A Personalized Recommendation System Based on PRML for E-Commerce

1
A Personalized Recommendation System Based on
PRML for E-Commerce

• Young Ji Kim, Hyeon Jeong Mun,
• Jae Young Lee and Yong Tae Woo
• Dept. of Computer Sciences, Kosin University,
  Korea
• yjkim_at_hibrain.net

2
Personalization

• Whats Personalization?
• The process of customizing the contents and
  structure of a web site to the specific and
  individual needs of each user taking advantage of
  the users behavior patterns.
• Why need Personalization?
• Technique to maintain closed relationships with
  clients.
• analyzing clients preferences.
• providing differentiated service to preferred
  clients for Internet based applications.
• Important role in a one-to-one marketing strategy
  to enhance both customer satisfaction and profits
  on an E-commerce site.

3
Personalization

• What is the need for personalization?
• Need to know clients preferences.
• What did clients buy?
• What did clients want or like?
• What things will the client be interested in?
• Steps to personalization.
• Collect users behavior.
• Analyze users behavior from collected data.
• Predict users behavior using analyzed results.
• Recommend things which client will be interested
  in.

4
Personalized Recommendation System

• Whats a personalized recommendation system?
• Analyze users behavioral patterns and recommend
  new products that best match the individual
  users preferences.
• Existing recommendation techniques
• Rule-based filtering technique
• Use demographic information
• Collaborative filtering technique
• Use other users rating value with similar
  preference
• Content-based filtering technique
• Compare user profile and product description
• Item-based filtering technique
• Analyze association among products

5
Personalized Recommendation System

• Problems of the existing techniques
• Some users are concerned about privacy issues
• Do not enter personal information.
• Enter incorrect information.
• Not easy to dynamically incorporate time-varying
  aspects of user preference using on existing log
  file.
• Existing log file does not contain enough
  personal information.
• Existing methods are tailored to particular
  applications.
• Lack ability to analyze user behavior patterns.
• Lack ability to dynamically generate and
  recommend web contents.

6
Proposed System

• Proposed system
• Propose a new personalized recommendation
  technique based on PRML.
• First, we make each users PRML instance.
• Users behaviors are collected from XML-based web
  sites.
• Save them as PRML instance.
• Second, we build each users profile.
• Analyze each users PRML instance.
• Make each users profile using them.
• Third, we recommend the products with Top-N
  similarities.
• Personalized recommendations are made by
  comparing the similarity between the information
  about new products and users profile.

7
Proposed System
8
Personal Information Collection System

• Whats PICS(Personal Information Collection
  System)?
• Collect users behavioral patterns while a user
  is connected.
• When the user connect.
• Where the user connect.
• What the user do.
• click, read and scrap contents, use shopping
  cart, purchase, etc.
• Save it as PRML instances.
• Existing method to collect users behavior
• Need to extract individual user's behavior
  patterns from mass web log.
• Various web log formats such as CLF(Common Log
  Format), IIS, W3C Ext. have been used in
  different web servers to record log information.

9
Personal Information Collection System

• Existing method to collect users behavior

10
Personal Information Collection System

• Existing method to collect users behavior
• Need to preprocess step such as referred in
  previous section.
• Use different log formats and need to remove
  unnecessary data such as images or scripts.
• Difficult to extract session information to
  identify an individual user.
• Difficult to collect users behaviors in real
  time.
• Proposed PICS
• Implement to collect the personalized information
  from individual client's behaviors in real time.
• Save personalized information as PRML instances.

11
Personal Information Collection System

• Configuration of personal information collection
  system

12
PRML for Personalized Services

• Whats PRML?
• Personalized Recommendation Markup Language.
• To efficiently store and manage individual
  clients behaviors.
• Conceptual diagram of PRML schema

PRML
???
1m
???
User Identification Information
0m
1m
0m
13
User Session Management Module

• Purpose
• To effectively identify and manage user
  information.
• What does it do?
• An agent at the server side collects user access
  information from each user session.
• User ID, session ID, IP address, URL, server
  status and etc.
• Convert user access information to PRML instance.
• PRML instance is summarized into user
  identification information and log information.
• Save the PRML instance in XML database.

