1

RECOMMENDER SYSTEMS

Jesse Davis
 Economics:
Traditional Retail vs. The Web
2

 Physical retail: Space is limited and expensive

 People are not willing to travel far for products
 Products stocked must make money

 Implication: Must focus on popular products

 Web: Storage space is cheap

 Sites cater to everyone
 Implication: Low cost and easy access mean
possible to offer more choice
Problem: Too many products, what interests me?

Solution: Systems that can recommend products

 Sales vs. Products: The Long Tail
3

Stores

Mixed
(e.g., Amazon)
Sales

Online only
(e.g., iTunes)

Products (e.g., songs, books, etc.)

 Making Recommendations
4

“Into Thin Air”

Books are similar

“Touching the Void”

The book “Into Thin Air” helped

make “Touching the Void” a bestseller
Wired article:
[Link]
 Recommendation Types
5

 Editorial
 List
of favorites
 Essential items

 Aggregates
 Top10 lists
 Most emailed articles
 Most recent posts

 Personalized user recommendations

 Amazon
 Movie sites
6 Key Challenges
 Key Challenges
7

 How do we get user feedback?

 How do we evaluate predictions?

 How do predict an unknown rating?

 Main interest: Highly rated products
 Options for Obtaining Ratings
8

 Explicit rating of products

 Bought items
 Items on “wish lists”
 Recently clicked product pages/links
 Length of time spent on product page
 Printed links
 Etc.
 Evaluation
9

Ratings
1 3 4
3 5 5
4 5 5
Users

3
3
2 ? ?
?
Most
2 1 ?
recent
3 ?
ratings
1
 Metrics: Compare with Known Rating
10

 Precision at top 10: % of those in top 10

𝑁
1
 Root-mean square error = 𝑁
෍(𝑃𝑖 − 𝑅𝑖 )2
𝑖=1

 Spearman’s rank correlation between

model’s and user’s complete rankings
6 σ𝑛𝑖=1 𝑈𝑖 − 𝑀𝑖 2
𝑅𝑠 =
𝑛(𝑛2 − 1)
 For 0/1 data
 Coverage: Number of items/users for which the
system can make a prediction
 Precision, ROC, etc.
 Weakness with Metrics
11

 Focusing on accuracy misses important points

 Prediction diversity
 Prediction context

 Order of prediction

 Only high ratings matter

 RMSE might penalize a method that does well for
high ratings and badly for others
 Recommending Products
12

 Idea 1: This is a machine learning problem

 Collect(a lot of) rating for the product per user
 Define a set of features (i.e., profile for item)

 Learn a model

 Idea 2: Content-based filtering

 Define a profile for an item
 Define a profile for a user

 Compare similarity between users and items

 Key Challenge: Building Item Profile
13

 Item profile: A way to describe each item

 These are usually hand-crafted
 For a movie:
 Actors

 Genre

 Director

 Etc.

 News articles:
 Words, title, author, etc.
 User Profiles and
Content-Based Prediction
14

 Main idea: Recommend items to customer x that

are similar to previous items rated highly by x

 User profile possibilities:

 Weighted average of rated item profiles
 Variation: weight by difference from average
rating for item

 Prediction heuristic: Given user profile x and

item profile i, estimate
𝒙·𝒊
𝑢(𝒙, 𝒊) = cos(𝒙, 𝒊) =
| 𝒙 |⋅| 𝒊 |
 Pros and Cons of
Content-Based Approaches
15

 Advantages
 Onlyneed data about one user
 More personalized approach

 More easily recommend new/unpopular items

 Can provide context for recommendation

 Disadvantages
 Must manually construct meaningful features
 Never recommends items outside of a
user’s content profile
 Hard to build a profile for a new user
16 Netflix and Recent Advances
 Present Recommendation in the
Context of Netflix Challenge
17

 While it is now older, it highlights a number of

important issues in data mining that are
applicable to many other problems

 Driven much of the advances in recommender

systems: We are still building off these ideas
 Netflix: The $1 Million Question
18

Given: A training set of 100 million rating

Do: Build a recommendation system that
improves the root-mean squared error by
10% over Netflix’s system
 Netflix Data and Evaluation
19

Each rating: (user, movie, rating, time stamp) tuple

Train set Test set
3M known to Netflix
100M ratings
• 99% sparse 1.5M for public
• 480,000 users leaderboard
• 17,700 movies
1.5M for
final winner

