Announcements

• Homeworks:
• HW2 (due: 11/06)
• HW3 (will be posted on 11/06; due: 11/27)
Recommender Systems 1 • Midterm:
EE412: Foundation of Big Data Analytics • Claim: 10/28 (Mon) – 10/29 (Tue), 19:00 – 20:00
Fall 2024 • Location: N1 113
• Classum:
• Please review the board before asking questions.
• Aim to ask questions that could benefit other students as well.

Jaemin Yoo 1 Jaemin Yoo 2

Recap Outline
• Singular Value Decomposition (SVD) 1. Recommender Systems
• Dimensionality Reduction with SVD 2. Content-based Recommendation
• CUR Decomposition 3. Collaborative Filtering
4. The Netflix Challenge
n r r n

Σ VT r
m M = U

Jaemin Yoo 3 Jaemin Yoo 4

Recommender Systems Personalized Recommender Systems
• Class of applications that predict user responses to options. • Two groups of recommender systems:
• Non-personalized recommendations: • Content-based: Focus on the profiles (or features) of users/items.
• Editorial and hand-curated: List of favorites, essential items, … • Collaborative filtering: Focus on the interactions between users/items.
• Simple aggregates: Top 10, most popular, recent uploads, …
• Personalized recommendations: like
like like
• Tailored to individual users: Amazon, Netflix, YouTube, …
similar

recommend
like recommend

Jaemin Yoo 5 Jaemin Yoo 6

From Scarcity to Abundance The Long Tail

• Physical delivery systems have scarcity of resources. • The distinction is called the long tail phenomenon.
• Stores have limited shelf space to show to users. • Online institutions provide the entire range of items.
• Not possible to tailor the store to individual customers. • The tail as well as the head (popular) items
• Force them to focus on personalized recommendation.
• Online stores can make anything available to users.
• Amazon offers millions of books. Recommend

Popularity
“Touching the void” “Into Thin Air”

Items
Jaemin Yoo 7 Jaemin Yoo 8
Utility Matrix Gathering Ratings
• We consider two classes of entities: Users and items. • Explicit feedback: Ask users to rate items.
• Utility matrix shows the preference of users for items. • E.g., Youtube asks for likes/dislikes of watched videos.
• The values come from an ordered set, e.g., 1-5 stars. • Users are generally unwilling to provide responses.
• Assumed to be sparse, i.e., most entries are unknown. • Biased as it comes from people willing to provide ratings.
• Implicit feedback: Learn ratings from user actions.
HP1 HP2 HP3 TW SW1 SW2 SW3 • If a user watches a movie, the user is said to “like” it.
A 4 5 1
• Hard to model low ratings: 0 (no rating) or 1 (like).
B 5 5 4
C 2 4 5
D 3 3

Jaemin Yoo 9 Jaemin Yoo 10

Outline Content-based Recommendation

1. Recommender Systems • Main idea: Use the profiles of items.
2. Content-based Recommendation • E.g., for a movie, { genre, director, actors, plot, release year }.
• Recommend items similar to previous highly-rated items.
3. Collaborative Filtering
• Examples:
4. The Netflix Challenge • Recommend movies with same actor(s), director, genre, …
• Recommend websites or blogs with “similar” content.

Jaemin Yoo 11 Jaemin Yoo 12

Plan of Action Item Profiles
• For each item, create an item profile as a set of features.
• Finding features is not straightforward for some classes of items.
• E.g., find a set of “important” words from a document.
1. Eliminate stop words.
2. Compute the TF-IDF scores of words.
3. Use the top 𝑘 words as its features. Common stop words The, and because

Less frequent terms car, drive

with small TF-IDF
More frequent terms auto repair
with higher TF-IDF
auto repair
Source: Stanford CS246 (2022)

Jaemin Yoo 13 Jaemin Yoo 14

Item Profiles as Vectors Example: Item Profiles

• If a feature is a discrete value, use a one-hot vector. • We can also combine discrete and numerical features.
• Length 𝐿 is the number of possible discrete values. • E.g., set of actors, director, genre, and average rating.
• If it is a set of values, e.g., actors, it can be a multi-hot vector.
• If a feature is numerical, use the exact value. Movie profiles: Actors Director Genre Avg. rating

• Its scale might have to be adjusted considering other features.

• E.g., average ratings for movies. M1: 0 1 1 0 1 1 0 1 0 1 0 0 1 0 2.5
M2: 1 1 0 1 0 1 1 0 0 0 1 1 0 0 4.1

E.g., Julia Roberts E.g., Science Fiction

Jaemin Yoo 15 Jaemin Yoo 16

User Profiles User Profiles
• We need a user profile that summarizes the history of a user. • If the matrix is not binary, weight item profiles by utility values.
• Take some aggregation of item profiles using the utility matrix. • E.g., say user 𝑈 rates 𝑀1 with rating 1 and 𝑀2 with rating 4:
• If the utility matrix is binary, the simple average is natural.
• Suppose that user 𝑈 watched movies 𝑀1 and 𝑀2 : Movie M1’s profile: 0 1 1 0 1 1 0 0 Weight = 1

