tech- niques were employed to provide personalized recom- mendations based on user preferences and past interac- tions.
• User Feedback Analysis: User feedback was contin- uouslyintegratedintotherecommendationengineto improve the accuracy and relevance of future recommen- dations.
3) WebApplication::
• User Authentication: User signup, sign-in, and Google OAuthfunctionalitieswereimplementedtoensuresecure user data storage and access.
• Map Integration: Leaflet was integrated for interactive map visualization of recommended places.
• User-Friendly Interface: The Flask web application provided a responsive and user-friendly interface for interaction and visualization, enhancing the overall user experience.
B. Improvements Over Existing Systems
1) Comprehensive Data Collection:: The system collects data from multiple reliable sources, ensuring a rich and up-to-date dataset. Existing systems often rely on a single data source,whichmaylimitthecomprehensivenessofthedataset.
2) AccurateDataPreprocessing::Thepreprocessingsteps, includingsentimentanalysisandfakereviewdetection,ensure thatthedatasetiscleanandreliable.Existingsystemsmaynot employsuchrigorouspreprocessingsteps,leadingtopotential biases or inaccuracies in the dataset.
3) Effective Recommendation Engine:: The recommenda- tionengineusesacombinationofmachinelearningalgorithms and collaborative filtering techniques to provide personalized and relevant recommendations. Existing systems may rely solely on collaborative filtering, which may not capture the nuances of user preferences as effectively [7].
4) Continuous User Feedback Analysis::The system con- tinuously improves its recommendations based on user feed- back, ensuring that the recommendations remain relevant and accurate over time. Existing systems may not integrate user feedback as effectively, leading to potential stagnation in recommendation quality.
C. Future Work and Enhancements
1) AdvancedMachineLearningTechniques:Futurework couldexploreadvancedmachinelearningtechniquessuch as deep learning and reinforcement learning to enhance the accuracy and relevance of recommendations [10].
2) Real-Time Data Updates: Integrating real-time data updates from the Google Places API to ensure that the recommendationsarealwaysup-to-date.Dynamicdataset and recommendation updates could be implemented.
3) Enhanced User Feedback Analysis: Implement more advanced techniques for user feedback analysis, such as NLP,tobetterunderstanduserpreferencesandfeedback[11][12]
4) Scalability and Performance: Optimize the system for scalability and performance to handle a large number of users and recommendations efficiently. Implement load balancing and caching mechanisms to improve respon- siveness and efficiency.
5) User Interface Enhancements: Enhance the user in- terfacetoprovideamoreintuitiveandengaginguser experience. Implement additional visualization tools and interactive features to help users explore recommenda- tions more effectively.
References
[1] Pu,Zihao&Du,Hongyu&Yu,Sizhe&Feng,Duanyu.(2020). Improved Tourism Recommendation System. 121-126.10.1145/3383972.3384074.
[2] Kawai, Yukiko & Zhang, Jianwei & Kawasaki, Hiroshi. (2009). Tourrecommendation system based on web information and GIS. 990 - 993.10.1109/ICME.2009.5202663.
[3] Osman, Nurul. (2020). Contextual Sentiment Based RecommenderSystem to Provide Recommendation in the Electronic Products Do-main. International Journal of Machine Learning and Computing. 9.10.18178/ijmlc.2019.9.4.821.
[4] Gong, Y. (2024). Research on recommendation algorithm based on usersentiment analysis. Proceedings of the 4th International Conference onSignal Processing and Machine Learning, 84-91.
[5] Zhang, H.-R., Min, F., He, X., & Xu, Y.-Y. (2015). A hybrid recom-mender system based on user-recommender interaction. MathematicalProblems in Engineering, 2015, Article ID 145636, 11 pages.
[6] Phan,Lan&Huynh,Hung&Huynh,Hiep.(2018).ImplicativeRating-BasedHybridRecommendationSystems.InternationalJour-nal of Machine Learning and Computing. 8. 223-228. 10.18178/i-jmlc.2018.8.3.691.
[7] S. M. Al-Ghuribi and S. A. Mohd Noah, ”Multi-Criteria Review-BasedRecommenderSystem–TheStateoftheArt,”inIEEEAccess,vol.7,pp.169446-169468,2019,doi:10.1109/ACCESS.2019.2954861.
[8] Ko,Hyeyoung&Lee,Suyeon&Park,Yoonseo&Choi,Anna.(2022).ASurvey of Recommendation Systems: Recommendation Models, Tech-niques, and Application Fields. Electronics. 11. 141. 10.3390/electron-ics11010141.
[9] Burke, Robin. (2002). Hybrid Recommender Systems: Survey andExperiments. User Modeling and User-Adapted Interaction. 12.10.1023/A:1021240730564.
[10] F.S¸eker,“EvolutionofMachineLearninginTourism:AComprehensiveReviewofSeminalResearch”,JournalofArtificialIntelligenceandDataScience, vol. 3, no. 2, pp. 54–79, 2023.
[11] Ma, S. (2024). Enhancing Tourists’ Satisfaction: Leveraging ArtificialIntelligence in the Tourism Sector. Pacific International Journal, 7(3),89–98.
[12] G.Linden,B.SmithandJ.York,”Amazon.comrecommendations:item-to-item collaborative filtering,” in IEEE Internet Computing, vol. 7, no.1, pp. 76-80, Jan.-Feb. 2003.
[13] Zheng, Yu & Xie, Xing & Ma, Wei-Ying. (2010). GeoLife: A Collabo-rative Social Networking Service among User, Location and Trajectory.IEEE Data Eng. Bull.. 33. 32-39.
[14] Onwuegbuzie, Anthony J., Nancy L. Leech, and Kathleen MT Collins.”Qualitative analysis techniques for the review of the literature.” Qual-itative Report 17 (2012): 56.
[15] Church,KennethWard.”Word2Vec.”NaturalLanguageEngineer-ing 23.1 (2017): 155-162.
[16] Ma, Long, and Yanqing Zhang. ”Using Word2Vec to process big textdata.” 2015 IEEE International Conference on Big Data (Big Data).IEEE, 2015.
[17] Tan, Kian Long, et al. ”RoBERTa-LSTM: a hybrid model for sentimentanalysiswithtransformerandrecurrentneuralnetwork.”IEEEAccess10(2022): 21517-21525.