Recommending

Products
Emily Fox & Carlos Guestrin
Machine Learning Specialization
University of Washington
2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
Where we see
recommender systems

2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Personalization is transforming
our experience of the world

Information overload

100 Hours a Minute Browsing is history
What do I care about? -Need new ways
to discover content

Personalization: Connects users & items

viewers videos

3 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Movie recommendations

Connect users with movies

they may want to watch

4 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Product recommendations

Recommendations combine
global & session interests

5 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Music recommendations

Recommendations form
coherent & diverse sequence
6 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
 Friend recommendations

Users and items

are of the same type
7 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
 Drug-target interactions
Cobanoglu et al. 13

What drug should we

repurpose for some disease?
8 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
Building a recommender system

2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

Solution 0: Popularity

2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Simplest approach: Popularity
What are people
viewing now?
-Rank by global
popularity

Limitation:
-No personalization

11 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

Solution 1: Classification model

2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Whats the probability Ill buy this product?
User info
Yes!
Purchase history

Product info
Classifier
No
Other info

Pros:
- Personalized:
Considers user info & purchase history
- Features can capture context:
Time of the day, what I just saw,
- Even handles limited user history: Age of user,
13 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
 Limitations of classification approach
User info
Yes!
Purchase history

Product info
Classifier
No
Other info

Features may not be available

Often doesnt perform as well as

collaborative filtering methods (next)

14 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

Solution 2: People who bought this
also bought

2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Co-occurrence matrix
People who bought diapers also
bought baby wipes

Matrix C:
store # users who bought both items i & j
- (# items x # items) matrix

- Symmetric: # purchasing i & j same as # for j & i (Cij = Cji)

16 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
Making recommendations using
co-occurences
User purchased diapers

1. Look at diapers row of matrix

2. Recommend other items with largest counts

- baby wipes, milk, baby food,

17 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Co-occurrence matrix must be
normalized
What if there are very popular items?
-Popular baby item:
Pampers Swaddlers diapers
-For any baby item (e.g., i=Sophie girae )
large count Cij for j=Pampers Swaddlers

Result:
-Drowns out other eects
-Recommend based on popularity
18 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
 Normalize co-occurrences:
Similarity matrix
Jaccard similarity: normalizes by popularity
- Who purchased i and j divided by
who purchased i or j

Many other similarity metrics possible, e.g., cosine similarity

19 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Limitations
Only current page matters, no history
- Recommend similar items to the one you bought

What if you purchased many items?

- Want recommendations based on purchase history

20 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 (Weighted) Average of purchased items
User bought items {diapers, milk}
- Compute user-specific score for each item j
in inventory by combining similarities:

Score( , baby wipes) =

(Sbaby wipes, diapers + Sbaby wipes, milk)

- Could also weight recent purchases more

Sort Score( , j ) and find item j with highest similarity

21 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Limitations
Does not utilize:
-context (e.g., time of day)
-user features (e.g., age)
-product features (e.g., baby vs. electronics)
Cold start problem
-What if a new user or product arrives?

22 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

Solution 3: Discovering hidden structure
by matrix factorization

2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Movie recommendation
Users watch movies and rate them
User Movie Rating

Each user only watches a few of the available movies

24 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
 Matrix completion problem
Xij known for black cells
Xij unknown for white cells

Users
Rating
X= Rows index movies
Columns index users
Movies

Data: Users score some movies

Rating(u,v) known for black cells
Rating(u,v) unknown for white cells

Goal: Filling missing data?

25 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Suppose we had d topics for
each user and movie
Describe movie v with topics Rv
- How much is it action, romance, drama,

Describe user u with topics Lu

- How much she likes action, romance, drama,

Rating(u,v) is the product of the two vectors

Recommendations: sort movies user hasnt watched by Rating(u,v)

26 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
 Predictions in matrix form
Xij known for black cells
Xij unknown for white cells
Rating
X= Rows index movies
Columns index users

R
X L
=

But we dont
know topics of
users and movies
27 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
 Matrix factorization model:
Discovering topics from data
Xij known for black cells
Xij unknown for white cells
Rating
X= Rows index movies
Columns index users

Parameters Many ecient algorithms
of model for factorization
Only use observed values to estimate topic vectors Lu and Rv
Use estimated Lu and Rv for recommendations
28 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
 Limitations of matrix factorization
Cold-start problem
- This model still cannot handle a new user or movie

Xij known for black cells

Xij unknown for white cells
Rating
X= Rows index movies
Columns index users

29 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

Bringing it all together:
Featurized matrix factorization

2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Combining features and discovered topics
Features capture context
- Time of day, what I just saw, user info, past purchases,
Discovered topics from matrix factorization capture
groups of users who behave similarly
- Women from Seattle who teach and have a baby

Combine to mitigate cold-start problem

- Ratings for a new user from features only
- As more information about user is discovered,
matrix factorization topics become more relevant

31 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Blending models
Squeezing last bit of accuracy by
blending models
Netflix Prize 2006-2009
-100M ratings
-17,770 movies
-480,189 users
- Predict 3 million
ratings to
highest accuracy
-Winning team blended over 100 models
32 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
A performance metric for
recommender systems

2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 The world of all baby products

34 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 User likes subset of items

35 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Why not use classification accuracy?
Classification accuracy =
fraction of items correctly classified
(liked vs. not liked)

Here, not interested in what a person

does not like

Rather, how quickly can we discover the

relatively few liked items?
- (Partially) an imbalanced class problem

36 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 How many liked items were
recommended?

Recall
# liked & shown
# liked

37 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 How many recommended items
were liked?

Precision
# liked & shown
# shown

38 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Maximize recall:
Recommend everything

Recall
# liked & shown
# liked

39 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Resulting precision?

Precision
# liked & shown
# shown

40 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Optimal recommender

Recall =1
Precision = 1

41 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 Precision-recall curve
Input: A specific recommender system
Output: Algorithm-specific precision-recall curve

To draw curve, vary threshold on # items recommended

- For each setting, calculate the precision and recall
precision

recall
42 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
 Which Algorithm is Best?
For a given precision, want recall as large as possible
(or vice versa)
One metric: largest area under the curve (AUC)
Another: set desired recall and maximize precision
(precision at k)
precision

recall
43 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization
Summary of
recommender systems

2015 Emily Fox & Carlos Guestrin Machine Learning Specialization

 What you can do now
Describe the goal of a recommender system
Provide examples of applications where recommender systems are useful
Implement a co-occurrence based recommender system
Describe the input (observations, number of topics) and output (topic
vectors, predicted values) of a matrix factorization model
Exploit estimated topic vectors (algorithms to come) to make
recommendations
Describe the cold-start problem and ways to handle it (e.g., incorporating
features)
Analyze performance of various recommender systems in terms of precision
and recall
Use AUC or precision-at-k to select amongst candidate algorithms

45 2015 Emily Fox & Carlos Guestrin Machine Learning Specialization