Score: Test set RMSE

 The Competition
20

 Register on competition site and download

anonymized data
 Submit predictions for test set at most one time
per day
 Winning algorithm: Software and non-exclusive
license to Netflix, must publish algorithm
 50k improvement prizes every year
 Once 10% threshold met: 30 day final
competition period
 Start October of 2006: More popular than
anticipated (eventually around 40,000 teams!)
 Pros and Cons of a Challenge?
21

 Advantages
 Expensive to develop internal system
 Publicity is good

 Awarding prize will pay for itself:

$1 million ≪ value of software to Netflix

 Disadvantages
 Privacy concerns (user backlash to data release)
 Prize won too quickly or no one wins the prize

 Time and effort to run the competition

 Results not useful in practice (e.g., too slow)

 What Worked Well:
Or General Data Analysis Advice
22

 Try the obvious thing first

 Predict the right thing

 Think outside the box

 Know your data

 The more (models), the merrier

23 Try Obvious Approaches First
 Classic Recommendation Approach:
Collaborative Filtering
24

 Intuition: Find users with similar tastes and

recommend products they liked

 Insight: Can do this just by looking at ratings!

 Big idea:
 Find other users whose ratings are similar to the
current user
 Propagate the (dis)likes to the current user
 Pictorial Overview
25

R1 R2 R3 R4 R5 R6
Alice 2 - 3 2 - 1
Bob 2 5 4 - - 2
Chris 4 3 - - - 5
Diana 3 - 2 4 - 5

Find Correlation R3 = 4

R1 R2 R3 R4 R5 R6
Eve 2 5 ? ? ? ?

Active user
 General Algorithm
26

 Step 1: Measure the similarity between

the user of interest and all other users

 Step 2 (Optional): Select a smaller subset

consisting of the most similar users

 Step 3: Predict ratings as a weighted

combination of “nearest neighbors”

 Step 4: Return the highest-rated items

 Step 1: Similarity Weighting
27

 Question: How can we compute the similarity

between two users?

𝑆𝑖 ∩ 𝑆𝑗
 Idea 1: Jaccard similarity: sim 𝑆𝑖 , 𝑆𝑗 =
𝑆𝑖 ∪ 𝑆𝑗

 Problem: Ignores the actual rating

𝑆𝑖 ⋅ 𝑆𝑗
 Idea 2: Cosine similarity: sim(Si, Sj) =
||𝑆𝑖 || ⋅ ||𝑆𝑗 ||
 Problem: Treats missing ratings as zero
 Step 1: Similarity Weighting
28

Pearson Correlation
σ𝑘 𝑅𝑖𝑘 − 𝑅ഥ𝑖 𝑅𝑗𝑘 − 𝑅ഥ𝑗
𝑊𝑖𝑗 =
2 2
σ𝑘 𝑅𝑖𝑘 − 𝑅ഥ𝑖 σ𝑘 𝑅𝑗𝑘 − 𝑅ഥ𝑗

σ𝑚 𝑅𝑖𝑒
𝑅ഥ𝑖 =
𝑒=1
(Average rating given by user i)
𝑚
Rik = User i’s rating on item k

Consider only items k rated by both users!!!

 What Does Pearson Correlation
Mean?
29

σ𝑘 𝑅𝑖𝑘 − 𝑅ഥ𝑖 𝑅𝑗𝑘 − 𝑅ഥ𝑗

𝑊𝑖𝑗 =
2 2
σ𝑘 𝑅𝑖𝑘 − 𝑅ഥ𝑖 σ𝑘 𝑅𝑗𝑘 − 𝑅ഥ𝑗
Positive if both predictions on same side of average

1 = perfect linear relationship with Rik increasing with Rjk

-1 = perfect linear relationship with Rik decreasing
when Rjk increases
 Step 2: Selecting Neighborhood
30

 Could use the whole database

 Could ignore “far away” users

 Pick “k” nearest users
 Include all users above a predetermined weight
threshold
 Note: Could use any k-NN strategy here
 Step 3: Predicting Ratings
31

 Predict a rating
 Account for different rating levels by looking at
difference from a user’s average rating
 Weight each user’s contribution by similarity

𝑃𝑖𝑘 = 𝑅ഥ𝑖 + 𝛼 ෍ 𝑊𝑖𝑗 𝑅𝑗𝑘 − 𝑅ഥ𝑗

𝑗
1
𝛼=
σ |𝑊𝑖𝑗 |
Wij = Pearson correlation
 Step 4: Return Items
32

 Usually only interested in highly rated items

 Do not want to overload the user

 There usually return the top 2, 5 or 10 items

 Can give user the option to view more items

 Practical Issues
33

 Potentially very large datasets

 Number of users: Millions
 Number of items: 100,000s

 Key issue: How to efficiently find similar users?

