Web Personalization and Recommender Systems

1
Web Personalization andRecommender Systems
Bamshad Mobasher School of CTI, DePaul University
2
What is Web Personalization

• Web Personalization personalizing the browsing
  experience of a user by dynamically tailoring the
  look, feel, and content of a Web site to the
  users needs and interests.
• Related Phrases
• mass customization, one-to-one marketing, site
  customization, target marketing
• Why Personalize?
• broaden and deepen customer relationships
• provide continuous relationship marketing to
  build customer loyalty
• help automate the process of proactively market
  products to customers
• lights-out marketing
• cross-sell/up-sell products
• provide the ability to measure customer behavior
  and track how well customers are responding to
  marketing efforts

3
Personalization v. Customization

• Its a question of who controls the users
  browsing experience
• Customization
• user controls and customizes the site or the
  product based on his/her preferences
• usually manual, but sometimes semi-automatic
  based on a given user profile
• Personalization
• done automatically based on the users actions,
  the users profile, and (possibly) the profiles
  of others with similar profiles

4
(No Transcript)
5
(No Transcript)
6
(No Transcript)
7
Challenges and Pitfalls

• Technical Challenges
• data collection and data preprocessing
• discovering actionable knowledge from the data
• which personalization algorithms
• Implementation/Deployment Challenges
• what to personalize
• when to personalize
• degree of personalization or customization
• how to target information without being intrusive

8
Web Personalization

• The Problem
• serve dynamic content to users based on their
  profiles or preferences
• Current Approaches
• Rule-based Filtering store a profile for users
  based on explicit registration information use
  prespecified rules to generate recommendations
• Collaborative Filtering requires explicit
  ratings from users to find profiles
• e.g., GroupLens, Firefly, PHOAKS, Syskill
  Webert
• Content-Based Filtering learn/store personal
  profiles locally or on server-side
  recommendations are based on content similarity
• e.g., WebWatcher, Letizia
• Limitations of Current Technologies
• content-based recommendations may be too narrow
• user input is subjective and prone to bias
• profiles may be static and can become outdated
  quickly
• problems with scalability and accuracy

9
Collaborative Filtering

• Examples
• FireFly Network (Shardanand Maes 95)
• Net Perceptions
• Users rate musical artists from like to dislike
• 1 detest 7 cant live without 4
  ambivalent
• There is a normal distribution around 4
• However, what matters are the extremes
• Nearest Neighbors Strategy Find similar users
  and predicted (weighted) average of user ratings
• Pearson r algorithm weight by degree of
  correlation between user U and user J
• 1 means very similar, 0 means no correlation, -1
  means dissimilar

10
Collaborative Filtering (Nearest Neighbor Example)
Sim(Sally, Karen) (776433) / SQRT(
(776633) (774433) )
Euclid(Sally, Karen) SQRT( (7-7)2 (6-4)2
(3-3)2 )
Pearson(Sally, Karen) ( (7-5.75)(7-4.67)
(6-5.75)(4-4.67) (3-5.75)(3-4.67) )
/ SQRT( ((7-5.75)2
(6-5.75)2 (3-5.75)2) ((7- 4.67)2 (4- 4.67)2
(3- 4.67)2))
11
Learning Interface Agents

• Add agents to the user interface and delegate
  tasks to them
• Use machine learning to improve performance
• learn user behavior, preferences
• Useful when
• 1) past behavior is a useful predictor of the
  future behavior
• 2) wide variety of behaviors amongst users
• Examples
• mail clerk sort incoming messages in right
  mailboxes
• calendar manager automatically schedule meeting
  times?
• Personal news agents
• portfolio manager agents
• Advantages
• less work for user and application writer
• adaptive behavior
• user and agent build trust relationship gradually

12
Letizia Autonomous Interface Agent (Lieberman 96)
user
letizia
heuristics
recommendations
user profile

• Recommends web pages during browsing based on
  user profile
• Learns user profile using simple heuristics
• Passive observation, recommend on request
• Provides relative ordering of link
  interestingness
• Assumes recommendations near current page are
  more valuable than others

13
Consequences of passiveness

• Weak heuristics
• example click through multiple uninteresting
  pages en route to interestingness
• example user browses to uninteresting page, then
  goes for a coffee
• example hierarchies tend to get more hits near
  root
• Cold start
• No ability to fine tune profile or express
  interest without visiting appropriate pages
• Some possible alternative/extensions to
  internally maintained profiles
• expose to the user (e.g. fine tune profile) ?
• expose to other users/agents (e.g. collaborative
  filtering)?
• expose to web server (e.g. cnn.com custom news)?

