Cloud Computing for Data Analysis: Movie Recommendation System

Authors:

• Aditya Gupta ([email protected])
• Rekhansh Panchal ([email protected])

Group 11

---

Overview:

• The project aims to analyze and extract insights from the Netflix data using the concepts of Cloud Computing.
• The goal of the project is to implement Pearson Correlation Coefficient & Alternating Least Squares algorithms with the help of PySpark.
• Movie Recommendations is implemented using Collaborative Filtering using pySpark on Netflix Data.
• This project’s primary aim is to provide movie recommendations to the user based on their preferences.

---

Tasks Involved and Steps Implemented:

• Configuring Jupyter Notebook and Spark
• Understanding the problem statement
• Understanding the algorithm
• Fetching the data
• Data cleaning
• Implementing PCS, ALS, and ALS with Library on Local Machine.
• Deploying the code and data on Amazon Web Services.
• Output generation
• Project Report

---

Motivation:

• Most of the online shopping is due to the personalised recommendations to users, reminding them about an item.
• It not only shows user interest but also helps the user to keep a price track of the items.
• This handy feature urged us to learn the technique of recommendation and algorithms behind it.
• Movie recommendation system is not something out-of-box project, infact, it has been already implemented by people. However, we considered this, more of a learning project and went with the movie recommendation option.

---

What is Collaborative Filtering:

Collaborative Filtering is a method of making automatic predictions (filtering) about the interests of a user by collecting preferences or taste information from similar users.

---

Data:

The original movie rating files contain over 100 million ratings from 480 thousand randomly-chosen, anonymous Netflix customers over 17 thousand movie titles.

The data were collected between October, 1998 and December, 2005 and reflect the distribution of all ratings received during this period. The ratings are on a scale from 1 to 5 (integral) stars.

However, use have worked on a part of the complete data for the project.

1. Number of Users: 750
2. Number of Movies: 1,000
3. Number of Ratings: 4,20,000

Input Ratings Data File contains (Before Cleaning): movie_id, user_id, ratings, date_of_rating

Input Movie Title file cotains: movie_id, year_of_release, movie_title

---

Data Link:

Movie Rating Files

---

Algorithm:

1. Pearson Correlation Coefficient

• We achieved movie recommendation results by calculating the Pearson Correlation Coefficient and thereby similarity between users based on the movies they watched and gave similar ratings.
• The coefficient value range from -1 to 1. Where -1 and 1 indicate a negative and positive correlation respectively.
• Coefficient with value 0 indicates no correlation between the two variables.
• Statistically it can be said that Pearson Correlation Coefficient between two variables can be calculated as the covariance of the variables divided by the product of their standard deviations.

2. Alternative Least Squares

• When we make the user-item matrix, we decompose into an lower dimensional matrix of user factors and item factors using Matrix Factorization.
• These lower dimension matrices are used to estimate the ratings by minimizing the cost function.
• Over multiple iterations, at the convergence point by reducing the Root Mean Square Error, ratings are predicted and displayed as results.

---

External Tools

• Amazon S3 for storage of data and program.
• Amazon EC2 (Spark 2.2.0) for running the program on cluster.
• Git for tracking the code changes.
• GitHub for hosting the website.

---

Framework:

• Apache Spark

Expectations/Aspects

1. What to expect?

• One can expect the implementation of both the algorithms and a proper documentation of outcomes of this project, which is the movie recommendations for users.

2. Likely to accomplish

• Result comparison and Performance Evaluation with respect to existing implementation of algorithms and project implementation.
• Documentation and online-publishing of the codebase.

3. Ideal Accomplishments.

• Suggested modifications/changes in the existing or project implementation.
• Creating a easy to use library that one can use for analysis purpose.

---

Tools Used:

• Jupyter Notebook
• pySpark
• Git and GitHub
• Amazon S3 and EC2

---

README

• Add all program files to the Amazon Storage S3 along with the reduced dataset of ratings.

• Create a Spark Cluster on Amazon EC2 and get the details of the cluster to use it on Terminal.

• Confirm Connection to the Cluster with obtained keypair.

ssh -i ~/keypair.pem -ND 8157 [email protected]

• Start Cluster Access

ssh -i keypair.pem [email protected]

• Import Pandas on the cluster:

sudo pip install pandas

• Run PCSalgorithm on input data consisting 4,20,000 ratings stored on S3 Storage. To recommend movies for user 1199825.

spark-submit s3://itcs6190/PCSalgorithm.py s3://itcs6190/movie_input_ratings.txt s3://itcs6190/movie_titles.csv 1199825

• Run ALS on input data consisting 4,20,000 ratings stored on S3 Storage. To recommend movies for user 1199825.

spark-submit s3://itcs6190/ALS.py s3://itcs6190/movie_input_ratings.txt s3://itcs6190/movie_titles.csv 1199825

• Run ALS using library for recommend movies for all users

spark-submit s3://itcs6190/ALSUsingLibrary.py s3://itcs6190/movie_input_ratings.txt

---

Outputs for User ID: 1488844 & Results

The programs to recommend were ran on Amazon EC2 Spark cluster. And satisfactory recommendations were obtained using 3 methods.

• Pearson Correlation Coefficient implementation.

• ALS implementation.

• ALS from ml Recommendation.

Root Mean Square Error: 2.015

---

Conclusion:

• Created a User - User based recommendation system using ALS and Pearson Correlation Coefficient techniques.

• Displayed top movies recommended a user by taking userId as input.

---

Future Scope:

• Recommend movies to new user and predict ratings for the same.
• Creating a easy to use library that one can use for analysis purpose.

---

Code Snippet:

• Data Cleaning

• Pearson Calculation

---

Challenges Faced:

• We started with implementing Singular Value Decomposition technique, but couldn't achieve anything potential with that due to multiple missing rating entries. Thus, we implemented ALS and ALS using ML library.
• Had no prior experience on implementing the code on PySpark, so had a lot of minor issues while handling the data.
• The data available, is huge to be considered, hence we had to limit it down to a lower scale.
• Hadoop DSBA Cluster was non-funcitonal during our project timeline.

---

Work Division:

The complete project has been accomplished together with inputs from both the team members.

Number	Task	Contribution
1	Pearson Correlation	Aditya
2	ALS Implementation	Rekhansh
3	ALS Using Library	Rekhansh
4	Cluter and Deployment	Aditya & Rekhansh
5	Project Report	Aditya & Rekhansh

---

References:

• Recommender Systems: Youtube

• Pearson Correlation

• ALS: Large-scale Parallel Collaborative Filtering for the Netflix

• Spark Documentation