The best indicator that a passenger - PowerPoint Presentation

Slide1

The best indicator that a passenger

will show up to board the flight

is that she called in for a special meal

Filtering and Recommender Systems

Content-based and Collaborative

4/15Slide2

Filtering and Recommender Systems

Content-based and CollaborativeSlide3

Filtering and Recommender Systems

Content-based and Collaborative

Some of the slides based

On Mooney’s SlidesSlide4

Personalization

Recommenders are instances of personalization software.Personalization concerns adapting to the individual needs, interests, and preferences of each user.

Includes:RecommendingFilteringPredicting (e.g. form or calendar appt. completion)

From a business perspective, it is viewed as part of Customer Relationship Management (CRM).Slide5

Feedback & Prediction/Recommendation

Traditional IR has a single user—probably working in single-shot modesRelevance feedback…

WEB search engines have:Working continuallyUser profilingProfile is a “model” of the user(and also Relevance feedback)

Many users

Collaborative filtering

Propagate user preferences to other users…

You know this oneSlide6

Recommender Systems in Use

Systems for recommending items (e.g. books, movies, CD’s, web pages, newsgroup messages) to users based on examples of their preferences.Many on-line stores provide recommendations (e.g. Amazon, CDNow).

Recommenders have been shown to substantially increase sales at on-line stores.Slide7

Feedback Detection

Click certain pages in certain order while ignore most pages.Read some clicked pages longer than some other clicked pages.

Save/print certain clicked pages.Follow some links in clicked pages to reach more pages.Buy items/Put them in wish-lists/Shopping Carts

Explicitly ask users to rate items/pages

Non-Intrusive

IntrusiveSlide8

Justifying Recommendation..

Recommendation systems must justify their recommendationsEven if the justification is bogus..For search engines, the “justifications” are the page synopses

Some recommendation algorithms are better at providing human-understandable justifications than othersContent-based ones can justify in terms of classifier features..Collaborative ones are harder-pressed other than saying “people like you seem to like this stuff”In general, giving good justifications is important..Slide9

Content-based vs. Collaborative

Recommendation

Needs description of items…

Needs only ratings from other usersSlide10

Content-Based Recommending

Recommendations are based on information on the content of items rather than on other users’ opinions.Uses machine learning algorithms to induce a profile of the users preferences from examples based on a featural description of content.

Lots of systems Slide11

Adapting Naïve Bayes idea for Book Recommendation

Vector of Bags modelE.g. Books have several different fields that are all textAuthors, description, …

A word appearing in one field is different from the same word appearing in anotherWant to keep each bag different—vector of m Bags; Conditional probabilities for each word

w.r.t

each class and bag

Can give a profile of a user in terms of words that are most predictive of what they like

Strengh

of a keyword

Log[P(

w|rel

)/P(w|~

rel

)]

We can summarize a user’s profile in terms of the words that have strength above some threshold.

Related to mutual informationSlide12

Collaborative Filtering

A 9

B 3

C

: :

Z 5

A

B

C 9

: :

Z 10

A 5

B 3

C

: :

Z 7

A

B

C 8

: :

Z

A 6

B 4

C

: :

Z

A 10

B 4

C 8

. .

Z 1

User

Database

Active

User

Correlation

Match

A 9

B 3

C

. .

Z 5

A 9

B 3

C

: :

Z 5

A 10

B 4

C 8

. .

Z 1

Extract

Recommendations

C

Correlation analysis

Here is similar to the

Association clusters

Analysis!Slide13

Item-User Matrix

The input to the collaborative filtering algorithm is an mxn matrix where rows are items and columns are users

Sort of like term-document matrix (items are terms and documents are users)Can think of users as vectors in the space of items (or vice versa

)

Can do vector similarity between users

Pearson correlation coefficient is a variation

And

find who are most similar users

..

Can

do scalar clusters over items etc..

And find what are most correlated items

Think users

docs

ItemskeywordsSlide14

A Collaborative Filtering Method(think

kNN)

Weight all users with respect to similarity with the active user.How to measure similarity?Could use cosine similarity; normally pearson coefficient is usedSelect a subset of the users (neighbors

) to use as predictors.

Normalize ratings and compute a prediction from a weighted combination of the selected neighbors’ ratings.

