Announcements

• Homeworks:
• HW2 (due extended: 11/08)
• Due to maintenance at Haedong Lounge (11/01 18:00 – 11/04 09:00).
Recommender Systems 2 • Enjoy the Netflix challenge!
• HW3 (will be posted on 11/06)
EE412: Foundation of Big Data Analytics
• Midterm:
Fall 2024
• Claim finished; thank you for your hard work!
• Classes:
• No in-person class on 11/04; video will be uploaded.

Jaemin Yoo 1 Jaemin Yoo 2

Recommend
Recap Outline
Popularity
“Touching the void” “Into Thin Air”
• Recommender Systems 1. UV Decomposition
• Content-based Recommendation 2. UV Decomposition: Computation
• Collaborative Filtering Items 3. UV Decomposition: Variants
• The Netflix Challenge

like
like like

similar

recommend
like recommend

Jaemin Yoo 3 Jaemin Yoo 4

Goal of Recommender Systems Latent Factor Models
• Recommendation is to fill in the blank in the utility matrix 𝑅. • We’ll learn latent factor models, which assume:
• The core operation is how to get user and item representations. • There are latent factors that can represent users and items well.
• Content-based approach creates user/item profiles. • Such latent factors can be extracted from the utility matrix.
• Collaborative filtering uses the rows and columns of 𝑅. • Many people consider latent factor models as a part of CF.
• Since they share the same philosophy.
HP1 HP2 HP3 TW SW1 SW2 SW3 • CF uses the rows and columns of 𝑅 without modification.
A 4 5 1 • Latent factor models extract (better) latent factors from 𝑅.
B 5 5 4
C 2 4 5
D 3 3

Jaemin Yoo 5 Jaemin Yoo 6

Latent Factor Models UV Decomposition

• Idea: Consider a utility matrix as the product of factor matrices. • Given an 𝑚 × 𝑛 utility matrix 𝑅 (i.e., 𝑚 users and 𝑛 items).
• Latent factors are underlying concepts/topics; same as in SVD. • Find an 𝑚 × 𝑘 matrix 𝑈 and 𝑛 × 𝑘 matrix 𝑉 such that:
• E.g., users react to certain genres, famous actors, or directors. • 𝑈𝑉 # closely approximates 𝑀 for the non-blank entries.
• UV decomposition decomposes a utility matrix into 𝑈 and 𝑉. • Use the elements of 𝑈𝑉 ! to estimate the blank entries in 𝑅.
• Each user and movie is summarized as a low-dimensional vector. ̂ = 𝑢$#𝑣% to predict 𝑟$% .
• Compute 𝑟$%
n k n

✖ 𝑉! k
m R ≈ U

R U 𝑉!

Jaemin Yoo 7 Jaemin Yoo 8

Error Function Recap: Singular Value Decomposition
• Root-mean-square error (RMSE) measures the difference. • Decomposition of any matrix into a product of three matrices.
• Let 𝐸 be the set of non-blank entries. • Choose any number 𝑟 of intermediate concepts (= latent factors).
• RMSE 𝑅, / 𝑅 = sqrt & ∑ $,% ∈' 𝑟$% ̂ − 𝑟$% * . • In a way that minimizes the reconstruction error.
'
• The error is zero when 𝑟 ≥ rank 𝑀 .
• RMSE is equivalent to the sum of squared errors (SSE).
/ 𝑅 = 𝐸 ⋅ RMSE* 𝑅,
• SSE 𝑅, / 𝑅 . n r r n

VT r
Σ
RMSE( )= m =
M U

R 𝑈𝑉 !

Jaemin Yoo 9 Jaemin Yoo 10

UV Decomposition vs. SVD Outline

• Note: 𝑈 and ΣV ! from SVD are also factor matrices. 1. UV Decomposition
• Differences of UV decomposition from SVD: 2. UV Decomposition: Computation
• UV ignores the missing entries, not treating them as zero. 3. UV Decomposition: Variants
• UV minimizes RMSE on only the training portion.
• Larger 𝑘 is not necessarily better.
• Larger 𝑘 does not guarantee decreasing RMSE on the test data.
• 𝑈 and 𝑉 need not to be orthonormal matrices.

