TasteLens 🍽️ : Your Taste, Our Lens ✨

---

✨ Overview

TasteLens is a personalized food recommendation system that suggests food items based on individual preferences. The system leverages multiple machine learning models to analyze user behavior and item characteristics, enabling more accurate and diverse recommendations.

The project explores different recommendation strategies and combines them to improve personalization and recommendation quality.

A web interface was built to display recommendations, allowing users to browse popular dishes and see personalized suggestions generated by the ML models.

---

⚠️ Problem Statement

• Food platforms offer a wide range of choices, making it difficult for users to find dishes that match their personal preferences.
• Most recommendation systems rely on generic or popularity-based suggestions, which fail to capture individual tastes.
• Personalized food recommendation is challenging due to subjective user preferences and the lack of suitable, real-world food preference datasets.
• Using a single recommendation approach often results in biased or less accurate suggestions.

TasteLens aims to address these challenges by applying multiple machine learning–based recommendation models and combining them through a hybrid approach to deliver more relevant and personalized food recommendations.

---

⚙️ Methodology

TasteLens follows a multi-model recommendation approach to deliver personalized food suggestions. Different recommendation techniques capture unique aspects of user behavior, and combining them improves overall recommendation quality.

1. Content-Based Filtering

• Recommends food items by analyzing item attributes and user preferences.
• Builds a preference profile for each user based on past interactions.
• Suggests food items with similar characteristics.
• Advantages: Effective personalization even when user interaction data is limited.

2. User-Based Collaborative Filtering (UBCF)

• Identifies users with similar taste patterns by comparing their rating behavior.
• Recommends items liked by similar users to the target user.
• Advantages: Introduces variety beyond a single user’s history by leveraging shared preferences.

3. Item-Based Collaborative Filtering (IBCF)

• Focuses on relationships between food items rather than users.
• Items receiving similar ratings from multiple users are considered related.
• When a user interacts with a food item, similar items are recommended.
• Advantages: Stable and efficient recommendation generation.

4. Hybrid Recommendation Model

• Combines content-based filtering with collaborative filtering techniques.
• Aggregates recommendations from individual models to produce the final output.
• Advantages: Reduces over-specialization, improves recommendation diversity, and provides balanced results compared to using a single model.

---

🔄 System Architecture

TasteLens is designed as a modular recommendation system, where each component transforms user preference data into meaningful food recommendations. The architecture supports multiple recommendation models and allows their outputs to be combined in a structured manner.

Architecture Overview

The system processes user–food interaction data through a sequence of well-defined stages, ensuring accuracy, scalability, and clarity in recommendation generation.

Workflow Description

1. User Preference Data Input

   • Accepts user interaction data including user IDs, food item IDs, ratings, and basic item attributes.
   • This data represents individual taste patterns and serves as the foundation for personalization.

2. Data Preprocessing and Preparation

   • Cleans and standardizes raw data for consistency across models.
   • Handles missing values, normalizes ratings, and transforms data into model-compatible formats.

3. Independent Model Training

   • Each recommendation model (Content-Based, User-Based CF, Item-Based CF) is trained independently.
   • Each model captures unique aspects of user behavior and item similarity.

4. Model-Specific Recommendation Generation

   • Trained models generate individual recommendation lists based on their methodology.
   • Each list reflects a different perspective on user preference and similarity.

5. Hybrid Recommendation Integration

   • Aggregates recommendations from all models using a hybrid strategy.
   • Balances personalization and diversity while reducing bias from any single model.

6. Final Recommendation Output

   • Integrated recommendation list is presented as the final output.
   • Delivers personalized and relevant food suggestions to users.

System Flow Diagram

User Interaction Data
        ↓
Data Preprocessing
        ↓
Model Training (CBF | UBCF | IBCF)
        ↓
Individual Recommendations
        ↓
Hybrid Integration
        ↓
Final Personalized Recommendations

---

🗃️ Dataset Description

The dataset used in the TasteLens project consists of 1,500 records representing user–food interactions. Each record captures detailed food attributes, nutritional information, user preferences, and contextual restaurant data. This enables the implementation of content-based, collaborative, and hybrid recommendation models.

