Upload
others
View
13
Download
1
Embed Size (px)
Citation preview
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 1
CHI 2003 1
Recommender Systems: Collaborating in Commerce and Communities
Joseph A. KonstanJohn Riedl
University of Minnesota{konstan,riedl}@cs.umn.edu
http://www.cs.umn.edu/Research/GroupLens
CHI 2003 2
Introduction
What are Recommender Systems?
Goals of this Tutorial
Brief History of Recommender Systems
CHI 2003 3
The Problem: Overload
CHI 2003 4
Too much stuff!
Too many books!Too many journal articles!
Too many movies!Too much content!
Too many product choices!Heck, even too many people!
CHI 2003 5
Recommenders
Tools to help identify worthwhile stuffFiltering interfaces
E-mail filters, clipping servicesRecommendation interfaces
Suggestion lists, “top-n,” offers and promotions
Prediction interfacesEvaluate candidates, predicted ratings
CHI 2003 6
Scope of Recommenders
Purely Editorial Recommenders
Content Filtering Recommenders
Collaborative Filtering Recommenders
Hybrid Recommenders
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 2
CHI 2003 7
Tutorial Goals and Outline
CHI 2003 8
Goals
When you leave, you should …Understand recommender systems and their applicationKnow enough about recommender systems technology to evaluate application ideasBe able to design and critique recommender application designsSee where recommender systems have been, and where they are going
CHI 2003 9
OutlineIntroductionRecommender Systems Application SpaceMovieLens Case StudyRecommender AlgorithmsEight Principles and Case StudiesDesigning Recommender ApplicationsPrivacy IssuesCommercial Tool Survey
CHI 2003 10
Highlights andFocus Areas
E-Commerce RecommendersCommunity RecommendersKnowledge Management RecommendersOff-The-Desktop RecommendersGroup Application Design ExercisePlenty of Discussion
CHI 2003 11
History of Recommender Systems
CHI 2003 12
The Early Years …
Why cave dwellers survived
Critics, critics, everywhere
The dreaded editor
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 3
CHI 2003 13
Automation
Information retrievalDynamic information needStatic content base
Information filteringStatic information needDynamic content base
CHI 2003 14
Collaborative Filtering
PremiseInformation needs more complex than keywords or topics: quality and taste
Small Community: ManualTapestry – database of content & commentsActive CF – easy mechanisms for forwarding content to relevant readers
CHI 2003 15
Automated CF
The GroupLens Project (CSCW ’94)ACF for Usenet News
users rate itemsusers are correlated with other userspersonal predictions for unrated items
Nearest-Neighbor Approachfind people with history of agreementassume stable tastes
CHI 2003 16
Usenet Interface
CHI 2003 17
ACF Blossomed
1995Ringo (later Firefly)Bellcore Video Recommender
1996 Recommender Systems Workshop
Early commercializationAgents Inc. (later Firefly)Net Perceptions
new issues of scale and performance!CHI 2003 18
Today
Broad commercial applicationWidely used in e-commerceavailable commercial tools
Diverse research communitylive research systemssubstantial integration with machine learning, information filtering, HCI, and more
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 4
CHI 2003 19
Introductions
CHI 2003 20
Who are We?
John RiedlCollaborative and distributed systems
Joe KonstanHuman-computer interaction
GroupLens ResearchNet PerceptionsWord of Mouse
CHI 2003 21
Who are You?
NameWhat you doWho you work for / where you studyBriefly
Your experience with recommender systemsOne key thing you want to get out of this tutorial
CHI 2003 22
The Recommender Application Space
CHI 2003 23
Dimensions of Analysis
DomainPurposeWhose OpinionPersonalization LevelPrivacy and TrustworthinessInterfaces<Algorithms Inside>
CHI 2003 24
Domains of Recommendation
Content to CommerceNews, information, “text”Products, vendors, bundles
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 5
CHI 2003 25
Purposes of Recommendation
The recommendations themselvesSalesInformation
Education of user/customer
Build a community of users/customers around products or content
CHI 2003 26
Whose Opinion?