Jaemin Yoo 11 Jaemin Yoo 12

Objective Function Mismatch in the Objective Function
• Goal: Find matrices 𝑈 and 𝑉 such that: • Our (true) goal is to minimize SSE for unseen test data.
• However, our objective function only considers training data.
𝑈 ∗ , 𝑉 ∗ = argmin$,& 𝐽 𝑅, 𝑈, 𝑉
• Increasing 𝑘 always decreases our objective function.
# * • Since the error decreases with more factors.
• 𝐽 𝑅, 𝑈, 𝑉 = ∑$,%∈' 𝑟$% − 𝑢$ 𝑣% measures SSE on training data 𝐸.
• However, SSE on test data can begin to rise with large 𝑘.
n k n

✖ 𝑉! k
• We call 𝑈, 𝑉 parameters.
m R ≈ U • Other choices are hyperparameters.

Jaemin Yoo 13 Jaemin Yoo 14

Overfitting Regularization
• This is a classical example of overfitting: • Regularization is a possible way to prevent overfitting:
• Model starts fitting noise with too many free parameters. • Allows a rich model when there is sufficient data.
• Model is not generalizing well to unseen test data. • Shrinks the model aggressively where data is scarce.
• We should carefully control the model complexity. • The new objective function with regularization is
• E.g., the number of clusters in 𝑘-means.
𝐽 ⋅ = 5 𝑟'( − 𝑢' 𝑣( +
+ 𝜆, 5 𝑢' +
+ + 𝜆+ 5 𝑣( +
+
',(∈* ' (

• 𝜆& and 𝜆* are hyperparameters. Limit the lengths of 𝑢! and 𝑣"

Source: Medium

Jaemin Yoo 15 Jaemin Yoo 16

 The Color
serious
Amadeus
Braveheart Validation Data
Purple

• Introducing validation data is also important.

Lethal • Split the training data into (new) training and validation data.
Sense and Weapon
Sensibility Ocean’s 11 Geared
• Check the performance on validation data during training.
Geared
towards towards • Use the validation performance as a proxy of the test performance.
Factor 1
females males

Validation dataset
Effect of regularization: The Lion King
Dumb and Training dataset
• Goes to the center unless Dumber

Factor 2
the signal is really strong. Independence
Day Test dataset

funny

Jaemin Yoo 17 Jaemin Yoo 18

Gradient Descent Gradient Descent for UV Decomposition

• Gradient descent (GD) aims to find an input 𝑥 to minimize 𝑓 𝑥 . • How to use GD to find 𝑈 ∗ and 𝑉 ∗ for UV decomposition:
• Compute the derivative ∇𝑓. • Step 1: Initialize 𝑈 and 𝑉 using SVD, treating missing entries as 0.
• Step 2: Update 𝑈 and 𝑉 to minimize the objective function 𝐽 ⋅ .
• Start at some point 𝑦 and evaluate ∇𝑓 𝑦 .
• Make a step in the reverse direction of the gradient: 𝑦 ← 𝑦 − ∇𝑓 𝑦 .
• 𝑈 ← 𝑈 − 𝜂 ⋅ ∇2 𝐽 ⋅ .
• Repeat until 𝑓 is sufficiently small.
𝑓 • 𝑉 ← 𝑉 − 𝜂 ⋅ ∇3 𝐽 ⋅ .
• 𝜂 is a hyperparameter (called a step size or a learning rate).
𝑓 𝑦 + 𝛻𝑓(𝑦)

Jaemin Yoo 19 Jaemin Yoo 20

Gradient Descent for UV Decomposition Stochastic Gradient Descent
• Perform the update step on every entry independently. • Observation: The gradient can be decomposed as
• Since 𝑈 and 𝑉 are matrices. • ∇𝐽 𝑅, 𝑈, 𝑉 = ∑%: $,% ∈' −2𝑣%6 𝑟$% − 𝑢$#𝑣% + 2𝜆&𝑢$6
• For each entry at row 𝑥, column 𝑐 of matrix 𝑈, we update: = ∑$,%∈' ∇𝐽 𝑟$% , 𝑈, 𝑉
• Stochastic gradient descent (SGD):
𝑢'/ = 𝑢'/ − 𝜂 ⋅ ∇0!" 𝐽 𝑈, 𝑉 • Evaluate the gradient on each (not all) rating and make a step.
• Needs more steps until convergence but each step is much faster.
where ∇4!" 𝐽 𝑈, 𝑉 = ∑%: $,% ∈' −2𝑣%6 𝑟$% − 𝑢$#𝑣% + 2𝜆&𝑢$6 . • GD: 𝑈 ← 𝑈 − 𝜂 ⋅ ∑% ∇𝐽 𝑟$% .
• SGD: 𝑈 ← 𝑈 − 𝜂 ⋅ ∇𝐽 𝑟$% .