 Pairwise similarity for 1M users: ~5 days
 Pairwise similarity for 10M users: ~1 year!

 Big trick: Locality Sensitive Hashing

(See programming for big data)
 Pros and Cons of
Collaborative Filtering
34

 Advantages
 Simple and intuitive approach for any item type
 No feature construction and selection
 Exploits information about other users

 Disadvantages
 Data is sparse: Hard to find similar users
 Cold start: Need to enough users in database
 First rater: Can’t recommend unrated items,
e.g., new or unique items
 Popularity bias: Favors items that lots of people like
(i.e., bad if you have unique taste)
 Performance of Various Methods
Global average: 1.1296

User average: 1.0651

Movie average: 1.0533

Netflix: 0.9514
Basic Collaborative filtering: 0.94

Grand Prize: 0.8563

35
36 Predict the Right Thing
 Predict the “Right Thing”
37

 Our task
 Predict(user_id, movie_id, ?)
 Minimize RMSE of predictions

Question: What will produce low RMSE?

 Two points:
 Obvious: Better results if model optimized
towards the given objective
 Subtle: To get good results, often have to derive
a new target variable to predict
 What Affects a User’s Rating?
38

 Hypothesis: Multiple factors affect rating

 Goal: Isolate the portion of the rating that
captures the user-movie effect
 Two big effects:

User – User’s rating scale

Bias – Values of other ratings user gave recently
(user’s mood, anchoring, multi-user accounts)

Movie – (Recent) popularity of movie i

Bias – Selection bias (“frequency”)
 Capturing Global Effects:
Better Baseline
39

overall mean rating Rating deviation of movie i

𝒓𝒙𝒊 = 𝒃𝒙𝒊 = 𝝁 + 𝒃𝒙 + 𝒃𝒊

rating deviation of user x

 𝝁: Average rating of all movies in data

 𝒃𝒙 : Average rating of user x − 𝝁
 𝒃𝒊 : Average rating of movie i − 𝝁
 Example of Baseline
40

 Mean movie rating: 3.7 stars

 The Sixth Sense is 0.5 stars above average

 Joe rates 0.2 stars below average

 Baseline estimation:
 3.7+ 0.5 + (-0.2) = 4
 Joe will rate The Sixth Sense 4 stars
 Three Problems with the
Collaborative Filtering Model
41

 Similarities are arbitrary: Many choices and

not tied to our ultimate objective
 Do not account for interactions between
neighbors: May “double count’’
 Weighted average: Ignores differences among
high & low confidence neighbors 3-NN, with sims:
0.1, 0.05, and 0.15
0.1(2) + 0.05 (3) + 0.15(2)
2+ = 3.27
(0.1 + 0.05 + 0.15)
3-NN, with sims:
0.6(2) + 0.3 (3) + 0.9(2) 0.6, 0.3, and 0.75
2+ = 3.27
(0.6 + 0.3 + 0.9)
 Idea: Weighted Sum
42

 Weighted sum rather than weighted avg.:

𝑟ෞ
𝑥𝑖 = 𝑏𝑥𝑖 + ෍ 𝑤𝑖𝑗 𝑟𝑥𝑗 − 𝑏𝑥𝑗
𝑗∈𝑁(𝑖;𝑥)
 A few notes:
 𝒃𝒙𝒊 is the new, more complex baseline with biases
 𝑵 𝒊; 𝒙 : movies rated by user x that are similar to
movie i
 𝒘𝒊𝒋 is the interpolation weight (some real number)

 𝒘𝒊𝒋 models the interaction between pairs of movies

(it does not depend on user x)
 Picking Weights: Optimization
43

 Competition objective: Sum of squared errors

𝟐
෍ 𝒓ො 𝒙𝒊 − 𝒓𝒙𝒊
(𝒊,𝒙)∈𝑹

 Idea: Pick weights to minimize this objective!