Present items with highest predicted ratings as recommendations.Slide15

Finding User Similarity with Person Correlation Coefficient

Typically use Pearson correlation coefficient between ratings for active user, a, and another user,

u.

r

a

and

r

u

are the ratings vectors for the

m

items rated by

both

a

and

u

r

i,j

is user

i

’s rating for item

jSlide16

Person Correlation Coefficient is the same as vector similarity over centered ratings vectors

It is easy to check for yourself that pearson correlation coefficient is the same as the cosine theta distance between centered ratings vectors

Covariance = dot productSqrt (Variance of each vector) = norm of each vectorSlide17

Neighbor Selection

For a given active user, a, select correlated users to serve as source of predictions.Standard approach is to use the most similar

k users, u, based on similarity weights, wa,u Alternate approach is to include all users whose similarity weight is above a given threshold.Slide18

Rating Prediction

Predict a rating, pa,i, for each item i

, for active user, a, by using the k selected neighbor users, u  {1,2,…k}.

To account for users different ratings levels, base predictions on

differences

from a user’s

average

rating.

Weight users’ ratings contribution by their similarity to the active user.

ri,j

is user

i

’s rating for item

jSlide19

Similarity Weighting=User Similarity

Typically use Pearson correlation coefficient between ratings for active user, a, and another user,

u.

r

a

and

r

u

are the ratings vectors for the

m

items rated by

both

a

and

u

r

i,j

is user

i

’s rating for item

jSlide20

Significance Weighting

Important not to trust correlations based on very few co-rated items.Include significance weights, sa,u, based on number of co-rated items,

m.Slide21

Covariance and Standard Deviation

Covariance:Standard Deviation:Slide22

Item-centered Collaborative Filtering

Starting with a “centered” user-item matrix, we found k-nearest users to the active user and used them to recommend unrated itemsWe can also use the centered U-I matrix to compute item-item correlations by starting with U-I’xU

-I, and doing (a) association clusters and (b) scalar clustersThis will give us, for each item, k-nearest itemsNow, given a new item In to be rated for a user U, we first find k items closest to In and, and take their (weighted) average rating from the user U as predictive of U’s rating of I

n

An advantage of this method over the “user-centered” idea is that the justifications for the recommendations can be more meaningful (you can tell the user that we are recommending I

n

because she rated the items in its association cluster high..)Slide23

LSI-style techniques for collaborative filtering

The NETFLIX prize was won by an approach that did “latent factor analysis” (aka LSI) on the u-i matrix, so that both users and items are seen as vectors in a k-dimensional factor space

One technical difficulty in doing LSI on u-i matrix is that it has many “null” valuesD-t matrix is sparse and that is good. U-I matrix has null values and that is bad (because null != 0)Two approaches:“fill in” the missing ratings (“Imputation” method) so we have no more null values

“compute distance between vectors only in terms of their common non-null dimensions

Problem:

Overfitting

. Solution: Regularization—penalize “large factor” values.

q

i

item in factor space

p

u

user in factor spaceSlide24

Problems with Collaborative Filtering

Cold Start: There needs to be enough other users already in the system to find a match.

Sparsity: If there are many items to be recommended, even if there are many users, the user/ratings matrix is sparse, and it is hard to find users that have rated the same items.First Rater: Cannot recommend an item that has not been previously rated.

New items

Esoteric items

Popularity Bias

: Cannot recommend items to someone with unique tastes.

Tends to recommend popular items.

WHAT DO YOU MEAN YOU DON’T CARE FOR BRITNEY SPEARS YOU DUNDERHEAD?

#$%$%$&^Slide25

Advantages of Content-Based Approach

No need for data on other users.No cold-start or sparsity problems.Able to recommend to users with unique tastes.

Able to recommend new and unpopular items No first-rater problem.Can provide explanations of recommended items by listing content-features that caused an item to be recommended.Well-known technology

The entire field of Classification Learning is at (y)our disposal!Slide26

Disadvantages of Content-Based Method

Requires content that can be encoded as meaningful features.Users’ tastes must be represented as a learnable function of these content features

.Unable to exploit quality judgments of other users.Unless these are somehow included in the content features.Slide27

Movie Domain

EachMovie Dataset [Compaq Research Labs]Contains user ratings for movies on a 0

–5 scale.72,916 users (avg. 39 ratings each).1,628 movies.Sparse user-ratings matrix – (2.6% full).Crawled Internet Movie Database (

IMDb

)

Extracted content for titles in

EachMovie.

