Enhanced Netflix Recommendation System

CS 410: Text Information Systems Final Project

This project is a content-based Netflix recommendation system that recommends titles based on a user's watch history. The recommender builds user profiles from previously watched titles and compares those profiles against a filtered Netflix title dataset using TF-IDF, Word2Vec, genre similarity, and production-country information.

About

Netflix's current recommendation system estimates how likely a user is to enjoy a title based on interactions with the service. This project explores an NLP-based approach to recommendation by using title metadata and user watch history to generate content-based recommendations. The goal was to apply concepts from CS 410, including TF-IDF, Word2Vec, and similarity-based retrieval.

Model Approach

The recommendation system follows a content-based retrieval approach. For each user, the model builds a profile from the titles they have already watched, then compares that profile against all unwatched titles in the filtered Netflix dataset.

The final score combines description similarity, semantic similarity, genre similarity, country similarity, and a genre-overlap bonus.

score =
    0.30 * description_tfidf_similarity
  + 0.30 * word2vec_description_similarity
  + 0.30 * genre_tfidf_similarity
  + 0.10 * country_tfidf_similarity
  + 0.10 * genre_overlap_bonus

Features

• Content-based title recommendation
• TF-IDF description similarity
• Word2Vec semantic similarity
• Genre and production-country profile matching
• Netflix-style recommendation UI
• Clickable recommendation detail windows
• Precision@K and Recall@K evaluation

Evaluation & Results

The recommender was evaluated using held-out titles from the synthetic user dataset. Precision@K measured how many of the top K recommendations appeared in the user's held-out evaluation list, while Recall@K measured how many held-out titles appeared in the recommendation set.

The model produced an average Precision@25 of 0.0040 and an average Recall@25 of 0.0250 across 60 users. Although these are strict evaluation measures, the system showed that it could capture broad content similarity across Netflix titles.

Limitations

• The system is fully content-based and does not use collaborative filtering.
• The user dataset is synthetic and only contains positive watch-history examples.
• The model does not use ratings, watch time, clicks, or explicit user feedback.
• Users with diverse watch histories can produce broad profiles that are harder to rank from.

Repository

OPEN_GITHUB_REPOSITORY

Code Contributors

Ian Puszynski
Jackson Fleege
Azaan Farooq

Report Contributors

Ian Puszynski
Jackson Fleege
Azaan Farooq
Raphael Ampon
Dennis Thomas