Building a 2-Stage Anime Recommender: From Naive TF-IDF to Industry-Grade Architecture

The Origin Story & The Problem

As an avid anime fan, my first recommendation engine used a basic TF-IDF vectorization and cosine similarity filter.

It quickly hit a massive accuracy wall:

• - Semantic blindness: Only matching exact keyword overlaps.
• - Popularity bias: Missing latent user preferences by defaulting to mainstream hits.
• - Scalability bottlenecks: Struggling with the single-stage problem of scaling catalog comparisons.

This failure pushed me to build a production-grade two-stage recommendation architecture (inspired by YouTube’s landmark research) that blends fast candidate retrieval with deep personalization to help fellow fans discover true hidden gems.

---

The Approach

To solve the scale and accuracy trade-off, I split the recommendation engine into a two-stage pipeline:

• 1. Candidate Generation (Retrieval):
  • - Content Filtering: Generated dense semantic embeddings of synopses and metadata using sentence-transformers/all-MiniLM-L6-v2 to capture deep thematic relationships.
  • - Collaborative Filtering: Leveraged user-item interaction matrices processed via Alternating Least Squares (implicit/ALS) to extract latent user preferences and peer behavior.
• 2. Re-Ranking:
  • Combined and filtered candidates passed into a LightGBM model (complete with hyperparameter tuning).
  • The ranker evaluates user history, metadata features, and cross-interaction scores to output a finely tuned top-10 list.

---

The Results

• 3.2x training speedup and efficiency boost through sparse-matrix optimizations and LightGBM.
• 45ms end-to-end inference latency via FastAPI, ensuring real-time responsiveness.
• Significantly higher recommendation relevance and diversity compared to the legacy TF-IDF approach.

---

Key Architectural Decisions

• Decoupled Retrieval from Ranking: While candidate generation casts a wide net to pull hundreds of relevant items instantly using vector search and matrix factorization, the heavy lifting is reserved for the gradient-boosting re-ranker.
• Feature-Rich Interactions: Allowing LightGBM to evaluate complex feature interactions without blowing past runtime latency limits.