06 - Collaborative Filtering and Matrix Factorization

Question 1

What are the problems of Collaborative Filtering?

Answer 1

• You need many ratings for every user and every movie
• Just a very simple view: Movies, user, ratings and nothing more behind

Question 2

What criteria can we use to decide if a user likes a movie or not?

Answer 2

• Users have certain preferences
• Movies fulfil these preferences to a certain extent
• Movies have general features such as genre, actors, director, budget, running time, etc.

Question 3

How could one calculate the user-movie-preference?

Answer 3

The user-movie-preference could be seen as a linear combination of the scores of the individual features

Question 4

What are the movie features and user preferences ultimately in matrix factorization in relation to collaborative filtering?

Answer 4

• The features are not real, they are so-called hypothetical “latent features” or “learned features”
• The number of these features is a hyperparameter

Question 5

Could one create the matrices with the movie features or the user preferences be created manually?

Answer 5

• Movie features with a lot of effort possibly possible but rather not
• User preferences completely impossible

Question 6

How can you automatically create matrices with movie features and user preferences?

Answer 6

• With (stochastic) gradient descent
• Initialise both matrices with random small values
• Update each entry in both matrices with the following formula until a certain criterion (runtime, threshold) is met

Question 7

What is factorization?

Answer 7

Divide a complex product into smaller, and often simpler, factors

Question 8

What is Matrix Factoristion?

Answer 8

A complex matrix is divided into two smaller, simpler matrices

Question 9

What is an advantage of factorized matrices?

Answer 9

They need less memory

Question 10

What do you do with missing entries in matrix factorization in relation to collaborative filtering?

Answer 10

• One possibility is to fill them with zeros or -1, but then you would waste a lot of time optimizing the algorithm to predict zeros
• Therefore: Optimise only for known ratings

Question 11

What are the options for evaluating whether the algorithm is good and which of them is the better choice?

Answer 11

• Option 1: Remove complete movies and their ratings from the matrix and put them into a separate test set
• Option 2: Assign random entries from the matrix to the test set
• Option 2 is the better one. With option 1 you have the cold start problem, how can you make a prediction for a movie for which you had no information in the training process?