Lecture 16.2 — Recommender Systems Intro
\n00:00 - How important recommendation system to the real world?
\n01:30 - Why should we study recommendation system? (Did or do academics pay much attention to it? Can it teach us the power of learning features of ML?)
\n03:00 - Andrew presented the recommendation system problem with a simplest example
Lecture 16.2 — Recommender Systems | Content Based Recommendations
\n00:00 - recap of the problem of recommendation system
\n00:23 - What is content-based recommendation system? How to describe the recommendation system problem in details with symbols (mathematically)?
\nHow To predict the rate of any user to any movie, we just need to find out the features of the user and the features of the movie; we need to use the existing ratings of movies by current users to find out the features of current users and current movies
\n06:12 - How to learn the features of a single user by designing a simple loss function?
\n10:06 - How to learn the features of all users by combining all of the losses based on the loss function above?
\n11:25 - How to calculate the derivaties or doing gradient descent on user features?
\n13:37 - Why called content based? using features of movies (contents) to find out features of users and then to make predictions on ratings
Lecture 16.3 — Recommender Systems | Collaborative filtering Recommendations Intro
\n00:00 - What's special about collaborative filtering compared with content-based recommendation system? It can learn all the features itself without giving features of movies first like content-based approach
\n00:21 - How to present collaborative filtering in terms of knowing user features and using rating data to find out features of movies and then make predictions on ratings (well, it is a just another content based approach using user features instead of movie features)
\n04:39 - What is the simplest loss function to help do gradient descent to find out features of all movies
\n06:56 - Why called collaborative filtering? what are collaborating? (between two content based approaches) why called filtering? What ensures such collaboration can learning possible? (data is good: a movie is rated by multiple users, a user has rated multiple movies)
\n09:15 - Andrew Ng explained why called collaboration filtering
Lecture 16.4 — Recommender Systems | Collaborative Filtering Algorithm
\n00:00 - How to combine the loss function of both content-based approaches to do gradient descent on both features of users and movies, instead of finding features of users and then movies sequentially?
\n04:40 - Andrew present the collaborative algorithm in steps
Lecture 16.5 — Recommender Systems | Vectorization Low Rank Matrix Factorization
\n00:00 - What is low-rank matrix factorization? put all features of all movies into a matrix, all features of all users into a matrix and put all rating prediction as a dot product of two matricies
\n05:07 - How can collaborative filtering enable recommendations of similar movies to a user? As it finds out all the features of all movies, we can compare the similarity of movie feature vectors to find out similar movies
Lecture 16.6 — Recommender Systems | Implementational Detail Mean Normalization
\n00:00 - How a new user without giving any rating create challenges on learning based on the loss of our recommendation system algo learnt above
\n03:23 - How mean normalization on all movies' ratings on the rows (movie on row, user on col) come to rescue? Why it makes sense? (assume the new user will give the average ratings to all movies)
\n07:29 - What about having movies without any user giving ratings? What problem does taking mean normalization on ratings on col? (mean normalization on a user's ratings of all movies does not make much sense )
Binary labels: favs, likes and clicks
\n00:00 - What is the binary label problem application in recsys?
\n01:50 - How to define label 1 and label 0?
\n04:04 - How to move from regression to binary prediction in recsys?
\n05:16 - How to modify loss function for binary labels?
Finding related items
\n00:00 - how to find related items using features of movies or products
\n03:24 - what are the limitations of collaborative filter? (cold start problem and not-using side information problem)
Collaborative filtering vs Content-based filtering
\n00:00 - How collaboratibe filtering differ from content based filtering? (content-based filtering use side info on both users and items)
\n01:57 - What do the features or side informations of users or items look like? What's the problem to solve in order to utilize the features?
\n05:44 - How does the features of users and items can help make predictions of ratings? The key is to convert the side info of users and items into the actual features (same length) of users and items in making predictions
\n09:10 - summary
Content-based Filtering | Deep learning for content-based filtering
\n00:00 - How deep learning (two neuralnets) help to convert side info into features (same length) of users and items?
\n03:07 - How to do the above with a single neuralnet?
\n06:02 - How to find similar items
\n08:09 - Sum up: deep learning can easily merge two neuralnets together; how to in practice make the computation of large amount of side info of users and items into actual features fast
9.10 Advanced implementation | Recommending from a large catalogue
\n00:00 - Why it is a challenge to make predictions when there are millions of users and items?
\n01:17 - How to make use of features of users and items to recommend for existing user daily? Retrieval of related items for the user
\n04:03 - How to use the model to predict on the retrieved items in order to speed up or ease the computation? Rank by using the neuralnet model
\n05:26 - How to decide how many items to select in retrieval step?