“Experts”
Ordinary “phoaks”
People like you
CHI 2003 27
Personalization Level
GenericEveryone receives same recommendations
DemographicMatches a target group
EphemeralMatches current activity
PersistentMatches long-term interests
CHI 2003 28
Privacy and Trustworthiness
Who knows what about me?Personal information revealedIdentityDeniability of preferences
Is the recommendation honest?Biases built-in by operator
“business rules”Vulnerability to external manipulation
CHI 2003 29
Interfaces
Types of OutputPredictionsRecommendationsFilteringOrganic vs. explicit presentation
Types of InputExplicitImplicit
CHI 2003 30
MovieLens Case Study
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 6
CHI 2003 31
Key Concepts
UsersItems – MoviesRatings – Explicit, 1 to 5 starsCorrelations – Relationships among usersPredictions – Personal predicted ratingsRecommendations – Suggested Items
CHI 2003 32
How It Works
C.F. Engine
Ratings Correlations
CHI 2003 33
GroupLens Model
C.F. Engine
Ratings Correlations
ratings
CHI 2003 34
GroupLens Model
C.F. Engine
Ratings Correlations
ratings
CHI 2003 35
GroupLens Model
C.F. Engine
Ratings Correlations
ratings
CHI 2003 36
GroupLens Model
C.F. Engine
Ratings Correlations
ratings
request
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 7
CHI 2003 37
GroupLens Model
C.F. Engine
Ratings Correlations
ratings
request
Neighborhood
CHI 2003 38
GroupLens Model
C.F. Engine
Ratings CorrelationsNeighborhood
ratings
requestpredictions
recommendations
CHI 2003 39
Understanding the Computation
Hoop Dreams
Star Wars
Pretty Woman
Titanic Blimp Rocky XV
Joe D A B D ? ? John A F D F Susan A A A A A A Pat D A C Jean A C A C A Ben F A F Nathan D A A
CHI 2003 40
Hoop Dreams
Star Wars
Pretty Woman
Titanic Blimp Rocky XV
Joe D A B D ? ? John A F D F Susan A A A A A A Pat D A C Jean A C A C A Ben F A F Nathan D A A
Understanding the Computation
CHI 2003 41
Hoop Dreams
Star Wars
Pretty Woman
Titanic Blimp Rocky XV
Joe D A B D ? ? John A F D F Susan A A A A A A Pat D A C Jean A C A C A Ben F A F Nathan D A A
Understanding the Computation
CHI 2003 42
Hoop Dreams
Star Wars
Pretty Woman
Titanic Blimp Rocky XV
Joe D A B D ? ? John A F D F Susan A A A A A A Pat D A C Jean A C A C A Ben F A F Nathan D A A
Understanding the Computation
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 8
CHI 2003 43
Hoop Dreams
Star Wars
Pretty Woman
Titanic Blimp Rocky XV
Joe D A B D ? ? John A F D F Susan A A A A A A Pat D A C Jean A C A C A Ben F A F Nathan D A A
Understanding the Computation
CHI 2003 44
Hoop Dreams
Star Wars
Pretty Woman
Titanic Blimp Rocky XV
Joe D A B D ? ? John A F D F Susan A A A A A A Pat D A C Jean A C A C A Ben F A F Nathan D A A
Understanding the Computation
CHI 2003 45
Hoop Dreams
Star Wars
Pretty Woman
Titanic Blimp Rocky XV
Joe D A B D ? ? John A F D F Susan A A A A A A Pat D A C Jean A C A C A Ben F A F Nathan D A A
Understanding the Computation
CHI 2003 46
ML-home
CHI 2003 47
ML-comedy
CHI 2003 48
ML-clist
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 9
CHI 2003 49
ML-rate
CHI 2003 50
ML-search
CHI 2003 51
ML-slist
CHI 2003 52
ML-groups
CHI 2003 53
Recommender Algorithms
CHI 2003 54
Collaborative Filtering
?
TargetTargetCustomerCustomer
Weighted Sum
3
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 10
CHI 2003 55
E-Commerce Scale
Millions of ProductsMillions of CustomersThousands of Clicks per Second
Scalability!