14
WebWatcher

• Dayne Freitag, Thorsten Joachims, Tom Mitchell
  (CMU)
• A "tour guide" agent for the WWW
• user tells agent what kind of information he/she
  is seeking (e.g., set of keywords)
• WebWatcher then accompanies user while browsing
  the web
• highlights hyperlinks that it believes will be of
  interest
• its strategy for giving advice is learned from
  feedback in earlier tours

15
Syskill Webert (Pazzani et al 96)

• User defines topic page for each topic
• User rates pages (cold or hot)
• Syskill Webert creates profile with Bayesian
  classifier
• accurate
• incremental
• probabilities can be used for ranking of
  documents
• operates on same data structure as picking
  informative features
• Only top k (100) informative words are used as
  features
• presence or absence of words provides information
  on classification of pages
• word occurs in a higher percentage of hot pages
  than cold pages

16
Syskill Webert Rating Pages
17
Syskill Webert Rating Pages
18
Usage-Based Web Personalization

• Basic Idea
• find aggregate user profiles by automatically
  discovering user access patterns through Web
  usage mining (offline process)
• match a users active session against the
  discovered profiles to provide dynamic content
  (online process)
• Advantages / Goals
• profiles are based on objective information (how
  users actually traverse the site)
• no explicit user ratings or interaction with
  users (to enter a profile, etc.)
• can preserve user privacy (mining from anonymous
  data)
• usage data captures relationships missed by
  content-based approaches
• Applications
• provide a customized navigational experience for
  users based on their interests
• targeted electronic advertising / personalized
  e-coupons / customer support

19
Clustering and User Profiles

• Collaborative Filtering and Clustering
• CF techniques attempt to match a set of user
  ratings against previous user ratings and find
  nearest neighbors
• Clustering can be used to pre-calculate typical
  user profiles
• Transaction clustering
• Pageviews used as features dimensionality
  problems arise for large sites
• Each cluster contains many transactions problem
  is how to derive useful aggregate profiles from
  large transaction clusters
• Pageview Clustering
• Find overlapping clusters of pageviews directly -
  clusters serve as aggregate profiles
• Can capture overlapping interests of different
  types of users (even those with potentially
  dissimilar transactions)
• Traditional clustering techniques fail due to
  very high dimensionality
• Related work Adaptive Web Sites by Perkowitz
  and Etzioni

20
Automatic Web PersonalizationOffline Process
Batch Process
Data Preparation
Usage Mining
Site Files/ Meta Data
Data Cleaning Page View Identification User/Sessio
n Identification Transaction Identification
Derivation of Aggregate Profiles
Usage Profiles
Server Logs
Transaction Clustering Pageview
Clustering Association Rule Discovery Sequential
Pattern Discovery
21
Automatic Web PersonalizationOnline Process
Online Process
Recommendation Engine
Recommendations
Active Session
22
Real-Time Recommendation Engine

• Keep track of users navigational history through
  the site
• a fixed-size sliding window over the active
  session to capture the current users
  short-term history depth
• Match current users activity against the
  discovered profiles
• profiles either can be based on aggregate usage
  profiles, or are obtained directly from
  association rules or sequential paterns
• Dynamically generated recommendations are added
  to the returned page
• each pageview can be assigned a recommendation
  score based on
• matching score to user profiles (e.g., aggregate
  usage profiles)
• information value of the pageview based on
  domain knowledge (e.g., link distance of the
  candidate recommendation to the active session)

23
Recommendations Based on Association Rules
Search for itemsets (of size w1) can be done
efficiently itemset are organized into a
lexicographic tree. Each node at depth d
corresponds to an itemset, I, of size d with its
children being itemsets of size d1 that contain
I.
24
Discovering Aggregate Usage Profiles

