Coms 573 Machine Learning Project How much someone is going to love a movie

1
Coms 573 Machine Learning ProjectHow much
someone is going to love a movie

• Hailin Tang

2
Contents

• Goal of the project
• Background
• Experiment design and results
• Discussion

3
Goal of the project

• The goal of this project is to predict whether
  someone will love a movie based on how much they
  liked or disliked other movies by predicting a
  rating on a scale of 1-5 for a given movie
• The ultimate goal is to help Netflix improve
  their current movie recommendation
  system-Cinematch

Movies related to a movie being recommended at
Netflix
4
Background

• Several challenges related to the data set of
  Netflix
• The huge data set (Uncompressed full dataset is
  about 2 Gigabytes)
• More than 100 million ratings (training set, on a
  scale of 1-5 stars)
• Over 480 thousand users
• Nearly 18 thousand movie titles
• The distribution of movies per user is skewed
• median number of ratings per user is 93
• 10 user rated 16 or fewer movies
• Two users rated as many as 17000 movies

5
Background

• The ratings per movie are also skewed
• 25 of the movies had 190 or fewer ratings
  associated with them
• 5 of the movies are rated fewer than 10 times
• The lack of user and movie attributes
• Nearly 99 percent of user-item entries being zero
  on user-item dataset
• Simply focused on the quality of recommendation
  system in terms of minimizing the error between
  the user rating and the predicted rating

6
(No Transcript)
7
Experiment Designs

• Simple Method Averaging methods
• To predict the ratings of a movie for a user
  based on average statistics
• Assign average rate to all movies for all users
• Calculate the average rate of rated movies for
  each user, assign it to the unrated movie
• Calculate averages over all users for any
  specific type of movies, assign it to the unrated
  movies

8
Experiment Designs

• Current wining solution was a linear combination
  of 107 sets of predictions
• KNN K-nearest neighbor
• RBMs Restricted Boltzmann Machines
• LSI Latent Semantic Indexing
• SVD (Factor models) singular value decomposition
• Conclusion
• A simple linear combination of three models
  K-NN,LSI and RBM could also get fairly good
  result
• Need a principled approach to optimizing the
  solution

9
Experiment Designs

• Maximum Margin Matrix Factorization for
  collaborative ranking
• Idea is take advantages of the collaborative
  effects rating patterns from other users are
  used to estimate rating for the current user
• It works without feature extraction
• Can handle very large, sparse and very imbalanced
  datasets and deal with users who have very few
  ratings
• Software package Cofirank

10
Experiment Designs

• Feature extraction and Traditional Classifiers
• Feature Extraction
• User related features
• Movie related features
• User-movie pair features
• Classifers
• K-nearest neighbor
• Logistic Regression
• Naïve Bayes Classifier

11
Feature extraction
User related features

• Number of historic user ratings
• Percentage of 1-star ratings of the user
• Percentage of 5-star ratings of the user
• Average rating of the user
• Standard deviation of the user ratings

12
Feature extraction
Movie related features

• number of historic ratings received by the movie
• percentage of 1-star ratings received by the
  movie
• percentage of 5-star ratings received by the
  movie
• average rating received by the movie
• standard deviation of the ratings received by
  movie

13
Feature extraction
User-movie pair features

• To include information on user and/or movies that
  behave similarly
• Identify groups of users which might have similar
  behaviors and also groups of movies that were
  rated in similar ways

14
Test Results
Grand Prize RMSElt0.8563, 10
improvement Current leading RMSE 0.8596, 9.65
improvement logistic regression provides the most
accurate results
15
Discussion

• The key to winning is to build models that can
  predict accurately where theres sufficient data,
  without over-fitting in the absence of adequate
  data
• Two main classes of algorithms can be applied
  into this project
• Content-based analysis, similar movie is
  recommended by computing the similarity in term
  vectors between various movies
• Collaborative analysis, each movie is treated as
  black box, and user interactions with the movie
  are used to compute the similarities

16
Thank You!