CHI 2003 56
Sparsity
Many customers have no relationshipMany products have no relationshipSynonymy
Similar products treated differentlyIncreases sparsity, loss of transitivityResults in poor quality
CHI 2003 57
Item-Item Collaborative Filtering
I
II
II I II
I
I
I
I
I
I
I
I
I
CHI 2003 58
Item-Item Collaborative Filtering
I
II
II I II
I
I
I
I
I
I
I
I
I
CHI 2003 59
Item-Item Collaborative Filtering
I
II
II I II
I
I
I
I
I
I
I
I
I
CHI 2003 60
Item-Item Discussion
Good quality, in sparse situationsPromising for incremental model
buildingSmall quality degradationBig performance gain
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 11
CHI 2003 61
Another Possible Solution:Dimensionality Reduction
Latent Semantic IndexingUsed by the IR communityWorked well with the vector space modelUsed Singular Value Decomposition (SVD)
Main IdeaTerm-document matching in feature spaceCaptures latent associationReduced space is less-noisy
CHI 2003 62
SVD: Mathematical Background
=R
m X n
U
m X r
S
r X r
V’
r X n
Sk
k X k
Uk
m X k
Vk’
k X n
The reconstructed matrix Rk = Uk.Sk.Vk’ is the closest rank-k matrix to the original matrix R.
Rk
CHI 2003 63
SVD for Collaborative Filtering
. 2. DirectPredictionm x n
1. Low dimensional representation O(m+n) storage requirement
m x k
k x n
CHI 2003 64
Singular Value Decomposition
Reduce dimensionality of problemResults in small, fast modelRicher Neighbor Network
Incremental UpdateFolding inModel Update
CHI 2003 65
Collaborative Filtering Algorithms
Non-Personalized Summary StatisticsK-Nearest Neighbor
User-UserItem-Item
Dimensionality ReductionSingular Value DecompositionFactor Analysis
Content + Collaborative FilteringCHI 2003 66
Collaborative Filtering Algorithms
Graph TechniquesHorting: Navigate Similarity Graph
ClusteringClassifier Learning
Rule-Induction LearningBayesian Belief Networks
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 12
CHI 2003 67
Case Studies in Personalization
CHI 2003 68
Overview
Commercial Agents that Treat You RightLands’ End, Clinique, Launch
We Know (Knew) Who You AreDoubleClick, Angarra, and their apps
Looking for An Opportunity?Product of the Month Clubs, TiVo
Straight from the LabCalendar Apprentice, Moods, and Effective Interrogation
CHI 2003 69
Lands’ End
The ChallengeGet people to buy clothes at a distance
Some Personal ToolsMy Personal Shopper
Learn preferences, recommend, support short-term interests
My Virtual ModelLearn body, use to help customers visualize
CHI 2003 70
My Personal Shopper
CHI 2003 71
Profile Data
CHI 2003 72
Results
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 13
CHI 2003 73
Today’s Interest
CHI 2003 74
My Virtual Model
CHI 2003 75
Seeing MVM
CHI 2003 76
Using MVM
CHI 2003 77
Clinique
The ChallengeBuy make-up online?Attract people to their brand online?
Some Personal ToolsPersonal Consultation
Interview to determine skin typeLooks Maker (of blessed memory)
Match products you already have
CHI 2003 78
Personal Consultation
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 14
CHI 2003 79
PC – 2
CHI 2003 80
Used
CHI 2003 81
Launch
Personalized RadioSome neat stuff
Runs with rules of real radioExplains why music is playingAllows ratings of Artists, Albums, SongsRepetition (to a point) is Good!Die! Die! Die! Button
SadlyNotion of “rating” is a bit strange
CHI 2003 82
Launch 1
CHI 2003 83
Launch 2
CHI 2003 84
Launch 3
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 15
CHI 2003 85
Launch 4
CHI 2003 86
Launch 5
CHI 2003 87
Launch Home
CHI 2003 88
Launch Levels
CHI 2003 89
What About First Timers?
DoubleclickPopular banner ad networkDecided in early 2002 to abandon full personalization
AngarraTracks users and compiles “non-identifying”personalizing informationMakes this data available for first-time personalization (to the segment level)Includes control cases to test effectiveness
CHI 2003 90
Brooks Brothers
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 16
CHI 2003 91
Could Do Better
Product of the Month ClubsForce customers into genre categoriesPotential of automatic shipment
Do you trust Dean and Deluca?
