1.
Project Title
Movie Recommendation System
2. Team Details
Name: Daksha Jha
Roll No: BETN1CS22012
Role: Project Lead (Overall coordination, Backend & ML
Development)
Name: Shalini
Roll No: BETN1CS22043
Role: (Frontend Development, Documentation)
3. Introduction
In the digital entertainment era, platforms such as Netflix, Amazon Prime, and
Disney+ offer extensive libraries with thousands of titles across multiple genres,
languages, and formats. While this abundance empowers users with choice, it also
creates a challenge known as the paradox of choice — users struggle to efficiently
identify content aligned with their tastes. Studies show that without effective
guidance, users spend more time browsing than watching, leading to dissatisfaction
and reduced platform engagement.
Recommendation systems address this challenge by leveraging machine learning,
data mining, and personalization algorithms to filter vast datasets and provide users
with context-aware, personalized movie suggestions. These systems enhance user
experience, extend watch time, and improve platform revenue.
This project proposes the development of a Movie Recommendation System that
integrates classical approaches like Collaborative Filtering (CF) and Matrix
Factorization (MF) with a scalable web-based implementation. The system will use
[Link] as a responsive frontend framework and Django API as the backend to
provide real-time recommendations. The architecture is designed to be extensible,
supporting future integration with hybrid models (content-based + CF), contextual
filtering (time of day, device usage), and even deep learning models (neural
collaborative filtering, attention networks).
1
4. Literature Review
The field of recommender systems has evolved through multiple paradigms:
Collaborative Filtering (CF):
User-based CF recommends items based on similarity between user
preference profiles, while item-based CF finds similarities between items and
recommends accordingly. Despite its popularity, CF struggles with issues like
data sparsity, cold start problem, and scalability when handling millions of
users and items.
Matrix Factorization (MF):
Popularized by the Netflix Prize Challenge (2006–2009), MF techniques such
as Singular Value Decomposition (SVD) and Alternating Least Squares (ALS)
achieved remarkable improvements in accuracy by uncovering latent factors
representing hidden user and movie features. These methods demonstrated
that beyond surface-level ratings, deeper preference dimensions could be
modeled.
Hybrid & Context-Aware Models:
Research has shown that combining CF with content-based features (e.g.,
genre, cast, language) significantly improves recommendation quality.
Additionally, incorporating temporal and contextual signals (e.g., weekend
vs weekday viewing, mobile vs TV usage) leads to more dynamic and
personalized experiences.
Deep Learning in Recommendations:
Recent advancements apply autoencoders, recurrent neural networks (RNNs),
and neural collaborative filtering (NCF) to model complex, nonlinear user-
item interactions. Models such as Wide & Deep Learning and Transformer-
based recommenders are increasingly being adopted in production
environments by platforms like YouTube and TikTok.
This literature establishes the importance of adopting hybrid and scalable
approaches in recommendation engines, motivating this project’s design.
2
5. Problem Statement
A movie recommendation system enhances user experience by suggesting films
aligned with individual preferences. Existing platforms rely on either simple
popularity-based listings or advanced proprietary models, but for academic and
small-scale use cases, there is a need for a system that balances practical
implementation with industry-standard practices.
This project aims to develop a movie recommendation system that uses
collaborative filtering and content-based filtering on a reliable dataset (such as
MovieLens). To avoid the cold-start problem, users will be asked to rate a few
genres and sample movies at the beginning, which provides the system with an
initial preference profile. Over time, likes, dislikes, and recent interactions will
further refine these preferences.
The model training process will be tracked through MLflow, while Airflow will be
used for scheduled retraining with updated data. A queue mechanism will capture
real-time user feedback (likes/dislikes). Recommendations will be delivered via an
API with caching, ensuring responsiveness and scalability.
6. Objectives
Model Development
o Implement collaborative filtering (user-user, item-item) and content-based
filtering models.
o Train on the MovieLens dataset to ensure realistic recommendations.
o Use genre and sample movie ratings during onboarding to create the first
user preference profile.
o Continuously refine recommendations using likes, dislikes, and recent watch
history.
System Design
o Build a Django API backend to handle authentication, movie ratings, and
recommendations.
o Set up a [Link] frontend for an interactive user interface.
o Use a queue mechanism to capture real-time user interactions for future
training.
o Employ Airflow (scheduled jobs) to retrain the model periodically.
o Track multiple training runs and models using MLflow.
