Recommendations with Text Embeddings (Synthetic Data)

Author
Sam Ginzburg samginzee@gmail.com

---

Business Problem and Motivation

Unstructured customer reviews contain rich qualitative insights, but they are difficult to analyze at scale using traditional rule-based or keyword approaches. Modern embedding models allow text to be transformed into numerical vectors that preserve semantic meaning, enabling similarity search, clustering, and recommendation-style retrieval.

The goal of this project is to demonstrate how text embeddings can be used to:

• Identify latent themes in customer reviews
• Group semantically similar feedback
• Retrieve representative reviews for high-level business topics (e.g., quality, comfort, price)

This approach is directly applicable to product analytics, customer experience analysis, and recommendation systems.

---

Data Source

The original project was inspired by a guided learning exercise on DataCamp using an e-commerce women’s clothing reviews dataset.

To respect licensing and redistribution constraints, this repository does not include the original dataset. Instead, a synthetic dataset was generated to mirror:

• Column structure
• Data types
• Missingness patterns
• Statistical distributions
• Categorical hierarchies (Division → Department → Class)

Synthetic dataset used in this repository:
data/womens_clothing_synthetic.csv

The synthetic text data is newly generated and does not contain verbatim or near-duplicate copies of the original reviews.

---

Methods and Skills Demonstrated

This project demonstrates the following technical skills:

• Natural Language Processing (NLP)
  • Text embedding generation using OpenAI embedding models
  • Semantic similarity via cosine distance
• Data Engineering & Reproducibility
  • Synthetic data generation for safe public sharing
  • Environment variable management for API keys
• Machine Learning & Geometry
  • Vector similarity search
  • Dimensionality reduction using t-SNE for visualization
• Python & Data Science Tooling
  • pandas, NumPy
  • scikit-learn
  • modular, readable notebook structure
• Responsible API Usage
  • Secure handling of API keys via .env
  • Clear documentation for local execution

---

Methodology Overview

1. Text Embedding

   • Customer review text is converted into high-dimensional embedding vectors using a transformer-based embedding model.
   • Each vector represents the semantic meaning of a review.

2. Topic Representation

   • High-level business topics (e.g., Quality, Comfort, Style, Price) are also embedded into the same vector space.

3. Similarity Search

   • Cosine distance is used to measure semantic similarity between topic vectors and review vectors.
   • For each topic, the closest reviews are retrieved as representative examples.

4. Visualization

   • t-SNE is applied to reduce high-dimensional embeddings into two dimensions for exploratory visualization.
   • This allows inspection of clustering behavior and semantic spread.

---

Quick Glance at Results

Embedding Space Visualization (t-SNE)

Each point represents a customer review embedded into semantic space and projected into two dimensions. Proximity reflects semantic similarity.

---

Representative Reviews by Topic

For each predefined business topic, the three most semantically similar reviews are retrieved using cosine similarity.

---

How to Interpret the Results

• Reviews that appear close together in the embedding space tend to express similar sentiments or themes, even when phrased differently.
• Topic-specific nearest-neighbor retrieval surfaces reviews that best exemplify abstract concepts such as comfort or quality.
• This approach enables qualitative insights at scale without manual labeling.

While this analysis does not explain why certain themes emerge, it provides a powerful foundation for:

• Customer feedback summarization
• Product issue discovery
• Recommendation and personalization systems

---

Environment Setup

To run this project locally:

1. Clone the repository
2. Install dependencies: pip install -r requirements.txt
3. Create a .env file in the project root and add: OPENAI_API_KEY=your_api_key_here
4. Open and run: notebooks/Recommendations_Embeddings.ipynb

The .env file is excluded from version control.

---

Credit

This project was inspired by a guided learning exercise on DataCamp and was extended with:

• Synthetic data generation
• End-to-end reproducibility
• Secure API usage
• Additional analysis and visualization

All implementation, restructuring, and extensions are original.