Skip to content

Latest commit

 

History

History
91 lines (60 loc) · 5.35 KB

README.md

File metadata and controls

91 lines (60 loc) · 5.35 KB
Raw

Movie-Recommender-System

A system incorporating collaborative, content-based, and hybrid techniques to offer personalised movie recommendations, thereby improving the overall user experience.

Table Of Contents

Introduction

This project develops a movie recommender system that leverages collaborative, content-based, and hybrid filtering techniques to deliver personalized movie recommendations. By analysing user interactions and movie features and combining multiple recommendation strategies, the system aims to enhance the user experience by suggesting movies that align with individual preferences.

Graphical User Interface (GUI)

A simple GUI for a movie recommender system is created using Tkinter, allowing users to interact with the recommendation engine easily. Users can enter their user ID and select their preferred recommendation technique. Upon submission, the GUI processes the input and displays a list of recommended movies tailored to the user's tastes.

The code for the GUI is located at recommender_system_GUI.py within the repository.

Methodology

Data Collection

The "ml-latest-small" dataset presents a comprehensive collection of movie ratings and tags, amassed from a diverse user base. The dataset comprises

  • 9,742 movies, encompassing various genres and periods.
  • 610 users, who have actively participated in rating movies and tagging content.
  • 1,589 unique tags, indicating the varied preferences and tastes of the user base.
  • A substantial number of ratings, amounting to 100,836, which illustrates the extensive engagement of users with the platform.

Recommendation Techniques

  1. Content-Based Filtering: Recommends items based on user preferences and item features.
  2. Memory-Based Collaborative Filtering: Uses user-item interactions to suggest similar items or users.
  3. Model-Based Collaborative Filtering: Utilises machine learning models to predict user preferences based on past data.

Evaluation Method

  1. RMSE (Root Mean Square Error): RMSE quantifies the average predictive error between actual and predicted ratings, with lower values indicating better accuracy.
  2. MAE (Mean Absolute Error): MAE calculates the average absolute difference between predicted and actual ratings, with lower values signifying improved prediction accuracy.
  3. Hit Rate: Hit Rate measures the proportion of recommended items that match user interactions or preferences.
  4. Coverage: Coverage assesses the percentage of items in the catalogue that the recommender system can suggest.
  5. Novelty: Novelty evaluates the uniqueness and diversity of recommendations to introduce users to unfamiliar items.
  6. Recall @k: Recall @ k quantifies the fraction of relevant items recommended within the top-k list.
  7. Precision@k: Precision @ k measures the accuracy of relevant items within the top-k recommendations.

Results

Root Mean Squared Error (RSME)

  • User-based and Item-based have the lowest RMSE, suggesting they are the most accurate in predicting exact ratings.

  • Year-based and Weighted year-based have the highest RMSE, implying they might not be as accurate in predicting ratings.

Mean Absolute Error (MAE)

  • User-based and item-based excel, indicating they are adept at accurately predicting ratings.

  • Year-based has the highest MAE, followed closely by combined content-based.

Hit Rate

  • SVD has an outstandingly high hit rate, implying it’s most effective at suggesting movies users will interact with.

  • NCF model anUser-based have very low hit rates, suggesting users might not find their recommendations as engaging.

Coverage

  • Almost all recommenders have a coverage of 1.0000 or near to 1. This means they can potentially recommend any movie in the dataset

  • NCF model has an unusually low coverage of 0.0001 or 0.01%. This means it can only recommend a tiny fraction of the available items.

Novelty

  • The NCF model stands out with an extraordinarily high novelty score, suggesting it recommends less popular items.

  • SVD has the lowest novelty, implying it tends to suggest more popular or mainstream movies.

Precision at K (P@k)

  • SVD dominates, implying that a high proportion of its top recommendations are items users have interacted with.

  • The NCF model has the lowest precision, suggesting its top recommendations are rarely hit with users.

Recall at K (R@k)

  • SVD filtering excels, suggesting it's able to capture most of the items users have interacted with in its top recommendations.

  • User-based and NCF model filtering have a very low recall, indicating they miss out on many movies users would interact with.

License

This project forms part of an academic course and is intended solely for educational purposes. It may include references to copyrighted materials and any such materials are utilised exclusively for scholarly use. For guidance on sharing or distributing this work, it is advisable to seek consultation from your instructor or institution.

For more details, see the LICENSE file.

Reference

Dataset: ml-latest-small