🍵 Matcha Café Recommendation System

A manually collected dataset and content-based recommendation system for exploring and recommending matcha drinks offered by cafés in Belgrade.

📌 Project Overview

This project explores matcha drinks available in cafés across Belgrade. The dataset was manually collected from café menus, websites, social media, Google Reviews, and food-delivery platforms such as Wolt.

The project aims to analyze the local matcha offering and develop a content- and aspect-based recommendation system. The recommender will suggest drinks based on characteristics such as sweetness, sentiment, ingredients, preparation style, and keywords extracted from customer reviews.

Installation and usage

git clone https://github.com/username/matcha-recommender.git
cd matcha-recommender
pip install -r requirements.txt

🎯 Objectives

• Create a structured dataset of matcha drinks offered in Belgrade.
• Analyze prices, ingredients, locations, and drink characteristics.
• Extract sentiment, sweetness level, and keywords from customer reviews.
• Explore spatial patterns in the local matcha offering.
• Develop a content- and aspect-based recommendation system.

📥 Data Collection

Data was collected manually from publicly available sources, including:

• Official café menus and websites
• Instagram and other social media pages
• Google Maps and Google Reviews
• Customer-submitted photographs on Google Reviews
• Wolt listings

Dataset Structure

Each row in the review-level dataset represents one customer review.

column_name	description	possible_values	origin
cafe_id	Unique numerical identifier for each café.	Integer (e.g. 1, 2, 3)	Manually assigned
cafe_name	Name of the café as listed on Google Maps.	Text (e.g. Wagokoro)	Google Maps
location	City district or neighborhood where the café is located.	Text (e.g. Stari Grad, Vračar, Savski Venac)	Google Maps / manually determined
latitude	Geographic latitude coordinate of the café.	Decimal number (e.g. 44.81787914205736)	Google Maps
longitude	Geographic longitude coordinate of the café.	Decimal number (e.g. 20.472490255822375)	Google Maps
is_chain	Whether the café belongs to a chain or franchise.	TRUE / FALSE	Manually determined
cafe_rating_google	Overall café rating displayed on Google Maps.	Decimal from 1.0 to 5.0 (e.g. 4.8)	Google Maps
num_reviews_cafe	Total number of Google Maps reviews for the café.	Integer (e.g. 735)	Google Maps
review_id	Unique identifier for each review, created using cafe_id and a sequential number.	Text (e.g. 1_01, 1_02, 2_01)	Manually generated
drink_name	Standardized English name of the drink.	Text (e.g. matcha latte, cold strawberry matcha)	Collected and manually standardized
flavor_type	Syrup used to sweeten the matcha or another distinct flavor or aroma.	Text (e.g. mango)	Café menu, social media, Wolt, or manually inferred
liquid_type	Type of milk or other liquid used in the drink.	Text (e.g. almond, water)	Café menu, reviews, images, or estimated
temperature	Serving temperature of the drink.	Text (e.g. hot, cold)	Café menu, reviews, images, or estimated
drink_size_ml	Estimated or stated volume of the drink in milliliters.	Integer (e.g. 350, 360)	Café menu, Wolt, social media, review images, or estimated
matcha_quality	Matcha grade stated on the menu or visually inferred from images.	Text (e.g. culinary, ceremonial)	Café menu or visual estimate
price_rsd	Price of the drink in Serbian dinars.	Integer (e.g. 350, 400)	Café menu, website, social media, or Wolt
price_per_100ml	Calculated price per 100 ml: price_rsd / drink_size_ml * 100.	Decimal number (e.g. 100.0, 114.28)	Calculated
menu_descriptors	Descriptive terms used by the café to describe the drink.	Comma-separated text (e.g. smooth, creamy, earthy)	Café menu, website, social media, or Wolt
drink_rating	Numerical rating given specifically to the drink when explicitly stated in a review.	Integer from 1 to 5	Google Reviews
review_text	Raw text of a Google Maps review mentioning the drink.	Text	Google Reviews; manually collected
sentiment_score	Sentiment score extracted from review_text.	Decimal from -1.0 to 1.0 / empty until processed	NLP-derived
sweetness_level	Sweetness level inferred from review_text.	high / medium / low / unknown / empty until processed	NLP-derived
keywords	Key descriptive terms extracted from review_text.	Comma-separated text / empty until processed	NLP-derived