TiVo – Digital Video RecorderGathers lots and lots of dataStill does a lousy job filling “time”
CHI 2003 92
From The Laboratory
Mitchell’s Calendar ApprenticeLearns rules of how to scheduleTakes over as confidence and user permit
Effective InterrogationValue-of-information analysisWhich items to ask about?
High popularity? High Entropy?Value to individual? Value to Community?System-Driven? User-Driven?
CHI 2003 93
FMRL Movie Recommendations
User submits “theme” queryTheme contains examples of items similar to those desired by the user
Set of potentially similar items identifiedUsing item-to-item correlation in ratings space
Potentially similar items ranked based on traditional ACF predictions
CHI 2003 94
Theme Creation
Meeting Ephemeral Needs
CHI 2003 95
Theme Selection
CHI 2003 96
Theme Query Interface – Query Results
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 17
CHI 2003 97
Principles LearnedBe a Customer Agent
Listen, Learn, and UseAnticipate PitfallsBring “Inside” Opportunities and InfoMake the Match
Box Products, Not PeopleIndividuals, not DemographicsEvolving PersonalizationReal-Time Updates
CHI 2003 98
Case Studies inKnowledge Management
CHI 2003 99
Overview
Recommending peopleeBay, ReferralWeb
Recommending DocumentsTacit
Privacy Issues
CHI 2003 100
Ebay feedback profile
CHI 2003 101
ReferralWeb
CHI 2003 102
Tacit
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 18
CHI 2003 103
Privacy Issues
Same as E-commerce, plusExtra sensitivity of profile data
E.g., Tacit’s dual profiles
Honesty/openness vs. edited content
CHI 2003 104
Principles
Recommending documentsEasier to “keep up to date”Automatic routing of content to peopleMore effective researchMore valuable portals and services
Advertising revenue!
Recommending peopleMore efficient organizationsSocial stickiness
CHI 2003 105
Principles
Key issues:Many techniques for analyzing data (keywords, associations, etc.)Control and data privacyGetting honest data when users have ulterior motives
CHI 2003 106
Case Studies in Expertise
CHI 2003 107
Overview
Overcoming ReluctancePriceline Hotels, Ticketmaster
Helping Users Explore PossibilitiesEntrée, Confidence
Wasted OpportunitySee’s Candies
CHI 2003 108
Priceline
The ChallengeGet you to commit money to a hotel they won’t identify
The ApproachesReassure that many have succeeded beforeDemonstrate knowledge of the localeCarefully explain rating systemHelp users with price guidelines
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 19
CHI 2003 109
PL1
CHI 2003 110
PL2
CHI 2003 111
PL3
CHI 2003 112
PL4
CHI 2003 113
TM Hockey
CHI 2003 114
Entrée
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 20
CHI 2003 115
ML with Conf
CHI 2003 116
Sees
CHI 2003 117
Principles LearnedDemonstrate Product Expertise
Use expertise and recommenders to build customer trustProvide deep product data, so that customers can make informed decisionsMake it fun!
CHI 2003 118
Case Studies in Community Recommenders
CHI 2003 119
Overview
Use Communities to Create ContentEpinions, Slashdot, Flutter.com, Social Navigation, Filterbots
Turn Communities into ContentGeocities, MatchMaker, LambdaMoo, MSN Games
CHI 2003 120
Epinions Sienna overview
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 21
CHI 2003 121
Epinions profile
CHI 2003 122
Epinions profile bottom
CHI 2003 123
Epinions earnings
CHI 2003 124
Slashdot Front page
CHI 2003 125
Slashdot Discussion
CHI 2003 126
Flutter.com (betfair.com)
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 22
CHI 2003 127
Social Navigation
Assumptions:Awareness of others (current or past) helps user find relevant pathPaths/location of others is distinctive enough for user to recognize
ApproachMake history or present visible
CHI 2003 128
Social Nav Example
CHI 2003 129
Filterbots
CHI 2003 130
Inspiration
Need selfless, consistent raters
Humans? No: ratings robots.
CHI 2003 131
Filterbot Integration
GroupLens
CHI 2003 132
Filterbots Alone
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 23
CHI 2003 133
GroupLens
Filterbots Apart
CHI 2003 134
Advantages of the FilterBot model
Combines best of agents and humansagents rate frequently, quickly, consistentlyhumans add subjective taste and quality
Framework pulls out the best of eachuse only the bots that work; ignore the othersuse only the people who agree; ignore the othersbalance people and bots based on available ratings and agreement
CHI 2003 135
Risks of Filterbots
• What if no humans read certain articles?“voluntary” censorship or quality control?
