Recommender Systems Session E

1
Recommender SystemsSession E

• Robin Burke
• DePaul University
• Chicago, IL

2
Roadmap

• Session A Basic Techniques I
• Introduction
• Knowledge Sources
• Recommendation Types
• Collaborative Recommendation
• Session B Basic Techniques II
• Content-based Recommendation
• Knowledge-based Recommendation
• Session C Domains and Implementation I
• Recommendation domains
• Example Implementation
• Lab I
• Session D Evaluation I
• Evaluation
• Session E Applications
• User Interaction
• Web Personalization
• Session F Implementation II
• Lab II

3
Web Personalization

• A subtype of recommendation
• 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.
• Items
• web pages
• Data
• instead of ratings
• log data from web sites

4
Raw data
5
Challenges and Pitfalls

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

6
Usage-Based Profiles

• Characteristics
• implicit ratings
• preferences inferred from actions
• passive stance
• system takes no activity to gather more rating
  information
• anonymity
• many users will be anonymous
• cannot rely on long-term identity
• data characteristics
• usually more voluminous than explicit rating data
• more noise than explicit rating data

7
Heuristic Preference Indicators

• How to know if user "likes" something?
• basic heuristic length of time spent
• the longer the more liked
• What if the user takes a phone call?
• threshold
• What if the page is at the end of the session?
• either ignore
• possibly lose crucial information
• or consider positive
• significant noise in failed interactions
• What if the page is really short?
• normalize on page length
• What if the back button doesn't reload page?
• infer session activity
• What if the "user" is actually a web crawler?
• recognize pattern of behavior

8
Problems

• Cold start
• cannot count on long-term profiling
• must be able to make predictions with short
  profiles
• Covert model
• user has no direct input
• can't express interests directly
• even if they are willing to
• Sparsity
• much worse than for explicit profiles
• many more users
• many more items (pages)
• profiles much shorter
• "Hidden web"
• many pages are produced by database queries
• page differences not reflected in the log
• can't recommend indistinguishable items

9
Advantages

• No explicit user ratings or interaction with
  users
• Helps preserve user privacy, by making effective
  use of anonymous data
• Large user base makes CF effective
• if we can implement it scalably
• Content-based / knowledge-based approaches hard
  to use on web data

10
Web Personalization Process

• Generate aggregate user models
• not single users but clusters of similar users
• guess why?
• off-line process
• Steps
• Clustering user transactions
• Clustering items / pageviews
• Association rule mining
• Sequential pattern discovery
• Provide recommendation
• on-line
• match a users active session
• to provide dynamic content

11
Off-line Process
12
On-line Process
13
Representation
Pageview/objects
Session/user data
Raw weights are usually based on time spent on a
page, but in practice, need to normalize and
transform.
14
Clustering

• k-means clustering algorithm
• specify a number of clusters
• system finds an arrangement of items that
  minimizes the mean distance within the items in
  each cluster is minimized
• for each cluster
• think of each items as a vector
• generate the centroid

15
Clustering

• Transaction clusters as Aggregate Profiles
• Each transaction is viewed as a pageview vector
• Each cluster contains a set of transaction
  vectors with a centroid
• Each centroid acts as an aggregate profile with
  representing the weight for pageview pi in
  the profile
• Compute similarity between a current users
  profile (or the active user session) and the
  cluster centroids

16
Recommendation Algorithm

• 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 patterns
• 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)

17
Matching Sessions

• 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

18
Recommendations

• 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
• profile matching score (similarity to session
  window) and
• the weight of the item in that profile
• can include novelty
• weight items farther away from the current
  location of user higher

19
Example
Sample cluster centroid from dept. Web site
(cluster size 330)
20
Using Clusters for Personalization
Original Session/user data
Given an active session A ? B, the best matching
profile is Profile 1. This may result in a
recommendation for page F.html, since it appears
with high weight in that profile.
Result of Clustering
PROFILE 0 (Cluster Size 3) ---------------------
----------------- 1.00 C.html 1.00 D.html PROFILE
1 (Cluster Size 4) ----------------------------
---------- 1.00 B.html 1.00 F.html 0.75 A.html 0.2
5 C.html PROFILE 2 (Cluster Size
3) -------------------------------------- 1.00 A.h
tml 1.00 D.html 1.00 E.html 0.33 C.html
21
Association Rules

• An alternative to clusters is to build
  association rules
• Association rule
• a tuple lti1, i2, .., ikgt
• all of which appear together with some frequency
• Can be used for prediction
• if a user has seen
• lti1, i2, ...ik-1gt then predict ik
• lti2, i3, ...ikgt then predict i1

22
Learning Association Rules

• Multiple passes through a database of
  transactions
• 1st time we collect all items that occur with a
  certain frequency
• 2nd time we collect all pairs
• both items must be in the set above
• 3rd time collect all triples
• until there are none left

23
Recommending with Association Rules

• Keep track of users navigational history through
  the site
• Match current users activity against the
  association rules
• find the longest rule with at least one
  non-matching entry
• sort by confidence
• predict the entry or entries not in the user's
  profile
• Dynamically generated recommendations are added
  to the returned page

24
Sequential methods

• Sequential patterns as profiles
• similar to association rules, but the ordering of
  accessed items is taken into account
• use Markov models
• systems with discrete states and probabilistic
  transitions
• commonly used for pre-fetching pages in web
  servers
• Characteristics
• high accuracy
• but usually low coverage
• few users get recommendations
• sometimes this is OK

25
Example Frequent Itemsets
Sample Transactions
Frequent itemsets (using min. support frequency
4)
26
Example Sequential Patterns
Sample Transactions
CSP (min. support frequency 4)
SP (min. support frequency 4)
27
Example An Itemset Graph
Frequent Itemset Graph for the Example
Given an active session window ltB,Egt, the
algorithm finds items A and C with recommendation
scores of 1 and 4/5 (corresponding to confidences
of the rules B,E gt A and B,E gt C ).
28
Example Frequent Sequence Trie
Frequent Sequence Trie for the Example
Given an active session window ltA,Bgt, the
algorithm finds item E with recommendation score
of 1 (corresponding to confidences of the rules
A,B gt E .
29
Impact of Window Size

• Increasing window sizes (using larger portion of
  users history) generally leads to improvement in
  precision

This example is based on the association rule
approach
30
Associations vs. Sequences

• Comparison of recommendations based on
  association rules, sequential patterns,
  contiguous sequential patterns, and standard
  k-nearest neighbor

Support threshold for Association, SP, CSP 0.04
31
Problems with Web Usage Mining

• New item problem
• Patterns will not capture new items recently
  added
• Bad for dynamic Web sites
• Poor machine interpretability
• Hard to generalize and reason about patterns
• No domain knowledge used to enhance results
• E.g., Knowing a user is interested in a program,
  we could recommend the prerequisites, core or
  popular courses in this program to the user
• Poor insight into the patterns themselves
• The nature of the relationships among items or
  users in a pattern is not directly available

32
Roadmap

• Session A Basic Techniques I
• Introduction
• Knowledge Sources
• Recommendation Types
• Collaborative Recommendation
• Session B Basic Techniques II
• Content-based Recommendation
• Knowledge-based Recommendation
• Session C Domains and Implementation I
• Recommendation domains
• Example Implementation
• Lab I
• Session D Evaluation I
• Evaluation
• Session E Applications
• User Interaction
• Web Personalization
• Session F Implementation II
• Lab II