Recipe Recommendation System

---

Introduction

The Recipe Recommendation System is designed to provide personalized and dynamic recipe recommendations by leveraging the Food.com Recipes and Interactions dataset (RAW_recipes.csv). This dataset contains a rich collection of recipes, including fields like recipe names, ingredients, cooking steps, and user reviews. With over 200K recipes and their associated interactions, the goal was to extract meaningful insights and deliver an intuitive user experience through a recommender system.

---

Dataset

The Food.com Recipes and User Interactions dataset contains recipes along with user interaction data, such as ratings and reviews. This dataset provides a comprehensive foundation for building a recommendation system by combining recipe details with user feedback.

• Dataset Highlights:

  • Includes recipe names, ingredients, cooking steps, and user reviews.
  • Over 200K recipes with associated interactions.

• Download:

  • The dataset can be downloaded from Kaggle.

---

Approach

1. Dataset Preprocessing

The RAW_recipes.csv dataset was first explored and analyzed for its structure and contents. The following steps were carried out to prepare the data:

• Handling Missing Values: Incomplete rows with missing recipe names, ingredients, or steps were filtered out.
• Outlier Detection: Recipes with unusually high numbers of steps or ingredients (e.g., >50 steps) were identified as potential outliers and excluded to reduce noise in the recommendations.
• Text Cleaning: Ingredients and steps were cleaned to remove punctuation, special characters, and unnecessary whitespace.

2. Feature Engineering

• TF-IDF Encoding:
  • Textual fields such as ingredients and steps were converted into numerical representations using TF-IDF (Term Frequency-Inverse Document Frequency) encoding.
  • This encoding allowed us to capture the importance of terms relative to the entire dataset, making the recommendation process more effective.

3. Model Implementation

• Content-Based Filtering:
  • Focused on recommending recipes similar to the ones the user interacted with, based on textual features. Several models were experimented with to generate recommendations:
• 1. K-Nearest Neighbors (KNN):
  • Implemented a similarity-based approach to recommend recipes based on their TF-IDF embeddings.
  • Recipes with similar ingredient profiles and cooking steps were identified and ranked.
• 2. Cosine Similarity:
  • Used cosine similarity metrics on TF-IDF vectors to find the most relevant recipes for a user.

---

Challenges

1. Curse of Dimensionality

• The dataset contained a vast amount of textual data, resulting in extremely high-dimensional TF-IDF vectors. This made it challenging to efficiently compute similarities without running into performance bottlenecks.

2. Memory Allocation Errors

• Due to the large dataset size and high-dimensional feature space, memory allocation errors were encountered during similarity computation, especially with models like KNN and cosine similarity.

3. Balancing Trade-offs

• Achieving a balance between reducing dimensionality and retaining the semantic richness of the recipes was a significant challenge. Over-reduction led to loss of important features, while under-reduction increased computational overhead.

---

Working Demo

Below are some screenshots showcasing the user interface and functionality of the Recipe Recommendation System:

1. Homepage:

2. Recipe Recommendation Results:

   

Future Work

1. Embedding-Based Models

• Explore embedding models like BERT, Sentence Transformers, or other LLM-based models (e.g., GPT) to encode recipes into dense, low-dimensional vectors.
• These embeddings can better capture the context and semantic meaning of recipes, improving recommendation quality.

2. LLM-Powered Recommender System

• Develop an advanced recommender system utilizing Large Language Models (LLMs) to:
  • Understand user preferences more effectively.
  • Provide personalized recommendations by leveraging the contextual richness of LLMs.
  • Overcome limitations of traditional approaches, such as sparse data and high dimensionality.

4. Collaborative Filtering

• Combine content-based filtering with collaborative filtering techniques to provide hybrid recommendations, leveraging both recipe data and user interaction data.

5. Cloud Deployment

• Deploy the system on scalable cloud platforms to handle large-scale datasets and real-time user requests.

---

Conclusion

The Recipe Recommendation System demonstrates the potential of combining TF-IDF encoding, similarity-based modeling, and efficient dimensionality reduction techniques for personalized recipe recommendations. By transitioning to embedding-based approaches and leveraging LLMs in the future, the system can achieve greater scalability, context-awareness, and user satisfaction.

---

How to Use

1. Run the Application

To set up and run the project, follow these steps:

1. Clone the repository to your local machine.

2. Navigate to the project directory.

3. Install the required dependencies by running the following command:

   pip install -r requirements.txt

4. Download the dataset (RAW_recipes.csv) into the current directory. You can find the dataset here.

5. Open and run the Data Preparing notebook to generate the recipe_cleaned.csv file in the current directory.

6. Start the Flask application by running the following command:

   flask run

7. Open your browser and navigate to http://127.0.0.1:5000 to access the application.

---