• Characteristics of Aggregate Profiles
• the goal is to effectively capture common usage
  patterns from potentially anonymous click-stream
  data
• profiles are represented as weighted collections
  of pageviews
• weights represent the significance of pageviews
  within each profile
• profiles are overlapping in order to capture
  common interests among different groups/types of
  users (e.g., customer segments)
• multiple profiles may contribute to the
  recommendation set for a given user
• Example Profiles from the ACR (Assoc. for
  Consumer Research) Site

1.00 Call for Papers 0.67 ACR News Special
Topics 0.67 CFP Journal of Psychology and
Marketing I 0.67 CFP Journal of Psychology and
Marketing II 0.67 CFP Journal of Consumer
Psychology II 0.67 CFP Journal of Consumer
Psychology I
1.00 CFP Winter 2000 SCP Conference 1.00 Call
for Papers 0.36 CFP ACR 1999 Asia-Pacific
Conference 0.30 ACR 1999 Annual
Conference 0.25 ACR News Updates 0.24 Conference
Update
25
Methodologies for the Discovery of Aggregate
Profiles

• Discovery of Profiles Based on Transaction
  Clusters
• cluster user transactions - features are
  significant pageviews identified in the
  preprocessing stage
• derive usage profiles (set of pageview-weight
  pairs) based on characteristics of each
  transaction cluster
• Cluster Pageviews
• directly compute overlapping clusters of
  pageviews based on co-occurrence patterns across
  transactions
• features are user transactions, so dimensionality
  poses a problem for traditional clustering
  algorithms
• we use Association-Rule Hypergraph Partitioning
  with an overlap factor

26
Profile Aggregation Based on Clustering
Transactions (PACT)

• Input
• set of relevant pageviews in preprocessed log
• set of user transactions
• each transaction is a pageview vector
• Transaction Clusters
• each cluster contains a set of transaction
  vectors
• for each cluster compute centroid as cluster
  representative
• Aggregate Usage Profiles
• a set of pageview-weight pairs for transaction
  cluster C, select each pageview pi such that
  (in the cluster centroid) is greater than a
  pre-specified threshold

27
Recommendations Based on Aggregate Profiles

• Matching score computed using cosine similarity
• Users active session (pageviews in the current
  window) is compared to each aggregate profile
  (both are viewed as pageview vectors)
• Weight of items in the profile vector is the
  significance weight of the item for that profile
• Weight of items in the session vector can be all
  1s, or based on some method for determining
  their significance in the current session
• Generating recommendations based on matching
  profiles
• from each matching profile recommend the items
  not already in the user session window, and not
  directly linked from the pages in the current
  session window
• the recommendation score for an item is based on
  a combination of profile matching score
  (similarity to session window) and the weight of
  the item in that profile
• additionally, we can weight items farther away
  from the current location of user higher (i.e.,
  consider them better recommendations)

28
PACT - An Example
PROFILE 0 (Cluster Size 4) ---------------------
------------- 1.00 D.html 0.50 A.html 0.50 C.html
0.50 E.html 0.00 B.html 0.00 F.html PROFILE 1
(Cluster Size 2) -------------------------------
--- 1.00 A.html 0.50 B.html 0.50 C.html 0.50 D.htm
l 0.50 E.html 0.50 F.html PROFILE 2 (Cluster
Size 4) ---------------------------------- 0.75
B.html 0.75 F.html 0.50 A.html 0.50 C.html 0.25 D.
html 0.00 E.html
Original Session/user data
Result of Clustering
29
Recommendations Based on PACT
PROFILE 0 (Cluster Size 4) ---------------------
------------- 1.00 D.html 0.50 A.html 0.50 C.html
0.50 E.html 0.00 B.html 0.00 F.html PROFILE 1
(Cluster Size 2) -------------------------------
--- 1.00 A.html 0.50 B.html 0.50 C.html 0.50 D.htm
l 0.50 E.html 0.50 F.html PROFILE 2 (Cluster
Size 4) ---------------------------------- 0.75
B.html 0.75 F.html 0.50 A.html 0.50 C.html 0.25 D.
html 0.00 E.html
Current User Session A.html gt B.html gt C.html
gt E.html
Assume session window size of 3 and unit
weights Sim(U, P0) (0.50.5) / SQRT (1.75
3) 0.44 Sim(U, P1) (0.50.50.5) /
SQRT(2.53) 0.20 Sim(U, P2) (0.750.5) /
SQRT(1.693) 0.25 Candidate Recommendations P
0 D.html (SQRT(0.441.00) 0.66) A.html
(SQRT(0.440.50) 0.47) P1 A.html
(SQRT(0.201.00) 0.45) D.html
(SQRT(0.200.50) 0.32) F.html
(SQRT(0.200.50) 0.32) P2 F.html
(SQRT(0.220.75) 0.41) A.html
(SQRT(0.220.50) 0.33) D.html
(SQRT(0.220.25) 0.23)
Recommendations
30
Integrating Content and UsageFor Personalization
31
Integration of Content Profiles