Inference & API
o Deploy an inference server that fetches the best model from MLflow.
o Use caching to speed up responses.
o Provide endpoints for:
Onboarding (genre & movie rating), Ongoing likes/dislikes
Personalized movie suggestions
3
7. Proposed Methodology
1. Dataset Collection
Use the MovieLens dataset containing movies, genres, and user ratings.
2. Preprocessing
Clean and normalize data (handle missing values, encode genres).
Create a user-item matrix for collaborative filtering.
3. Model Training
Train baseline models: user-user CF, item-item CF, content-based
filtering.
Compare models using evaluation metrics (RMSE, MAE, Precision@K).
Track experiments and versions using MLflow.
4. Feedback Integration
At onboarding, users rate genres and sample movies.
Store feedback (like/dislike/recent watch) in a database + queue system.
Update the model incrementally or during retraining cycles (scheduled via
Airflow).
5. System Deployment
Backend (Django API): Serves recommendations via endpoints.
Frontend ([Link]): Allows users to log in, rate movies, and view
recommendations.
Inference Server: Loads the current best model from MLflow.
Cache Layer: Improves latency for frequently accessed recommendations.
8. Expected Outcomes
A functioning movie recommendation system that provides personalized
suggestions.
Onboarding flow where users rate genres and movies to avoid cold-start
problems.
Recommendations continuously improve with likes/dislikes and recent
interactions.
A Django API integrated with a basic frontend for real-time use.
Reproducible ML workflow using MLflow (tracking) and Airflow (retraining).
A scalable, industry-inspired system without unnecessary complexity.
4
9. Resources Required
Software:
Python, Django, [Link], Pandas, NumPy, Scikit-learn
MLflow (model tracking), Airflow (scheduling), Redis/PostgreSQL
(database/cache)
Hardware:
Standard laptop/desktop (8 GB RAM, i5 processor or equivalent)
Optional: Cloud (AWS/GCP/Azure) for scalability testing
Dataset:
MovieLens dataset (GroupLens Research, free and widely used).
10. Ethical Considerations
User Privacy: Ensure no sensitive data is exposed. Use anonymized datasets
(MovieLens is already anonymized).
Bias Avoidance: Prevent overfitting to popular movies by ensuring diversity
in recommendations.
Transparency: Keep recommendation logic explainable for educational use.
Fair Use: Dataset is used under research/academic license only.
11. Learning Outcomes
Hands-on understanding of collaborative and content-based filtering.
Practical experience with end-to-end ML pipeline (training, tracking,
deployment).
Knowledge of MLflow for experiment tracking and Airflow for automation.
Skills in building web-integrated machine learning applications using
Django API and a frontend.
Awareness of ethical implications in recommendation systems.
5
12. Reference
[1] Y. Koren, R. Bell, and C. Volinsky, "Matrix factorization techniques for recommender
systems," IEEE Computer, vol. 42, no. 8, pp. 30–37, 2009.
[2] J. Bennett and S. Lanning, "The Netflix prize," Proceedings of KDD Cup and
Workshop, 2007.
[3] X. Su and T. M. Khoshgoftaar, "A survey of collaborative filtering techniques,"
Advances in Artificial Intelligence, vol. 2009, Article ID 421425.
[4] GroupLens, "MovieLens Dataset," Available: [Link]
[5] F. Ricci, L. Rokach, and B. Shapira, "Recommender Systems: Introduction and
Challenges," in Recommender Systems Handbook, Springer, 2015, pp. 1–34.
[6] C. C. Aggarwal, Recommender Systems: The Textbook, Springer, 2016.
[7] P. Resnick and H. R. Varian, "Recommender systems," Communications of the ACM,
vol. 40, no. 3, pp. 56–58, 1997.
[8] S. Zhang, L. Yao, A. Sun, and Y. Tay, "Deep learning based recommender system: A
survey and new perspectives," ACM Computing Surveys (CSUR), vol. 52, no. 1, pp. 1–
38, 2019.
[9] R. Burke, "Hybrid recommender systems: Survey and experiments," User Modeling
and User-Adapted Interaction, vol. 12, no. 4, pp. 331–370, 2002.
[10] M. Jahrer, A. Töscher, and R. Legenstein, "Combining predictions for accurate
recommender systems," in Proceedings of the 16th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining (KDD '10), pp. 693–702, 2010.