14
User Session Management Module

• Schema structure of personal identification
  information section in PRML

15
User Session Management Module

• Example of personalized identification
  information section in PRML instance

lt?xml version"1.0" encoding"UTF-8"?gt ltPRML xmlnsxsihttp//www.w3.org/2001/XMLSchema-instance xsinoNamespaceSchemaLocation "http//www.hibrain.net/prml/PRML.xsd"gt ltUSER ID"gdhonggt ltSESSIONID ID"JHPWDWORDS" LOGIN_DATE "2005/06/26 102158" LOGOUT_DATE" 2005/06/26 104012 "/gt ltIPADDR IP"203.246.6.121"/gt ltAGENT TYPE"Mozilla/4.0"/gt     ltREQUEST_SETgt       ltREQUESTgt         ltREQUEST_URL URL"/serviet/RecruitManager?RecruitCmdRecruitSummaryView"/gt         ltTIME DATE"2005/06/26 102322"/gt         ltBYTES SIZE"1024"/gt         ltHTTPCODE METHOD"GET" NAME"HTTP/1.1" STATUS_CODE"200"/gt ..       lt/REQUESTgt ltREQUESTgt.lt/REQUESTgt.     lt/REQUEST_SETgt   lt/USERgt lt/PRMLgt
16
Implicit Rating Information Collection Module

• Purpose
• Implicitly collect rating information from
  XML-based web sites utilizing hierarchical
  characteristics of XML documents.
• Preparation
• Elements in the XML documents are assigned
  different weights based on their importance in
  the documents.
• Store these weights in the element weight
  database.
• What does it do?
• When a user visits a web site, the module
  collects the XML elements in the XML contents
  which the user accessed.
• Save them as PRML instance.

17
Implicit Rating Information Collection Module

• Configuration of implicit rating collection
  technique
• Schema of implicit rating information collection
  section

18
Experimental XML document

• XML schema structure of faculty contents

19
Experimental Element Weight Database

• Element weight database
• In the element weight database, each element has
  a level weight and element weight.
• The level weight of an element.
• Determine by its position in the hierarchy of the
  XML documents.
• The element weight of an element.
• Reflect the importance of XML documents.
• An experimental element weight database

20
Implicit Rating Information Module
21
CBR feature Information Collection Module

• Purpose
• Collect CBR feature information to extract users
  preference on web site contents.
• Preparation
• Select feature elements.
• Some elements in an XML document are considered
  important characteristics.
• Store them in the characteristics of XML document
  database.
• What does it do?
• When a user accesses XML document, the feature
  information in the XML document is collected.
• Save it as PRML instance along with the users
  implicit rating information.

22
CBR feature Information Collection Module

• Configuration of CBR feature collection technique
• Schema structure of CBR feature collection
  section

23
CBR feature Information Collection Module
24
Proposed Personalized Recommendation System

• Personalized Recommendation System
• Use a CBR-based learning technique.
• Create user profile based on the PRML instance
  and save in the user profile database.
• Compute the similarity between the user profile
  and each new product.
• Recommend to the user the new products with Top-N
  similarities.

25
Proposed Personalized Recommendation System

• Configuration of proposed system using CBR
  technique

26
Personalized Rating Information Calculation Module

• Purpose
• Compute users preference of each contents a user
  accessed.
• Use implicit rating information collection
  section in the PRML instance and element weight
  database.
• Steps to calculate implicit rating information
• Group all the elements by contents id.
• all the elements collected by the implicit rating
  information collection module are divided into
  groups based on their contents.
• Retrieve element weights and level weights from
  the element weight database.
• Compute rating information of the each contents.

27
Personalized Rating Information Calculation Module

• Rating information of the content
• V is the set of elements in the XML content the
  user accessed.
• le is the level weight of the element e.
• ke is the element weight of e.
• Rc is the implicit rating information.

28
CBR-based Learning technique

• Traditional case-based reasoning system
• When a new problem appears, the system retrieves
  the most similar case, reuses the case to solve
  the problem.
• Revises the proposed solution if necessary, and
  retains the new solution as a part of a new case.
• Proposed the CBR-based Learning technique
• Make users profile analyzing users behavior
  patterns.
• Suggest the recommendation of the most similar
  ones using the past preference information stored
  in the user profile.
• Update the user profile for learning the new case.

29
User Profile Management Module

• Select contents
• Select contents whose implicit rating value(Rc)
  is high.
• Build user profile using CBR feature information
  refer to selected contents.
• User profile
• P (u, A, R, D)
• u is a user ID.
• A is the set of attributes in the web contents.
• R is a set of intra-attribute weights.
• D is a set of inter-attribute weights.