📁 Project Structure

matcha/
├── data/
│   ├── processed/
│   └── raw/ 
├── images/
├── notebooks/
│   ├── 01_data_cleaning.ipynb
│   ├── 02_eda.ipynb
│   ├── 03_nlp_feature_extraction.ipynb
│   └── 04_recommender_system.ipynb
├── src/
│   ├── feature_extraction.py
│   └── recommender.py
├── README.md
└── requirements.txt 

⚙️ Methodology

The project methodology consists of the following stages:

1. Manual data collection and source documentation
2. Data validation, cleaning, and preprocessing
3. Exploratory data analysis
4. NLP-based feature extraction:
   • Sentiment analysis using a pretrained sentiment model
   • Sweetness-level classification using a rule-based or model-based approach
   • Keyword extraction using TF-IDF, KeyBERT, or another suitable method
5. Aggregation of review-level features into drink-level profiles
6. Content- and aspect-based recommendation using cosine similarity or another suitable similarity measure
7. Evaluation of the NLP pipeline and recommendation quality

📊 Exploratory Data Analysis

🚧 Status: In progress

This section will include key visualizations and findings related to cafés, drinks, prices, locations, and NLP-derived review features.

🤖 Recommendation Approach

Since the dataset does not contain user identifiers or interaction histories, collaborative filtering is not applicable. The system therefore uses a content- and aspect-based approach.

Each drink is represented using its structured attributes and review-derived features, including sweetness, sentiment, ingredients, and descriptive keywords. User preferences are compared with drink profiles to generate ranked recommendations.

📊 Results

This section will include:

• Key findings from the exploratory data analysis
• An example user query and the corresponding top five recommendations
• Evaluation results for sentiment, sweetness-level, and keyword extraction
• Evaluation of recommendation relevance and quality
• A screenshot or demonstration of the final application

Limitations

• It is impossible to determine the exact milk_type used for each drink. Many cafés only state on their menus that they offer a plant-based option, without specifying the exact type of milk. In some instances, the milk type could be derived from Google Reviews. Otherwise, cow’s milk was assumed to be the default option. For drinks such as matcha lemonade or yuzu matcha, the milk type was set to water, tonic water, orange juice or grapefruit juice for bumble.
• It was also difficult to determine the exact drink temperature from Google Reviews. When users included a picture with their review, it was sometimes possible to determine whether the drink was hot or cold. Otherwise, the default value was set to cold, as iced matcha lattes are the most popular option.
• Drink size was also impossible to determine with complete accuracy. When the drink_size_ml was stated on the menu, that value was used. Otherwise, it was estimated based on the cup size shown on social media, Instagram posts, or images submitted with Google Reviews. Wolt listings were also checked when available, as they sometimes included the drink size. Consequently, price_per_100ml is also an estimate in most cases.
• Matcha type was estimated based on images posted on social media or included in Google Reviews. If the color appeared bright green, the matcha_type was classified as ceremonial grade, while duller or more yellow matcha was classified as culinary grade.
• menu_descriptors were scarce. When available, they were derived from social media posts by local users and online menus found on Google, social media, Wolt, or the café’s website.
• review_text and the other values described above were collected manually for this first version of the dataset.

🚧 Future Work

Task	Description	Status
Spatial analysis	Split latitude_longitude into separate latitude and longitude variables.	🟢 Done
Data preparation	Group all unspecified plant-based milk types under the plant category in milk_type.	🟢 Done
NLP feature extraction	Derive and populate sentiment_score, sweetness_level, and keywords using NLP.	🟢 Done
Exploratory data analysis	Conduct additional analyses and create informative visualizations.	🟡 In progress
Recommendation system	Develop a content- and aspect-based recommender system.	🟡 In progress

🔁 Reproducibility

To ensure that the project can be reproduced:

• All required dependencies are listed in requirements.txt.
• A fixed random seed is used where applicable.
• The recommended order for running the notebooks is documented.
• Raw and processed datasets are stored separately in the data/raw and data/processed directories.
• NLP-derived columns can be regenerated by running the corresponding feature-extraction notebook.

⚖️ Ethical Considerations

The dataset contains publicly available Google Review text collected for research and educational purposes. No reviewer names, usernames, profile links, photographs, or other personal information are stored.

Review texts remain attributable to their original authors and source platform.

📄 License

This project is licensed under the MIT License.

👤 Author

Ana Petrović
Master Engineer of Information Systems with a focus on data science and machine learning.