galsolomon9/final.assignment

Exploratory Data Analysis (EDA) and Embeddings Evaluation

This dataset contains synthetic venue recommendations for London, generated to support a semantic recommendation system. Each record represents a recommended venue with metadata such as venue type, area, price level, and a natural language explanation.

---

Dataset Overview

• Number of user queries: 10,000
• Total recommended venues: 30,000
• Unique venues: 513
• Columns:
  • venue_name
  • venue_type
  • area
  • price_level
  • why

Each venue recommendation includes a short textual explanation describing why it matches the user's intent.

---

Exploratory Data Analysis (EDA)

Venue Type Distribution

The dataset is imbalanced across venue types, with cafés and wine bars being the most frequent categories. This reflects common urban recommendation patterns and should be considered when designing downstream recommendation logic.

Area Distribution

Venues are distributed across a wide range of London neighborhoods. No single area dominates the dataset, ensuring geographical diversity in recommendations.

Price Level Distribution

The three price levels (budget, mid, splurge) are relatively balanced, supporting recommendations for users with different spending preferences.

Text Length Analysis

The explanation texts (why field) show a consistent length distribution, indicating uniform quality in the synthetic data generation process.

---

3

Embeddings Generation

To enable semantic similarity search, a text representation (item_text) was constructed by concatenating:

• venue type
• area
• price level
• explanation text

This combined text was used as input for multiple sentence embedding models.

---

Embedding Models Evaluation

Three different Hugging Face sentence-transformer models were evaluated:

Model	Embedding Dim	Time (seconds)	Notes
all-MiniLM-L6-v2	384	~273	Fast, lightweight, strong baseline
paraphrase-MiniLM-L6-v2	384	~265	Optimized for semantic similarity
all-mpnet-base-v2	768	~1859	High quality but slow and heavy

---

Model Selection

The final model selected for the recommendation system is:

sentence-transformers/all-MiniLM-L6-v2

Justification:

• Provides strong semantic representations
• Significantly faster than larger models
• Lower embedding dimensionality (384)
• More suitable for scalable recommendation systems
• Best balance between performance and efficiency

The larger MPNet-based model was included for comparison but is less practical for production-scale usage due to its computational cost.

---

Saved Artifacts

• embeddings_all_MiniLM_L6_v2.npy – venue embeddings
• items_metadata.parquet – metadata mapping embeddings to venues

These files enable efficient similarity search and recommendation generation.

---

4

Input–Output Recommendation Pipeline

An end-to-end recommendation pipeline was implemented to transform free-text user input into meaningful venue recommendations.

Pipeline Overview

1. The user provides a natural language query describing their preferences (e.g., desired venue type, location, and budget).
2. The user query is encoded into a semantic embedding using the selected sentence-transformer model.
3. Cosine similarity is computed between the user embedding and all precomputed venue embeddings.
4. Venues are ranked based on similarity scores.
5. The top-K most similar venues are returned as the AI-generated recommendation output.

Example

For a user query such as:

"Looking for a cozy cafe to journal in Soho, budget friendly"

The system successfully returns cafés located in Soho with a budget price level, demonstrating semantic understanding of both location and intent.

Design Considerations

• Embeddings are computed offline and stored to improve runtime efficiency.
• The same embedding model is used for both dataset items and user queries to ensure consistency.
• Cosine similarity was chosen as a simple and effective metric for semantic comparison.

This pipeline demonstrates a complete USER INPUT → PIPELINE → AI OUTPUT flow, fulfilling the requirements of the final project.