Jaemin Yoo 21 Jaemin Yoo 22

Convergence of SGD vs. GD SGD for UV Decomposition

• GD improves the value of the objective function at every step. • How to use SGD to find 𝑈 ∗ and 𝑉 ∗ for UV decomposition:
• SGD improves the value but in a “noisy” way. • Step 1: Initialize 𝑈 and 𝑉 using SVD, treating missing ratings as 0.
• GD takes fewer steps to converge but each step takes much longer. • Step 2: Iterate over the ratings and update factors until convergence:

for each 𝑟$% ∈ 𝐸:

Objective function

𝜇& and 𝜇*: Learning rates.

𝜖$% ← 2 𝑟$% − 𝑢$#𝑣%
𝑢$ ← 𝑢$ + 𝜇& 𝜖$% 𝑣% − 𝜆&𝑢$
𝑣% ← 𝑣% + 𝜇* 𝜖$% 𝑢$ − 𝜆*𝑣%

Iteration/step
Jaemin Yoo 23 Jaemin Yoo 24
23
SGD with Mini-batches Outline
• In practice, people do not apply SGD for individual samples. 1. UV Decomposition
• Instead, they create (mini-)batches of several samples. 2. UV Decomposition: Computation
• GD: 1 step using 𝑁 samples. 3. UV Decomposition: Variants
• (True) SGD: 𝑁 steps using 1 sample.
• Batch SGD: 𝑁/𝐵 steps using a batch of 𝐵 samples.
• Makes a better balance between speed and stability in training.
• 𝐵 is a hyperparameter to choose.

Jaemin Yoo 25 Jaemin Yoo 26

Modeling Biases and Interactions Model with Biases

• There are global effects (biases) of users and movies. • Let’s predict a rating as 𝑟'(
̂ = 𝜇 + 𝑏' + 𝑏( + 𝑢'! 𝑣( .
• Rating 𝑟$% is not only about the interaction between 𝑥 and 𝑖. • 𝜇 is the overall mean rating.
• E.g., a critical reviewer 𝑥& vs. a generous person 𝑥*. • 𝑏$ and 𝑏% are biases for user 𝑥 and movie 𝑖, respectively.
• E.g., The Godfather 𝑖& vs. some bad movie 𝑖* on Netflix. • 𝑢$#𝑣% is their interaction modeled by factor matrices.
• Example:
user bias movie bias user-movie interaction
• Mean rating of training data is 𝜇 = 3.7.
• You are a critical reviewer: 𝑏$ = −1.
• Star Wars is favored by many people: 𝑏% = +0.5.
• Final score is 3.7 − 1 + 0.5 + 𝑢$#𝑣% .

Jaemin Yoo 27 Jaemin Yoo 28

Fitting the New Model Further Improvements
• Update the parameters {𝑈, 𝑉, 𝐵1234 , 𝐵5637 } minimizing 𝐽′: • Any idea for modifying the three components is okay to try:
• Set 𝜃 of learnable parameters.
𝐽 ⋅ = 5 𝑟'( − 𝜇 + 𝑏' + 𝑏( + 𝑢'! 𝑣( + regularizer
8 + • Objective function 𝐽 to minimize.
• (Optional) regularizer on 𝜃 with 𝜆.
',(∈*
• SGD (or GD) will take care of finding the optimal parameters.
• There are 4 regularization hyperparameters: 𝜆&, 𝜆*, 𝜆7, 𝜆8. • We believe in the power of gradient-based optimization!
• 𝜇 is the simple average of ratings; it needs not to be learned.

Jaemin Yoo 29 Jaemin Yoo 30

Hyperparameter Search Dealing with Implicit Feedback

• How can we search for optimal hyperparameters 𝜆, , 𝜆+ , 𝜆9 , 𝜆: ? • What if the utility matrix 𝑅 contains only implicit feedback?
• We pick the one with the best validation performance. • Each entry is either 0 (not watched) or 1 (watched).
• Grid search: Create sets of values and try for each combination. • Using the same model with RMSE 𝐽 leads to some limitations.
• E.g., 𝜆& in 0.001, 0.01, 0.1, 1, 10 . • Limitation 1: Our prediction can be any number, i.e., 𝑢$#𝑣% > 1.
• Limitation 2: We may have a better gradient curve of 𝐽.
• Random search: Try for each randomly selected combination. • Limitation 3: Our loss 𝐽 assumes 0 as a dislike, not “not watched.”
• E.g., 𝜆& from 0.001, 10 (but in a log scale).

