dmckinney-ml/movie-recommender-artifacts

🎬 Two-Tower + FAISS + XGBoost Recommender — Inference Artifacts

This dataset contains the serialized inference artifacts for the Two-Stage Movie Recommendation System built using TensorFlow Recommenders, FAISS, and XGBoost.

These artifacts support low-latency retrieval and ranking for the deployed Hugging Face Space:

🔗 Space: two-tower-faiss-xgb-recommender

---

📌 Overview

The recommendation system follows a two-stage architecture:

Retrieval (Stage 1)

• Two-tower neural embedding model (TensorFlow Recommenders)
• Vector similarity search using FAISS (IndexFlatIP)
• Cosine similarity implemented via inner-product normalization

Ranking (Stage 2)

• XGBoost learning-to-rank model
• Input features:
  • User embedding
  • Movie embedding
  • Genre one-hot features

Artifacts are structured for inference-only usage.

---

🗂️ Artifact Contents

File	Description
faiss.index	FAISS index containing 87K+ movie embeddings
xgb_ranker.json	Trained XGBoost ranking model
two_tower/user_model/	Saved user embedding sub-model
two_tower/movie_model/	Saved movie embedding sub-model
two_tower/genre_model/	Saved genre projection model
two_tower/rating_model/	Saved rating prediction head
movie_id_to_row.json	Mapping from MovieLens ID → FAISS index row
movies.parquet	Movie metadata (title + features)

---

📊 Training Data

Models were trained on the MovieLens 32M dataset:

• 32 million ratings
• ~87,585 movies
• ~200,948 users

Preprocessing performed using:

• Apache Beam (Cloud Dataflow)
• BigQuery
• Kubeflow Pipelines (Vertex AI)

---

🧠 Embedding Details

Property	Value
Embedding dimensionality	Tunable (selected via Hyperband)
Vector space	Normalized for cosine similarity
FAISS index type	IndexFlatIP
Total indexed items	~87K

---

📈 Ranking Model

Property	Value
XGBoost objective	rank:pairwise
Evaluation metric	NDCG
Hyperparameter tuning	Optuna (50 trials)
Validation strategy	User-level cold-start split

---

🚀 Intended Usage

These artifacts are designed for:

• ✅ Inference inside the Hugging Face Space
• ✅ Reproducible local experimentation
• ✅ Demonstration of production-style ML architecture

⚠️ Not intended for retraining without the full GCP pipeline.

---

🛠️ Loading Example

import faiss
import xgboost as xgb
import tensorflow as tf

# FAISS index
index = faiss.read_index("faiss.index")

# XGBoost ranker
booster = xgb.Booster()
booster.load_model("xgb_ranker.json")

# Two-tower sub-models
user_model = tf.keras.models.load_model("two_tower/user_model")
movie_model = tf.keras.models.load_model("two_tower/movie_model")

---

📄 License

• Project code: MIT License
• Training data: MovieLens dataset license applies → https://grouplens.org/datasets/movielens/