2
𝐽 𝑤 =෍ 𝑏𝑥𝑖 + ෍ 𝑤𝑖𝑗 𝑟𝑥𝑗 − 𝑏𝑥𝑗 − 𝑟𝑥𝑖
𝑥,𝑖 𝑗∈𝑁 𝑖;𝑥
True
Posed as Predicted rating
rating
optimization
problem!
 Performance of Various Methods
Global average: 1.1296

User average: 1.0651

Movie average: 1.0533

Netflix: 0.9514
Basic Collaborative filtering: 0.94

CF+Biases+learned weights: 0.91

Grand Prize: 0.8563

44
45 Think Outside the Box
 Let’s Think about the Task
46

Movies  Goal: Fill in “missing entries”

1 3 ? 5
5 4 ? ?
Users

 Likely some shared hidden

2 4 1 2 structure for users and movies
2 4 5
4 3 4 2  Implication: Could represent the
1 3 3 ? data using “fewer” dimension
 Latent Factor Models
Drama
Sense and Amadeus
Sensibility

Braveheart
Serious Factor 1Light
Dumb and hearted
Dumber
Lethal Weapon
Factor 2

47

Syrianna
Ocean’s 11
Action
 Latent Factor Model
48

Movies
= R ≈ Q x PT
1 3 ? 5 Movies “Topics’’ Movies
5 4 ? ?

“Topics’’
Users

Users

Users
2 4
2 4
1 2
5 ͌ X
4 3 4 2
R=uxm Q=uxd P=mxd
1 3 3 ?
matrix matrix matrix
Key idea: d ≪ m, u
“Topics”: shared hidden structure
(e.g., how much each users likes each genre)
 Latent Factor Models
49

Movies factors factors

1 3 5 5 4 .1 -.4 .2 1.1 -.8 2.1
5 4 ? 4 2 1 -.2 .7 -.4
Users

-.5 .6 .5
2 4 1 2 3 4 3 5 .3 .5 .6

Users
≈

Items
-.2 .3 .5
2 4 5 4 2 1.1 2.1 .3 .5 1.4 1.7
4 3 4 2 2 -.7 2.1 -2 -2 .3 2.4
1 3 3 2 4 -1 .7 .3 -.5 -1 .9
.8 1.4 -.3
? = -.5 * -2 +.6 * .3 + .5 * 2.4 = 2.38 -.4 2.9 .4
.3 -.7 .8
Note: Item-factor should be 1.4 1.2 .7
transposed, but not due to space 2.4 -.1 -.6
 Optimization problem
50

Goal: Find Matrices P and Q according to:

2
min ෍ 𝑅𝑖𝑗 − 𝑄𝑃T 𝑖𝑗
P,Q
𝑖𝑗∈𝑅

Three things to think about

 R has many missing entries

 Objective is non-convex

 R is sparse, so worried about overfitting

 Intuition of a Solution
51

1 3 5 ? ? ? ? ? ? ? ?
5 4 ? ? X ? ? ? ? ? ?
2 4 1 2 ? ?
2 4 5
4 3 4 2
͌ ? ?
? ?
1 3 3 ? ?
 Intuition of a Solution
52

1 3 5 ? ? 0.1 0.1 0.5 1.2 0.1 1.5

5 4 ? ? X -1.4 0.8 -0.5 0.4 -0.3 0.9
2 4 1 2 ? ?
2 4 5
4 3 4 2
͌ ? ?
? ? Fit least
X1 X2 Y squares
1 3 3 ? ?
0.1 -1.4 2 solution
0.1 0.8 4
1.2 0.4 1
0.1 -0.3 2
 Idea: Alternating Least Squares
53

Sum over non-missing entries

2
2 2
min ෍ 𝑅𝑖𝑗 − 𝑄𝑃T 𝑖𝑗
+ λ1 𝑄 𝐹 + λ2 𝑃 𝐹
P,Q
𝑖𝑗∈𝑅
Frobenius norm = Matrix equivalent of L2
Input: R, d
Randomly initialize Q,P
for i = 0, 1,… do Solve m d-dimensional
ridge regression problems
Fix Q: Optimize P
Fix P: Optimize Q Solve u d-dimensional
ridge regression problems
 Latent Factor Models with Biases
54

users factors users

factors
1 3 5 5 4 .1 -.4 .2 1.1 -.2 .3 .5 -2 -.5 .8 -.4 .3 1.4 2.4
5 4 4 2 1 -.5 .6 .5 -.8 .7 .5 1.4 .3 -1 1.4 2.9 -.7 1.2 -.1
items