Dataset Structure

Each row corresponds to a single interaction between a user and a food item, and includes the following attributes:

User Information

• user_id: Unique identifier for each user

Food Item Details

• food_id: Unique identifier for each food item
• dish_name: Name of the dish
• cuisine: Cuisine category
• course: Type of course (starter, main, dessert, etc.)
• ingredients: Key ingredients used in the dish
• img_url: URL of the dish image for visual representation

Nutritional Information

• calories: Energy content
• protein_g, fat_g, carbs_g, fiber_g, sugar_g: Macronutrients
• sodium_mg: Sodium content

Preparation & Dietary Context

• prep_time_min: Estimated preparation time
• diet_type: Dietary category
• allergies: Known allergens
• goals: User dietary goals

User Interaction Data

• rating: User rating
• liked: Binary preference indicator
• times_ordered: Number of times ordered

Restaurant Information

• restaurant_id: Unique identifier
• restaurant_name: Name of the restaurant
• location: Restaurant location
• cuisine_type: Cuisine offered
• delivery_time: Estimated delivery time
• average_cost: Average order cost

Temporal Data

• timestamp: Time of interaction

Dataset Usage

• Dish attributes and img_url support content-based recommendations and UI-level visualization.
• Ratings and interaction data are used for collaborative filtering.
• Combined features contribute to the hybrid recommendation model.

---

🛠️ Feature Engineering & Preprocessing

The TasteLens dataset was preprocessed and transformed to extract meaningful features for all four recommendation models: Content-Based, User-Based Collaborative Filtering (UBCF), Item-Based Collaborative Filtering (IBCF), and Hybrid. The goal was to ensure that each model could effectively capture user preferences and item similarities.

1. Content-Based Filtering

• Feature Combination: Multiple text-based features were combined into a single representation with weighted importance:

combined = cuisine (2x) + diet_type (1.5x) + course + ingredients + goals

• Missing Values: Filled any missing values in the combined text column with empty strings.
• Vectorization: Used TfidfVectorizer to convert textual features into numerical vectors.
• Similarity Calculation: Computed cosine similarity between dishes to identify similar items for each user.

2. Item-Based Collaborative Filtering (IBCF)

• Relevant Columns: Extracted user_id, dish_name, and rating.
• Data Cleaning: Converted ratings to numeric and dropped rows with missing values.
• Index Mapping: Created numeric mappings for users and items to build a sparse user-item matrix.
• Similarity Computation: Applied cosine similarity on the transposed matrix (items × users) to measure item-item similarity.

3. User-Based Collaborative Filtering (UBCF)

• User-Item Matrix: Rows = user_id, Columns = dish_name, Values = rating.
• Missing Values: Handled appropriately to ensure matrix completeness.
• Similarity Computation: Applied cosine similarity on the user-item matrix to identify similar users.
• Evaluation: Split dataset into train and test sets to assess model performance.

4. Hybrid Recommendation Model

• Approach: Combined explicit collaborative filtering (SVD-based) with content-based similarity features.
• Text Features: Concatenated dish_name, cuisine, and restaurant_name and vectorized using TF-IDF.
• Similarity Calculation: Computed cosine similarity for content-based features.
• SVD Collaborative Filtering: Trained model on user-food ratings using the surprise library.
• Final Recommendations: Aggregated predictions from both approaches to generate robust recommendations.

Key Insights

• Preprocessing ensured all models received clean, structured, and meaningful input.
• Transformed text features, numerical ratings, and user interactions into model-compatible formats.
• Cosine similarity and matrix factorization (SVD) formed the backbone of the recommendation calculations.

---

💻 Tech Stack

The TasteLens project leverages frontend, backend, and machine learning technologies to deliver a complete food recommendation system.

1. Frontend 🖥️

• React.js – For building a responsive, dynamic user interface.
• JSX – To structure UI components and seamlessly integrate JavaScript with HTML.