30
User Profile Management Module

• Intra-attribute weights
• The intra-attribute weights R of Ai is ri1, ri2,
  , rim.
• kij is the number of times aij is accessed.
• rij represents how much a user prefers the
  attribute value aij to other attribute values.

i 1, 2, , n, and j 1, 2, , m.
31
User Profile Management Module

• Intra-attribute weights

User profile User profile User profile User profile User profile
Userid (u) Userid (u) gdhong gdhong gdhong
Attribute (A) Attribute Value (ai1..aim) Appear Count (kij) Intra- attribute weight (R) Inter-attribute Weight (D)
Major Database 7 - -
Major Animation 1 - -
Major Network 2 - -
position Professor 4 - -
position Researcher 3 - -
position Post-Doc 3 - -
Location Pusan 2 - -
Location Seoul 8 - -
rij ?
Compute rij of A1(Major)
Attribute value Attribute value Appear count Appear count Intra-attribute weight Intra-attribute weight
a11 Database k11 7 r11 0.7
a12 Animation k12 1 r12 0.1
a13 Network k13 2 r13 0.2
32
User Profile Management Module

• Inter-attribute weights
• The inter-attribute weights D of A is d1, d2,
  , dn.
• each di represents how much Ai is preferred by
  the user.
• If di is large,
• the attribute Ai is more important to the user
  than other attributes.

33
User Profile Management Module

• Inter-attribute weights

di ?
User profile User profile User profile User profile User profile
Userid (u) Userid (u) gdhong gdhong gdhong
Attribute (A) Attribute Value (ai1..aim) Appear Count (kij) Intra- attribute weight (R) Inter-attribute Weight (D)
Major Database 7 0.7 -
Major Animation 1 0.1 -
Major Network 2 0.2 -
Position Professor 4 0.4 -
Position Researcher 3 0.3 -
Position Post-Doc 3 0.3 -
Location Pusan 2 0.2 -
Location Seoul 8 0.8 -

• d1 of Major(A1) 0.7 (1/3) 0.4
• d2 of Position(A2) 0.4 (1/3) 0.1
• d3 of Location(A3) 0.8 (1/2) 0.3
• each di of Ai(Attribute)

Attribute Attribute Inter-attribute Weight Inter-attribute Weight
A1 Major d1 0.4
A2 Position d2 0.1
A3 Location d3 0.3
34
Contents Recommendation Module

• Contents Recommendation Module
• Analyze individual users behavioral pattern to
  generate recommendation for the user.
• Use nearest-neighbor approach to compute the
  similarities between the attributes of user
  profile(P) and new products(I).
• To compute similarity
• aij is the attribute value of Ai in P
• aij is that of I
• if aij aij , f (aij, aij) returns 1 and
  otherwise, 0.

35
Experimental Results

• Experiment
• Experimental content
• XML contents of a faculty position recruiting web
  site.
• Number of User
• 824 person.
• Accessed contents
• 1,144 XML faculty contents.
• New contents
• 1,484 faculty contents.

36
Experiment for Personal Information Collection
System

• PRML instance

37
Experiment for Proposed Recommendation System

• User profile

User profile User profile User profile User profile User profile
Userid (u) Userid (u) gdhong gdhong gdhong
Attribute Of item (A) Attribute Value (ai1..aim) Appear Count (kij) Intra-attribute weight (R) Inter-attribute Weight (D)
Major Database 7 0.7 0.4
Major Animation 1 0.1 0.4
Major Network 2 0.2 0.4
Position Professor 4 0.4 0.1
Position Researcher 3 0.3 0.1
Position Post-Doc 3 0.3 0.1
Location Pusan 2 0.2 0.3
Location Seoul 8 0.8 0.3
38
Experiment for Proposed Recommendation System

• Experimental Results of recommendation
• Use MAE(Mean Absolute Error) and ROC(Receiver
  Operating Characteristic)

39
Conclusion

• Proposed System
• Personalized recommendation system
• Use the PRML approach.
• Define the inter-attribute weights and
  intra-attribute weights.
• Build user profile based on the behavioral
  patterns of a user.
• Recommend the products with Top-N similarities.
• Future work
• Research a Personalized recommendation system
  using ontology.
• Research User Ontology extending the proposed
  user profile.
• Research Domain Ontology to represent contents
  feature.
• Research Log Ontology to represent users
  behavior patterns.