• What about rogue filterbots?
• What if people “prefer” filterbots to humans?
CHI 2003 136
The Filterbot Rainbow
Pure CF PersonalAgent
CF + Filterbot Agents + CF
MixedCommunity
CHI 2003 137
Yahoo & GeoCities
Yahoo bought GeoCities, which than tried to change rules on ownership of content; customers protested by “blacking out” all their content until site returned to old rules.
CHI 2003 138
Slashdot article about Geocities
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 24
CHI 2003 139
Geocities Personal Page
CHI 2003 140
Turn Communities into Content
CHI 2003 141
Matchmaker: Seeker Features
CHI 2003 142
Lambda Moo Map
CHI 2003 143
Game site: MSN
CHI 2003 144
Principles
Editorial process is value addedHelp customers find interesting
information from other customersFree is better than paying for it
customers trust what they produceReward creativelyEnable customer control (e.g., social nets)Encourage response (e.g., eBay)
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 25
CHI 2003 145
Case Studies in Using Implicit Preference Data
CHI 2003 146
Overview
Gather “Work Product” DataGoogle, PHOAKS, OWL
Make Implicit Ratings VisibleAmazon
Don’t Ignore Me!My Yahoo!
CHI 2003 147
Google PageRank
• Ranks pages based on incoming links• Links from higher ranked pages matter
more• Combines text analysis with importance
to decide which pages to show you• Runs on network of thousands of PCs!• Works to be hard to trick (e.g., citation
trading)
CHI 2003 148
CHI 2003 149
PHOAKS
Read Usenet news to find web sites!Implicit ratingsFilter URLs to find endorsementsCreate top-n lists of web sites for a Usenet newsgroup community
Links to endorsements (with age shown)Tested against hand-maintained FAQ lists
CHI 2003 150
PHOAKS
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 26
CHI 2003 151
MITRE’s OWL
Recommender for word processorsMonitored word processor command useIdentified common patterns of useRecommended commands to learn
CHI 2003 152
OWL
CHI 2003 153
Amazon.com
Broad collection of recommendersLots of use of purchase data
But what if it was a gift?What if you didn’t like it?What if your account is shared?
Makes this data visibleImprove your recommendationsExplain a recommendation
CHI 2003 154
Amazon Improve Your Recommendations
CHI 2003 155
Amazon Explanation
CHI 2003 156
My Yahoo!
Typical portalUser control over “streams” of contentStreams managed independentlyDoesn’t track usage to update
Already seenWhat I seem to like/dislike
Doesn’t suggest streams I might like
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 27
CHI 2003 157
My Yahoo
CHI 2003 158
Principles Learned
Watch What I DoActions speak louder than wordsDetermine actions by contextMake implicitly gathered data visibleRespond to customers’ reactions to your recommendations
CHI 2003 159
Case Studies around Touchpoints
CHI 2003 160
Outline
KiosksCall CentersMobileZagats
CHI 2003 161
Kiosks
Alienware PC's Now Offered on Best Buy ``Computer Creation Stations'‘
Blockbustercustomer identityprivacy issues
Music Storesampling versus “listening”
CHI 2003 162
Call Centers
Inbound“screen-pops”Legacy systemsappropriateness
OutboundPredict who will buyPredict what they will buyPredict when to contact themOnline campaign management
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 28
CHI 2003 163
GUS Call Center Experiment
Consumers call in purchasesImplemented personalization, tooExperiment:
one group of agents with old methodone group of agents with personalization
CHI 2003 164
With traditional With cross-sell realtime methods recommendations
Avg Cross-Sell Value $19.50* 60%higher
Cross Sell Success Rate 9.8% 50%higher
* Converted from British Pounds
Directed Buyers: Increasing Cross-Sell
CHI 2003 165
WMLLens Login
CHI 2003 166
Consumer Desires
You will love Attack of the Giant
Leeches!
I have 20,000Copies
If Only I couldBring thisShopping!