\n07:09 - sum up
9.11 Advanced implementation | Ethical use of recommender systems
\n00:00 - goals of recommendation system? what are the suspicous ones?
\n04:28 - Ad examples
\n07:51 - other problematic cases
06:00 - lecture outline
\n08:09 - Why recommendation system is so important? information overload, commercial value of recommendation
\n13:32 - what is the \"Recommendation problem\"?
\n14:35 - what is the \"Recommendation problem\" formally?
\n15:48 - The two-step: offline and online process
\n17:53 - approaches to recommendations
\n21:23 - recommendation as data mining
\n22:33 - Serendipity: unsought findings
\n25:00 - a safe bet on what works: collaborative filtering + data preprocessing + matrix factorying
\n27:05 - Ingredients of Collaborative filtering: implicit vs explicit data - rating data is very noisy, contradiction
\n30:34 - Other ingredients
\n31:04 - Collaborative filtering in a sentence
\n32:05 - The basic steps of doing CF and other approaches
\n32:57 - The Pros and Cons of CF
\n36:06 - Personalized vs Non-Personalized CF (what is your base line when you don't know the active user to predict at hand)
\n45:18 - User-based collaborative filtering
\n52:22 - User-based collaborative filtering example
\n54:30 - Challenges to User-based CF (intro to latent models)
\n56:43 - Item-based CF
\n58:01 - Item-based CF example
\n59:48 - neighborhood based CF problems: Sparsity
\n1:03:39 - Scalability problem
\n1:05:21 - Model-based CF
\n1:09:08 - SVD/MF
\n1:12:36 - SVD for rating prediction
\n1:17:54 - What is RBM for Recommendation
\n1:21:30 - Hinton's paper on RBM for recommendation and How it put together in production
\n1:29:01 - How to do clustering for recommendation and its pros and cons
\n1:33:09 - LSH (locality sensitivity Hashing)
\n1:34:48 - Association Rule mining
\n1: 39:22 - classifiers
\n1:43:31 - limitations of CF: cold start and popularity bias
\n1:45:50 - cold start problem
\n1:47:28 - content-based recommenders
\n1:50:45 - CF advantages and disadvantages
\n1:52:27 - Content-based methods: TF-IDF, content-based User profile, and a paper recommended (a word of caution)
\nLecture 16.2 — Recommender Systems Intro
\n00:00 - How important recommendation system to the real world?
\n01:30 - Why should we study recommendation system? (Did or do academics pay much attention to it? Can it teach us the power of learning features of ML?)
\n03:00 - Andrew presented the recommendation system problem with a simplest example
Lecture 16.2 — Recommender Systems | Content Based Recommendations
\n00:00 - recap of the problem of recommendation system
\n00:23 - What is content-based recommendation system? How to describe the recommendation system problem in details with symbols (mathematically)?
\nHow To predict the rate of any user to any movie, we just need to find out the features of the user and the features of the movie; we need to use the existing ratings of movies by current users to find out the features of current users and current movies
\n06:12 - How to learn the features of a single user by designing a simple loss function?
\n10:06 - How to learn the features of all users by combining all of the losses based on the loss function above?
\n11:25 - How to calculate the derivaties or doing gradient descent on user features?
\n13:37 - Why called content based? using features of movies (contents) to find out features of users and then to make predictions on ratings
Lecture 16.3 — Recommender Systems | Collaborative filtering Recommendations Intro
\n00:00 - What's special about collaborative filtering compared with content-based recommendation system? It can learn all the features itself without giving features of movies first like content-based approach
\n00:21 - How to present collaborative filtering in terms of knowing user features and using rating data to find out features of movies and then make predictions on ratings (well, it is a just another content based approach using user features instead of movie features)
\n04:39 - What is the simplest loss function to help do gradient descent to find out features of all movies
\n06:56 - Why called collaborative filtering? what are collaborating? (between two content based approaches) why called filtering? What ensures such collaboration can learning possible? (data is good: a movie is rated by multiple users, a user has rated multiple movies)
\n09:15 - Andrew Ng explained why called collaboration filtering
Lecture 16.4 — Recommender Systems | Collaborative Filtering Algorithm
\n00:00 - How to combine the loss function of both content-based approaches to do gradient descent on both features of users and movies, instead of finding features of users and then movies sequentially?