Basic movie information:

Title, Director, Cast, Genre, etc.

Popular opinions:

User comments, Newspaper and Newsgroup reviews, etc.Slide28

Content-Boosted Collaborative Filtering

IMDb

EachMovie

Web Crawler

Movie

Content

Database

Full User

Ratings Matrix

Collaborative

Filtering

Active

User Ratings

User Ratings

Matrix (Sparse)

Content-based

Predictor

RecommendationsSlide29

Content-Boosted CF - I

Content-Based

Predictor

Training Examples

Pseudo User-ratings Vector

Items with Predicted Ratings

User-ratings Vector

User-rated Items

Unrated ItemsSlide30

Content-Boosted CF - II

Compute pseudo user ratings matrix

Full matrix – approximates actual full user ratings matrixPerform CFUsing Pearson corr. between pseudo user-rating vectorsThis works better than either!

User Ratings

Matrix

Pseudo User

Ratings Matrix

Content-Based

PredictorSlide31

Why can’t the pseudo ratings be used to help content-based filtering?

How about using the pseudo ratings to improve a content-based filter itself? (or how access to unlabelled examples improves accuracy…)Learn a NBC classifier C

0 using the few items for which we have user ratingsUse C0 to predict the ratings for the rest of the itemsLoop

Learn a new classifier C

1

using all the ratings (real and predicted)

Use C

1

to (re)-predict the ratings for all the unknown items

Until no change in ratings

With a small change, this actually works in finding a better classifier!

Change: Keep the class posterior prediction (rather than just the max class)

This means that each (unlabelled) entity could belong to multiple classes—with fractional membership in each

We weight the counts by the membership fractions

E.g. P(A=

v|c

) = Sum of class weights of all examples in c that have A=v

divided by

Sum of class weights of all examples in c

This is called

expectation maximization

Very useful on web where you have tons of data, but very little of it is

labelled

Reminds you of K-means, doesn’t it

?

(no coincidence—K-means is “hard-assignment” EM)

Unlabeled examples help only when they are drawn

from the same distribution as the labeled ones..Slide32
Slide33

(boosted) content filtering Slide34

Co-Training Motivation

Learning methods need labeled dataLots of <x, f(x)> pairsHard to get… (who wants to label data?)But unlabeled data is usually plentiful…Could we use this instead??????Slide35

Co-training

Suppose each instance has two parts:

x = [x1, x2]

x1, x2 conditionally independent given f(x)

Suppose each half can be used to classify instance



f1, f2 such that f1(x1) = f2(x2) = f(x)

Suppose f1, f2 are learnable

f1

 H1,

f2

 H2, 

learning algorithms A1, A2

Unlabeled Instances

[x1, x2]

Labeled Instances

<[x1, x2], f1(x1)>

A1

f2

Hypothesis

~

A2

Small labeled data needed

You train me—I train you… Slide36

It really works!

Learning to classify web pages as course pages

x1 = bag of words on a pagex2 = bag of words from all anchors pointing to a pageNaïve Bayes classifiers12 labeled pages1039 unlabeledSlide37

Observations

Can apply A1 to generate as much training data as one wantsIf x1 is conditionally independent of x2 / f(x),then the error in the labels produced by A1 will look like random noise to A2 !!!

Thus no limit to quality of the hypothesis A2 can makeSlide38
Slide39

Focussed Crawling

Cho paper Looks at heuristics for managing URL queueAim1: completeness

Aim2: just topic pagesPrioritize if word in anchor / URLHeuristics: Pagerank#backlinksSlide40

Modified Algorithm

Page is hot if:Contains keyword in title, orContains 10 instances of keyword in body, orDistance(page, hot-page) < 3Slide41

ResultsSlide42

More ResultsSlide43

Conclusions

Recommending and personalization are important approaches to combating information over-load.Machine Learning is an important part of systems for these tasks.Collaborative filtering has problems.

Content-based methods address these problems (but have problems of their own).Integrating both is best.Which lead us to discuss some approaches that wind up using unlabelled data along with labelled data to improve performance.Slide44

Discussion of the Google News Collaborative Filtering PaperSlide45

Advertising

Advertising is a sort of paid recommendationWhile