CHI 2003 167
Consumer Desires
Recommendations wherever I amRecommendations that I can trustControl over what preferences I shareControl over who I share with
CHI 2003 168
Recommendations Unplugged
What movieShould I
see?
What good movies
are close by?
What DVDShould I
buy?
Tell me what I should
See!
Wireless PDA
Voice
WML
Experimental questions•How do users interact?•What usage patterns?•What happens as users gain experience?
•How do different modalities compare?
•How does usage compare with web?
Avant Go
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 29
CHI 2003 169
Peer-to-Peer Recommenders
Builds on model based item-item algorithm
Separate model construction from usageKey Question: How to find neighbors?Incrementally build a model just for me
CHI 2003 170
Zagat
CHI 2003 171
Zagat What it Takes
What happened to my favorite guide?They let you rate the restaurants!
What should be done?Personalized guides, from the people who “know good restaurants!”
CHI 2003 172
Principles Learned
Maintain excellent service across touchpoints
It’s still you however your customers get thereDifferent strokes for different folksEquivalent service does not mean equal service
CHI 2003 173
Designing Recommender Applications
CHI 2003 174
The O-I-P Model
OutputsPresenting recommendations
InputsGathering preference and product data
ProcessProducing recommendations from inputs
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 30
CHI 2003 175
Outputs
TypesSuggestions, predictions, ratings, reviews
DeliveryPush, pull, organic/passive
PresentationLists, annotations, annotations
CHI 2003 176
Inputs
From the targeted userExplicit vs. implicitRatings, purchase history, keyword/attribute, navigation
From the communityProduct attributes, popularity, ratings, purchase histories, reviews, navigation
CHI 2003 177
Process
Manual recommendation – editorsStatistical summarizationAttribute-based selection
Ephermeral personalizationCurrent selectionMarket Basket
Long-Term personalization
CHI 2003 178
Exercise
Work in groups of 3-4
Identify a recommender applicationE-commerce or community application (broadly defined)Ideally tied to one of your research interests, businesses, or personal interests
Outline desired outputs, available inputsDiscuss appropriate processes
CHI 2003 179
Privacy
The conflict between recommender systems and
privacy
CHI 2003 180
Privacy versus Personalization
Personalization
Privacy
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 31
CHI 2003 181
Some Stories
ToysmartAmazon Pricing
CDnow email
CHI 2003 182
Privacy or Trust?
Customers are willing to share dataIf they trust the recipientIf they know what it is being used forIf they benefit
Customers are very waryIf they fear their data is being soldIf they fear it is being used to take advantage of them
CHI 2003 183
P3P from W3C
Web Server PersonalizationEngine
UserProfile
Profile Parser
PrivacyPolicy
ProfileData
CHI 2003 184
Where are the Profiles?
Web client
Web server
Smart Card
PDA
Encrypt
CHI 2003 185
Microsoft Hailstorm
P3P approach to PrivacyMicrosoft volunteers to keep the profiles
Consumer response?Great silence
Business response!Not with our customers!
CHI 2003 186
Commercial Tools
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 32
CHI 2003 187
Tools
Data mining for offline analyticsCampaign management so marketer has
something to doPersonalization for real-time decision-makingDynamic content for Web site Dynamic Content
+ PersonalizationApproach: Goals, Relate to Recommender
Systems, Sample Vendors
CHI 2003 188
Commercial ToolsGoals
• Analyze customer and traffic data to drive business decisions
• Simplify deployment of analytic solutions
• “Close the loop” between the marketer and the customer
DynamicContent
Data Mining
CampaignMgmt Personalize
CHI 2003 189
Commercial ToolsRelate to Personalization
•Similar analytic techniques•Migrating towards real-time
solutions•Lessons learned: human-in-
the-loop stops the loop
DynamicContent
Data Mining
CampaignMgmt Personalize
CHI 2003 190
Commercial ToolsSample Vendors
• digiMine• Accrue• SPSS (+NetGenesis)
DynamicContent
Data Mining
CampaignMgmt Personalize
CHI 2003 191
Commercial ToolsGoals
• Run inbound and outbound campaigns
• Derive more revenue from customers