\n04:40 - Andrew present the collaborative algorithm in steps
Lecture 16.5 — Recommender Systems | Vectorization Low Rank Matrix Factorization
\n00:00 - What is low-rank matrix factorization? put all features of all movies into a matrix, all features of all users into a matrix and put all rating prediction as a dot product of two matricies
\n05:07 - How can collaborative filtering enable recommendations of similar movies to a user? As it finds out all the features of all movies, we can compare the similarity of movie feature vectors to find out similar movies
Lecture 16.6 — Recommender Systems | Implementational Detail Mean Normalization
\n00:00 - How a new user without giving any rating create challenges on learning based on the loss of our recommendation system algo learnt above
\n03:23 - How mean normalization on all movies' ratings on the rows (movie on row, user on col) come to rescue? Why it makes sense? (assume the new user will give the average ratings to all movies)
\n07:29 - What about having movies without any user giving ratings? What problem does taking mean normalization on ratings on col? (mean normalization on a user's ratings of all movies does not make much sense )
Binary labels: favs, likes and clicks
\n00:00 - What is the binary label problem application in recsys?
\n01:50 - How to define label 1 and label 0?
\n04:04 - How to move from regression to binary prediction in recsys?
\n05:16 - How to modify loss function for binary labels?
Finding related items
\n00:00 - how to find related items using features of movies or products
\n03:24 - what are the limitations of collaborative filter? (cold start problem and not-using side information problem)
Collaborative filtering vs Content-based filtering
\n00:00 - How collaboratibe filtering differ from content based filtering? (content-based filtering use side info on both users and items)
\n01:57 - What do the features or side informations of users or items look like? What's the problem to solve in order to utilize the features?
\n05:44 - How does the features of users and items can help make predictions of ratings? The key is to convert the side info of users and items into the actual features (same length) of users and items in making predictions
\n09:10 - summary
Content-based Filtering | Deep learning for content-based filtering
\n00:00 - How deep learning (two neuralnets) help to convert side info into features (same length) of users and items?
\n03:07 - How to do the above with a single neuralnet?
\n06:02 - How to find similar items
\n08:09 - Sum up: deep learning can easily merge two neuralnets together; how to in practice make the computation of large amount of side info of users and items into actual features fast
9.10 Advanced implementation | Recommending from a large catalogue
\n00:00 - Why it is a challenge to make predictions when there are millions of users and items?
\n01:17 - How to make use of features of users and items to recommend for existing user daily? Retrieval of related items for the user
\n04:03 - How to use the model to predict on the retrieved items in order to speed up or ease the computation? Rank by using the neuralnet model
\n05:26 - How to decide how many items to select in retrieval step?
\n07:09 - sum up
9.11 Advanced implementation | Ethical use of recommender systems
\n00:00 - goals of recommendation system? what are the suspicous ones?
\n04:28 - Ad examples
\n07:51 - other problematic cases
06:00 - lecture outline
\n08:09 - Why recommendation system is so important? information overload, commercial value of recommendation
\n13:32 - what is the \"Recommendation problem\"?
\n14:35 - what is the \"Recommendation problem\" formally?
\n15:48 - The two-step: offline and online process
\n17:53 - approaches to recommendations
\n21:23 - recommendation as data mining
\n22:33 - Serendipity: unsought findings
\n25:00 - a safe bet on what works: collaborative filtering + data preprocessing + matrix factorying
\n27:05 - Ingredients of Collaborative filtering: implicit vs explicit data - rating data is very noisy, contradiction
\n30:34 - Other ingredients
\n31:04 - Collaborative filtering in a sentence
\n32:05 - The basic steps of doing CF and other approaches
\n32:57 - The Pros and Cons of CF
\n36:06 - Personalized vs Non-Personalized CF (what is your base line when you don't know the active user to predict at hand)
\n45:18 - User-based collaborative filtering
\n52:22 - User-based collaborative filtering example
\n54:30 - Challenges to User-based CF (intro to latent models)
\n56:43 - Item-based CF
\n58:01 - Item-based CF example
\n59:48 - neighborhood based CF problems: Sparsity
\n1:03:39 - Scalability problem
\n1:05:21 - Model-based CF
\n1:09:08 - SVD/MF
\n1:12:36 - SVD for rating prediction
\n1:17:54 - What is RBM for Recommendation
\n1:21:30 - Hinton's paper on RBM for recommendation and How it put together in production
\n1:29:01 - How to do clustering for recommendation and its pros and cons
\n1:33:09 - LSH (locality sensitivity Hashing)
\n1:34:48 - Association Rule mining
\n1: 39:22 - classifiers
\n1:43:31 - limitations of CF: cold start and popularity bias
\n1:45:50 - cold start problem
\n1:47:28 - content-based recommenders
\n1:50:45 - CF advantages and disadvantages
\n1:52:27 - Content-based methods: TF-IDF, content-based User profile, and a paper recommended (a word of caution)
\n