Jaemin Yoo 31 Jaemin Yoo 32

Idea 1: Sigmoid Function Implication of the Sigmoid
• We want to limit the predictions of our model to be 0, 1 . • Without sigmoid: Signals are mixed regardless of 𝑟'( .
• Simple solution is to use 𝜎 𝑢'! 𝑣( instead of 𝑢'! 𝑣( as output. • If 𝑟$% = 1 and 𝑢$#𝑣% > 1, the model is updated to decrease 𝑢$#𝑣% .
= 1 and 𝑢$#𝑣% < 1, the model is updated to increase 𝑢$#𝑣% .
• Sigmoid function 𝜎 is defined as follows:
• If 𝑟$%
• If 𝑟$% = 0 and 𝑢$#𝑣% > 0, the model is updated to decrease 𝑢$#𝑣% .
1 • If 𝑟$% = 0 and 𝑢$#𝑣% < 0, the model is updated to increase 𝑢$#𝑣% .
𝜎 𝑥 = • With sigmoid: Signals are consistent with 𝑟'( .
1 + 𝑒 ;'
• If 𝑟$% = 1, the model is always updated to increase 𝑢$#𝑣% .
• Maps −∞, ∞ to 0, 1 with 𝜎 0 = 0.5. • If 𝑟$% = 0, the model is always updated to decrease 𝑢$#𝑣% .
• Monotonically increasing for all ranges of 𝑥.

Jaemin Yoo 33 Jaemin Yoo 34

Idea 2: Binary Cross Entropy Gradient Curves

• RMSE is mainly designed for continuous, unbounded targets. • BCE has a gradient curve different from that of RMSE.
• We may use binary cross entropy (BCE) instead of RMSE: • Suppose that 𝑟'( = 0 and 𝑎 = 𝜎 𝑢'! 𝑣( for simplicity.
• (RMSE) ∇9 𝑎 * = 2𝑎: Gradient increases linearly with 𝑎.
𝐽<=> 𝑢' , 𝑣( , 𝑟'( = − 𝑟'( log 𝜎 𝑢'! 𝑣( + 1 − 𝑟'( log(1 − 𝜎 𝑢'! 𝑣( ) • (BCE) ∇9 − log 1 − 𝑎 = 1/ 1 − 𝑎 : Gradient increases rapidly.

• If 𝑟$% = 1, we minimize − log 𝜎 𝑢$#𝑣% by making 𝜎 𝑢$#𝑣% to 1. • BCE makes the model focus more on wrong samples.
• If 𝑟$% = 0, we minimize − log(1 − 𝜎 𝑢$#𝑣% ) by making 𝜎 𝑢$#𝑣% to 0. • Not already accurate ones.

Jaemin Yoo 35 Jaemin Yoo 36

Idea 3: Ranking Losses Bayesian Personalized Ranking
• Both RMSE and BCE assume 0 entries as “dislike,” not “unknown.” • We may use the Bayesian personalized ranking (BPR) loss:
• Since they are computed for individual elements 𝑟$% .
• Idea: Let’s consider the task as ranking, not elementwise prediction. 𝐽<BC 𝑈, 𝑉 = 5 − log 𝜎 𝑢'! 𝑣( − 𝑢'! 𝑣D
• Given a user 𝑥, suppose that 𝑟$% = 1 while 𝑟$: = 0. ',(,D
• Then, we can safely assume that user 𝑥 likes 𝑖 more than 𝑗.
• If 𝑥 really likes 𝑗, they would have watched it before 𝑖. • Item 𝑗 is randomly selected from the negative samples { 𝑗 | 𝑟$: = 0 }.
• Let’s update the model by comparing items, so that 𝑢$% > 𝑢$: . • In BPR, we don’t have to compare 𝜎(𝑢'! 𝑣( ) and 𝜎(𝑢'! 𝑣D ).
• Since 𝜎 is a monotonically increasing function.

Jaemin Yoo 37 Jaemin Yoo 38

Summary
1. UV Decomposition
• Latent factor models
2. UV Decomposition: Computation
• Overfitting and regularization
• Stochastic gradient descent
3. UV Decomposition: Variants
• Modeling biases
• Dealing with implicit feedback
• BCE and BPR losses

Jaemin Yoo 39