• Evaluate campaign performance
DynamicContent
Data Mining
CampaignMgmt Personalize
CHI 2003 192
Commercial ToolsRelate to Personalization
• Choose people to campaign at• Choose products to offer• Work “for” rather than “at”
customers
DynamicContent
Data Mining
CampaignMgmt Personalize
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 33
CHI 2003 193
Commercial ToolsSample Vendors
• E.piphany (www.epiphany.com)• Kana (www.kana.com)• Teradata (NCR, www.teradata.com)
DynamicContent
Data Mining
CampaignMgmt Personalize
CHI 2003 194
Commercial ToolsGoals
• Recommend products or content to site visitors
• Help customers find value from site• Create strong relationships with
customers
DynamicContent
Data Mining
CampaignMgmt Personalize
CHI 2003 195
Commercial ToolsRelate to Personalization
• Recommender systems can be used to personalize
• Personalization has very broad meaning: “Hello John” to “I know what you did last summer”
DynamicContent
Data Mining
CampaignMgmt Personalize
CHI 2003 196
Commercial ToolsSample Vendors
• Net Perceptions (www.netperceptions.com)
• LikeMinds (www.macromedia.com)• TripleHop (www.triplehop.com)
DynamicContent
Data Mining
CampaignMgmt Personalize
CHI 2003 197
Commercial ToolsGoals
• Produce HTML content from database
• Simplify authoring, editing, and management of site
• Support interactivity and personalization
DynamicContent
Data Mining
CampaignMgmt Personalize
CHI 2003 198
Commercial ToolsRelate to Personalization
• Often provide simple rules to decide how to deliver content
• Other vendors must work in dynamic content environment
DynamicContent
Data Mining
CampaignMgmt Personalize
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 34
CHI 2003 199
Commercial ToolsSample Vendors
• Vignette (www.vignette.com)• Broadvision(www.broadvision.com)• Microsoft CMS
DynamicContent
Data Mining
CampaignMgmt Personalize
CHI 2003 200
Holy Grail
Combine technologies across entire spectrum
Data mining for offline analyticsCampaign management to drive marketing decisions to customersPersonalization for real-time decision-makingDynamic content for Web site
CHI 2003 201
Conclusions
CHI 2003 202
Trust and Usability
Recommendation is essentially task-basedTricky issues of balance and focus
Correctness vs. ValueDiversity
People notice poor or manipulated recommendations (Amazon, MovieLens)
Interfaces are evolvingMore integrated into task flow
CHI 2003 203
Real-world ExperienceLarge international catalog retailer
17% hit rate, 23% acceptance rate in call centerMedium European outbound call center
17% hit rate, 6.7% acceptance rate from an outbound telemarketing call$350.00 price of average item soldItems were in an electronics over-stocked category and were sold-out within 3 weeks
Medium American online children’s product store (e-mail campaign)
19% click-thru rate vs. 10% industry average14.3% conversion to sale vs. 2.5% industry average
CHI 2003 204
The Golden Rule
“The one who has the gold, makes the rules.”
Price BotsJunglee (Excite), Jango (Amazon), MySimon
Opinion LeadersDeja.com (Half.com), AskIda (Best Buy), Epinions
RecommendersOwned by businesses
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 35
CHI 2003 205
Ida1
CHI 2003 206
Ida2
CHI 2003 207
Ida3
CHI 2003 208
IdaRecs
CHI 2003 209
IdaZoom
CHI 2003 210
IdaMoreRecs
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 36
CHI 2003 211
Detail
CHI 2003 212
Compare
CHI 2003 213
Compare2
CHI 2003 214
Recommender System for User
Find what I wantKnow I will like itTrust system to help meTeam up with my friends to defeat evil
marketers
CHI 2003 215
Recommender System for Marketer
Show people what they will buyLearn what people want so you have itLearn how much they want it so you get
as much as possible
CHI 2003 216
Who will Prevail?
Who is deploying recommender systems?Who has the money?
Consumers will react if trickedAlternatives exist, and will be deployed if
necessary
Recommender Systems: Interfaces and Technology
Copyright 2002 John Riedl and Joseph A. Konstan
April 7, 2003
CHI 2003 37
CHI 2003 217
Recommender Systems
• Becoming necessary for e-commerce• Create value for businesses• Create value for customers• Many open research problems
TechnologyDeployment and InterfacesEffectiveness
CHI 2003 218
Discussion