PROJECT 03 — RECOMMENDATION SYSTEMS

Game Recommendation Engine

Semantic similarity content recommendation leveraging dense BERT transformer embeddings.

MAR–MAY 2025
PYTHON · HUGGING FACE BERT · COSINE SIMILARITY
KAGGLE ENRICHED WITH RAWG API
~60% HIT RATIO (50 CAMPUS USERS)

01. Executive Overview

Traditional keyword-matching recommendation engines fail when game descriptions use different vocabularies to describe similar gameplay dynamics. By extracting dense vector embeddings from Hugging Face transformer models (`bert-base-uncased`), this system understands high-level semantic context to recommend games with similar thematic and mechanical patterns.

02. Embedding Space & Cosine Similarity Distribution

Below is a representation of semantic distance similarity scores calculated across target embedding vectors. Higher cosine similarity scores represent closer neighborhood proximity in latent embedding space.

Top Nearest-Neighbor Semantic Similarity Scores
[REAL PROJECT DATA REQUIRED] Evaluated ~60% hit ratio across 50 campus benchmark users.

03. Vector Pipeline Architecture

The vector engine executes three primary steps:

  1. Tokenization & contextual embedding extraction via `transformers.BertModel` ($768$-dimensional vector per game summary)
  2. Pre-computation and indexing of embedding matrix
  3. Cosine similarity ranking: $\text{similarity}(\vec{A}, \vec{B}) = \frac{\vec{A} \cdot \vec{B}}{\|\vec{A}\| \|\vec{B}\|}$

04. Campus Benchmark & Streamlit Interface

The system was evaluated in a campus benchmark trial with 50 participants, achieving a ~60% top-5 hit ratio. The application is hosted as an interactive Streamlit tool where users can input any game title to explore semantic neighborhood recommendations.