Movie M2’s profile: 1 1 0 1 0 1 1 0 Weight = 4

Movie M1’s profile: 0 1 1 0 1 1 0 0
U’s profile: 0.5 1 0.5 0.5 0.5 1 0.5 0
Movie M2’s profile: 1 1 0 1 0 1 1 0
U’s weighted profile: 2 2.5 0.5 2 0.5 2.5 2 0
U’s profile: 0.5 1 0.5 0.5 0.5 1 0.5 0

Jaemin Yoo 17 Jaemin Yoo 18

Prediction Heuristic Pros and Cons: Content-based Approach

• Compute the cosine distance between user and item profiles. • Pros:
• For efficiency, one can apply LSH to quickly find candidate items. • No need for data on other users.
• Recall the random hyperplane approach. • Able to recommend items to users with unique tastes.
• Able to recommend new or unpopular items.
• Able to provide explanations (by listing content features).
U’s profile: 0.5 1 -1 0.5 -1 1 0.5 -1.5 Movie to
recommend U’s profile • Cons:
Movie to recommend: 0 1 0 0 0 1 0 0 • Finding the appropriate features may be difficult.
Movie not to • New users may not have a profile.
Movie not to recommend: 0 0 1 0 1 0 0 0 recommend • Overspecialization: Cannot recommend beyond user’s profile.

Jaemin Yoo 19 Jaemin Yoo 20

Outline Collaborative Filtering
1. Recommender Systems • Create user and item profiles solely from the utility matrix.
2. Content-based Recommendation • E.g., an item profile is the set of users who purchased it.
= The profile of item 𝑖 is the 𝑖-th column of the matrix.
3. Collaborative Filtering • Don’t care about the content of items.
4. The Netflix Challenge • Comes in two flavors:
• User-user collaborative filtering like like
• Item-item collaborative filtering

like recommend

Jaemin Yoo 21 Jaemin Yoo 22

User-User Collaborative Filtering Finding Similar Users

• Given user 𝑈, find users whose ratings are similar to 𝑈’s ratings. • There are various ways for defining which users are “similar.”
• Estimate 𝑈’s ratings based on similar users. • Jaccard similarity:
• Treat ratings as sets, ignoring the values (i.e., likes vs. dislikes).
• For this example, it seems intuitively wrong.
• Since Jaccard 𝐴, 𝐵 = 1/5 < Jaccard 𝐴, 𝐶 = 2/4. Does it make sense?
HP1 HP2 HP3 TW SW1 SW2 SW3

A 4 5 1
B 5 5 4
C 2 4 5
D 3 3

Source: Stanford CS246 (2022)

23 Jaemin Yoo 24
Finding Similar Users Finding Similar Users
• There are various ways for defining which users are “similar.” • There are various ways for defining which users are “similar.”
• Cosine similarity: • Person’s correlation coefficient:
• Treat ratings as points (or vectors), considering blanks as 0. • Consider only the ratings on items rated by both users.
• Questionable, since no rating doesn’t mean dislike. σ𝑠∈𝑆 𝑟𝐴,𝑠 −𝑟ҧ𝐴 𝑟𝐵,𝑠 −𝑟ҧ𝐵
• In this example, 𝑣𝐴⊤ 𝑣𝐵 /‖𝑣𝐴 ‖‖𝑣𝐵 ‖ = 0.380 and 𝑣𝐴⊤ 𝑣𝐶 /‖𝑣𝐴 ‖‖𝑣𝐶 ‖ = 0.322. • sim 𝐴, 𝐵 =
2 2
σ𝑠∈𝑆 𝑟𝐴,𝑠 −𝑟ҧ𝐴 σ𝑠∈𝑆 𝑟𝐵,𝑠 −𝑟ҧ𝐵
HP1 HP2 HP3 TW SW1 SW2 SW3
• 𝑆: Items rated by both users 𝐴 and 𝐵
A 4 5 1 • 𝑟𝐴ҧ and 𝑟𝐵ҧ : Average ratings of users 𝐴 and 𝐵, resp.
B 5 5 4 • Limitation: The size of 𝑆 should be large enough.
C 2 4 5
D 3 3

Jaemin Yoo 25 Jaemin Yoo 26

Improving Similarity Metrics Improving Similarity Metrics

• How to improve Jaccard similarity: Rounding the data. • How to improve cosine similarity: Normalize ratings.
• Idea: Distinguish between high and low ratings. • Subtract from each rating the average rating of that person.
• Consider only the high ratings 3, 4, 5 as “1”. • Then, we have sim 𝐴, 𝐵 = 0.092 > sim(𝐴, 𝐶) = −0.559.
• Then, Jaccard 𝐴, 𝐵 = 1/4 = Jaccard 𝐴, 𝐶 = 0. • Notice that 𝐷’s ratings are not worth taking seriously.
• Better than the previous result. • Does it make sense? What if the two ratings are both 5?