2. Backend ⚙️

• Python – Handles server-side logic, API requests, and ML model integration.
• Flask / FastAPI – Serves APIs that connect the frontend with machine learning models.

3. Machine Learning 🤖

• Content-Based Filtering – Uses TF-IDF vectorization and cosine similarity.
• Collaborative Filtering (User-Based & Item-Based) – Computes cosine similarity on user-item matrices.
• Hybrid Model – Combines SVD-based collaborative filtering with content-based features.

4. Libraries & Tools 📦

• Pandas, NumPy – For data manipulation and preprocessing.
• Scikit-learn – For TF-IDF vectorization and similarity calculations.
• Surprise – Implements SVD collaborative filtering.
• Scipy (csr_matrix) – For sparse matrix operations in collaborative filtering.

---

📊 Results & Observations

The TasteLens system was evaluated using Precision@5, Recall@5, and F1 Score@5 metrics for all four recommendation models. These metrics measure how accurately the system recommends relevant dishes to users.

Model	Precision@5	Recall@5	F1 Score@5
Content-Based Filtering	0.026	0.061	0.036
User-Based Collaborative Filtering	0.490	1.000	0.658
Item-Based Collaborative Filtering	0.5607	0.7791	0.6521
Hybrid Recommendation	0.6187	1.0000	0.7645

🔑 Key Observations

• Content-Based Filtering (CBF): Performs poorly due to limited user-specific feature information, resulting in low precision and recall.
• User-Based Collaborative Filtering (UBCF): Achieves high recall but moderate precision, indicating it recommends most relevant items but may include less precise suggestions.
• Item-Based Collaborative Filtering (IBCF): Balances precision and recall better than UBCF, showing effective similarity-based item recommendations.
• Hybrid Model: Outperforms all other models with the highest F1 Score, demonstrating that combining content-based and collaborative filtering techniques provides the most balanced and accurate recommendations.

Final Observations

• Individual models capture different aspects of user preferences, but a hybrid approach significantly improves overall recommendation quality.
• The system is capable of delivering personalized and relevant food suggestions despite limitations in dataset size and scope.

---

✅ Conclusion

TasteLens demonstrates the effectiveness of a multi-model recommendation system for personalized food suggestions. By combining content-based filtering, user-based and item-based collaborative filtering, and a hybrid approach, the system can:

• Capture individual user preferences.
• Recommend items based on similar users and similar dishes.
• Balance personalization and diversity through hybrid aggregation.

Evaluation Insights:

• Content-Based Filtering (CBF) alone performs poorly due to limited user-specific features.
• Collaborative Filtering techniques (UBCF & IBCF) significantly improve recommendation quality.
• The Hybrid Model outperforms all individual approaches, achieving the highest F1 Score, highlighting that combining multiple strategies is crucial for accurate and relevant recommendations.

---

🚀 Future Scope

1. Dataset & Personalization

• Larger and Real-World Dataset: Incorporate a more extensive user dataset to improve model accuracy and generalization.
• Dynamic User Preferences: Use temporal modeling to capture changing tastes over time.
• Enhanced Personalization: Analyze user behavior, preferences, and context for highly tailored recommendations.
• Context-Aware Recommendations: Consider time of day, location, or dietary restrictions for smarter suggestions.

2. Platform & Features

• Full-Fledged Website with Dashboards: Build a web platform with multiple dashboards for users, admins, and restaurants, providing a smooth and interactive experience.
• Integration with Live Platforms: Deploy TasteLens as a web or mobile app with real-time recommendation updates.
• Visual Recommendation Features: Use the img_url column for image-based similarity or deep learning-driven recommendations.

3. Unique & Advanced Features

• Unique Feature Ideas:
  • Recipe Mood Match: Recommend dishes based on mood, weather, or cravings.
  • Image-Based Search: Suggest dishes similar to a user-uploaded photo.
  • Gamification / Rewards: Introduce points or achievements to increase engagement.
• Advanced Hybrid Techniques: Experiment with ensemble methods, deep learning, or graph-based recommendation models to further enhance recommendations.

---