2 4 1 2 3 4 3 5 2.1 -.4 .6 1.7 2.4 .9 -.3 .4 .8 .7 -.6

≈
-.2 .3 .5

2 4 5 4 2 1.1 2.1 .3
PT
items
4 3 4 2 2 -.7 2.1 -2
1 3 3 2 4 -1 .7 .3

R Q
𝒓ො 𝒙𝒊 = 𝒒𝒊 ⋅ 𝒑𝒙 = ෍ 𝒒𝒊𝒇 ⋅ 𝒑𝒙𝒇
𝒇
𝑟𝑥𝑖 = 𝜇 + 𝑏𝑥 + 𝑏𝑖 + 𝑞𝑖 ⋅ 𝑝𝑥 T

Overall Bias for Bias for User-Movie

mean rating user x movie i interaction
 Koren, Bell, Volinksy, IEEE Computer, 2009
55
 Notes On Latent Factors
56

 Regularization is key to avoiding overfitting

 Often want to preprocess or normalize the data

(Note: Must convert back to original scale when
making a prediction)

 Alternating least squares is easy to parallelize

 Can also solve via stochastic gradient descent

 Non-Negative Matrix Factorization
57

 Often want the entries of P, Q to be

non-negative: Easier to interpret and visualize

 Easy solution: Projected gradient methods

∂l(w)
 Normal gradient wi+1 ⟵ wi - η
∂wi
 Projected gradient wi+1 ⟵ max(0, wi – η ∂l(w) )
∂wi
 Performance of Various Methods
Global average: 1.1296

User average: 1.0651

Movie average: 1.0533

Netflix: 0.9514
Basic Collaborative filtering: 0.94

CF+Biases+learned weights: 0.91

Latent factors: 0.90
Latent factors+Biases: 0.89

Grand Prize: 0.8563

58
59 Know Your Data
 Temporal Effect: Early 2004 Sudden
Jump in Average Movie Rating
60

◼ Pre 2004: 3.4 stars

◼ Post 2004: >3.6 stars

 Netflix improved matching, leading to higher

rankings?
 People biased towards higher ratings: Meaning
of rating changes?
 Temporal Effect: Movie Age
61

◼ High initial ratings

◼ Ratings increase with

movie age

 Older movies rated by users better matched to

them?
 Older movies are inherently better than newer
movies?
 Temporal Biases Of Users
[Koren, KDD’09]
62

 Original model: rxi = m +bx + bi + qi ·px

 Add time dependence to biases:
rxi = m +bx(t)+ bi(t) +qi · px
 Make parameters bx and bi depend on time
 (1) Parameterize time-dependence by linear trends
(2) Each bin corresponds to 10 consecutive weeks

 Add temporal dependence to factors

 px(t)… user preference vector on day t
 Performance of Various Methods
Global average: 1.1296

User average: 1.0651

Movie average: 1.0533

Netflix: 0.9514
Basic Collaborative filtering: 0.94

CF+Biases+learned weights: 0.91

Latent factors: 0.90
Latent factors+Biases: 0.89

Latent factors+Biases+Time: 0.876

Grand Prize: 0.8563

63
64 The More (Models) the Merrier
 What Now??
65

 Tried lots of things, but still have not reached the

10% threshold
 Getting desperate…
 Idea: “Kitchen Sink Approach”
 Build lots and lots of predictors
◼ Classifiers
◼ Collaborative filters
◼ Ensembles

 Come up with clever ways to blend the results

66
 2009
67

 At the end of June, the leading team

(BellKorPragmaticChaos) submits results that
exceed the 10% threshold
 Competition enters 30 day final period
 New “Ensemble” team formed based on
collaboration for others near at top of
leaderboard and quickly beat 10% threshold too
 The race is on…
 Standing on June 26th 2009
68

June 26th submission triggers 30-day “last call”

 The Final Countdown
69

 Direct competition between two teams

 Can only submit results once per day so each
team can submit once on final day
 A day before deadline, BellKorPragmaticChaos
notice that Ensemble is in the lead
 Each team prepares final results
 BellKorPragmaticChaos submits 40 mins early
 Ensemble submits 20 mins early

 And they wait…

70
 September 2009: Prize Awarded
71

 Teams tied BellKorPragmaticChaos wins by

submitting earlier
72 Conclusions and Perspectives
 Perspectives
73

 Must account for unpredictable users behavior

 Think about what you really want to predict and

whether you need a new target

 Knowing your data is crucial

 When in doubt, combine lots of models

 Privacy and ethics are tricky but important

 Summary
74

 Recommender systems an important and active

area of research and use
 Data is really messy
 Paradigms
 Content: Based on designing features and
applying classification/regression algos
 Collaborative: Based on comparing users/items
(aka nearest neighbor)
 Latent factor approaches

 Baselines important in practice

 Questions
75

 A number of slides are based on the [Link]

lectures on recommender systems
[Link]