• Content Profile Representation
• content profiles are also represented as
  overlapping collections of pageview-weight pairs
• cluster features over the n-dimensional space of
  pageviews
• for each feature cluster derive a content profile
  by collecting pageviews in which these features
  appear as significant
• Integration with Recommendation Engine
• Usage and content profiles have similar
  representation, so they can be used by the
  recommendation engine in the same way
• Item weights within profiles must be normalized,
  so that content and usage profiles can be
  compared on the same scale
• One approach match active user session with all
  profiles (both content and usage) then use the
  maximal recommendation score for candidate
  recommendations
• Another approach use content profiles for
  generating recommendations only if no matching
  usage profiles (with sufficient confidence) is
  found

32
How Content Profiles Are Generated
1. Extract important features (e.g., word stems)
from each document
2. Build a global dictionary of all
features along with relevant statistics
Total Documents 41 Feature-id Doc-freq Total
-freq Feature 0 6 44 1997 1 12 59 1998 2 13 76 199
9 3 8 41 2000 123 26 271 confer 124 9 24 c
onsid 125 23 165 consum 439 7 45 psycholog
i 440 14 78 public 441 11 61 publish 549 1
6 vision 550 3 8 volunt 551 1 9 vot 552 4 23 vote
553 3 17 web
33
How Content Profiles Were Generated
3. Construct a document-feature matrix with
normalized tf-idf weights
4. Now we can perform clustering on features (or
documents) using one of the techniques described
earlier (e.g., k-means clustering on features).
34
How Content Profiles Were Generated
Examples of feature clusters obtained using
k-means
CLUSTER 0 ---------- anthropologi anthropologist a
ppropri associ behavior ...
CLUSTER 4 ---------- consum issu journal market ps
ychologi special
CLUSTER 10 ---------- ballot result vot vote ...
CLUSTER 11 ---------- advisori appoint committe co
uncil ...
5. Content profiles are now generated from
feature clusters based on centroids of each
cluster (similar to usage profiles, but we have
feature instead of users/sessions).
35
Comparison of Recommendations(Example Based on
ACR Site)
Recommendations based on usage profiles only
(window size 2)
36
Comparison of Recommendations(Example Based on
ACR Site)
Recommendations based on content profiles only
(window size 2)
37
Prediction Accuracy - Precision (Example Based on
ACR Site)

• 18342 transactions, 62 pageview URLs (after
  filtering)
• Data set divided into training and evaluation
  sets
• Portion of each transaction in evaluation set
  used to generate a recommendation set (based on a
  given recommendation threshold)
• Precision percentage of recommendations
  actually visited in the transaction

38
Prediction Accuracy - Coverage (Example Based on
ACR Site)

• Coverage percentage of visited pageviews
  recommended by the personalization engine

39
Example - ACR Demo Site(http//aztec.cs.depaul.ed
u/scripts/acr2)
Recommendations are presented in a separate frame
at the bottom of each page. Experiment Set UP -
18430 transactions - 129 unique URLs - Session
window 3 - 28 URL clusters after post
processing
40
Automatic Web PersonalizationExample - ACR Demo
Site
Recommendations after navigating through
Presidents Column and Online
Archives Recommendation score is determined
based on matching with discovered user profiles
(clusters), current users session, and the
distance between the current session and the
recommended pages
41
Automatic Web PersonalizationExample - ACR Demo
Site
Recommendations after navigating through
Conference Update and Call for Papers System
recommends pages (which may or may not appear in
the page being viewed) based on what other users
(with similar navigational patterns) have found
interesting
42
Automatic Web PersonalizationExample - ACR Demo
Site
Recommendations after navigating through
Conference Update, Call for Papers, and ACR
2000 Asia-Pacific Conference Note that as the
user navigates to more specific pages, the
recommendations also become more specific