HP1 HP2 HP3 TW SW1 SW2 SW3 HP1 HP2 HP3 TW SW1 SW2 SW3

A 1 1 A 2/3 5/3 -7/3

B 1 1 1 B 1/3 1/3 -2/3
C 1 1 C -5/3 1/3 4/3
D 1 1 D 0 0

Jaemin Yoo 27 Jaemin Yoo 28

Rating Predictions Item-Item Collaborative Filtering
• From similarity metrics to recommendations: • Another view: Item-item collaborative filtering.
• Let 𝑟𝑥 be the vector of user 𝑥’s ratings. • For item 𝑖, find other similar items.
• Let 𝑁 be the set of 𝑘 users most similar to user 𝑥. • Estimate rating for item 𝑖 based on ratings for similar items.
• Let 𝑁 ′ ⊆ 𝑁 be the subset of users who rated item 𝑖. • Can use the same similarity metrics as in the user-user model.
• Prediction for item 𝑖 of user 𝑥: σ𝑗∈𝑁 sim 𝑖, 𝑗 ⋅ 𝑟𝑥𝑗
𝑖;𝑥
• Simple version: 𝑟𝑥𝑖 =
1
σ ′𝑟 .
𝑟𝑥𝑖 =
𝑘′ 𝑦∈𝑁 𝑦𝑖 σ𝑗∈𝑁 𝑖;𝑥 sim 𝑖, 𝑗
σ𝑦∈𝑁′ sim 𝑥,𝑦 ⋅ 𝑟𝑦𝑖
• Complex version: 𝑟𝑥𝑖 = σ
.
𝑦∈𝑁′
sim 𝑥,𝑦 • 𝑁 𝑖; 𝑥 is the set of items similar to item 𝑗 and rated by user 𝑥.

Jaemin Yoo 29 Jaemin Yoo 30

Item-Item vs. User-User Pros and Cons: Collaborative Filtering

• Item-item similarity is often more reliable. • Pros:
• Intuitively, items are classifiable in simple terms, e.g., one genre. • Do not have to come up with features (or profiles).
• Users may like multiple genres, so harder to compute similarity. • Cons:
• However, there is no clear advantage of one from another. • Need enough users in the system to find a match.
• E.g., user-user is better for relatively new items. • Cannot recommend new or unpopular items that have not been rated.
• Cannot recommend items to someone with unique taste.
HP1 HP2 HP3 TW SW1 SW2 SW3
• I.e., tends to recommend popular items.
A 4 5 1
B 5 5 4
C 2 4 5
D 3 3

Jaemin Yoo 31 Jaemin Yoo 32

Outline The Netflix Challenge (2006)
1. Recommender Systems • Netflix offered $1M to beat its movie recommendation algorithm.
2. Content-based Recommendation • The prize was awarded after three years of work.
3. Collaborative Filtering
4. The Netflix Challenge

Jaemin Yoo 33 Jaemin Yoo 34

Description of the Challenge Utility Matrix from Netflix

• Training data: • Goal: Minimize RMSE on the test data.
• 100M 1-5 ratings from 480K users and on 18K movies. • RMSE = sqrt
1
σ 𝑟𝑥𝑖 − 𝑟𝑥𝑖
Ƹ 2
where 𝑇 is the test set.
𝑥,𝑖 ∈𝑇
• Collected for 6 years (2000 to 2005). 𝑇

• Test data: 17K movies

• Set of the last few ratings of each user (2.8M in total).

• Evaluation criterion: Root mean square error (RMSE). Training dataset
• Netflix’s system RMSE: 0.9514.
0.5M users
• Competition: 2,700+ teams.
Test dataset

Jaemin Yoo 35 Jaemin Yoo 36

Training and Test Data Key Takeaways
• Test data: Notable difference from the previous topics. • Not easy to outperform very simple approaches.
• Goal of data mining is often to find patterns in given data. • Naïve approach: Predict the rating 𝑟𝑢𝑖 as 𝜇𝑢 + 𝜇𝑖 /2.
• Clustering, frequent itemsets, dimensionality reduction. • 𝜇𝑢 = The average rating of 𝑢.
• The Netflix challenge is about generalizing to unseen data. • 𝜇𝑖 = The average rating for 𝑖 by all users.
• Most machine learning problems are this case. • CineMatch was only 3% better than the naïve approach.
• Focus on performance, rather than scalability or running time. • CineMatch: Netflix’s previous solution.

Jaemin Yoo 37 Jaemin Yoo 38

Key Takeaways Key Takeaways

• Additional data, other than the utility matrix, can help or not. • The winning solution was a combination of several algorithms.
• Movie information: • These algorithms are developed independently.
• Used IMDB to get information of actors, directors, and genres. • The second team also used a blend of independent algorithms.
• Not effective, maybe because they didn’t give useful information. • Combining different algorithms is effective in general.
• Time information: • Called an ensemble approach.
• Decreasing vs. increasing reputation over time. • Cons: Difficult interpretation, extensive computation, etc.
• Examined the ratings of a movie to see its rating trend in time.

Jaemin Yoo 39 Jaemin Yoo 40

Summary
1. Recommender Systems
• The long tail phenomenon
• Utility matrix
2. Content-based Recommendation
• User / item profiles
3. Collaborative Filtering
• Defining “similarity”
4. The Netflix Challenge
• Key takeaways

Jaemin Yoo 41