Data Mining for Web Personalization

Presented by the Highflyers group

Who are the Highflyers?

Irfan Butt Introduction and Traditional approaches to Web Personalization Joel Gascoigne Data Collection, Preprocessing and Modelling James Silver Pattern Discovery Predictive Web User Modelling Part 1 Aaron John-Baptiste Pattern Discovery Predictive Web User Modelling Part 2 Asad Qazi Evaluating Personalized Models and Conclusion

Introduction
Paper titled: Data Mining for Web Personalization Author: Bamshad Mobasher

Irfan Butt Introduction and Traditional approaches to Web Personalization

Introduction to Web Personalization

Personalization
Delivery of content tailored to a particular user

Web Personalization
Delivery of dynamic content, such as text, links tailored to a particular user or segments of user

Automatic Personalization Vs Customization

Similarity: Both refer to delivery of content Difference: Creation and updating of user profile Examples
Customization: My Yahoo, Dell Website Automatic Personalization: Amazon

Personalization in Traditional Approaches

Two phases in the process of personalization
1) Data Collection Phase 2) Learning Phase

Classification based on learning from data

1. Memory Based Learning (Lazy) Examples: User-based collaborative system, Content-based filtering system 2. Model Based Learning (Eager) Examples: Item-based System

Memory Based Learning VS Model Based Learning

Memory Based Learning (Lazy)

Huge memory required Scalability issue Adaptable to changes

Model Based Learning (Eager)

Limited memory required Easily scalable Learning phase offline Not adaptable to changes

Traditional Approaches to Web Personalization

Rule Based Personalization Systems
Rules are used to recommend item Rules based on personal characteristics of user Static profiles result in degradation of system

Traditional Approaches to Web Personalization

Content-based Filtering Systems
User profile built on content descriptions of items Profile based on previous rating of items

Traditional Approaches to Web Personalization

Collaborative Filtering Systems
Single profile is built in the same way i.e. contentbased filtering Systems Items from more than one profile is used to recommend new item or content These profiles are K Nearest Neighbors based on previous ratings of items of each profile Poor results as the system grows

Data Mining Approach to Personalization

Data Mining (or Web Usage Mining)
The automatic discovery and analysis of patterns in click stream and associated data collected or generated as a result of user interactions with Web resources on one or more Web sites

Data Mining Cycle:

Data preparation and transformation phase. Pattern discovery phase Recommendation phase

Joel Gascoigne Data Collection, Preprocessing and Modelling

Data Modelling and Representation

Assume the existence of a set of m users:
U = {u1, u2, , um}

Set of n items:
I = {in, in, , in}

Data Modelling and Representation

The profile for a user u U is an n-dimensional vector of ordered pairs:
u(n) = {(i1, su(i1)), (i2, su(i2)), , (in, su(in))}

Typically, such profiles are collected over time and stored

Can be represented as an n x m matrix, UP

Data Modelling and Representation

A Personalisation System, PS can be viewed as a mapping of user profiles and items to obtain a rating of interest The mapping is not generally defined for the whole domain of user-item pairs
System must predict interest scores

Data Modelling and Representation

This general framework can be used with most approaches to personalisation In the data mining approach:
A variety of machine learning techniques are applied to UP to discover aggregate user models These user models are used to make a prediction for the target user

Data Sources for Web Usage Mining

Main data sources used in web usage mining are server log files
Clickstream data

Other data sources include the site files and meta-data

Data Sources for Web Usage Mining

This data needs to be abstracted
Pageview
Representation of a collection of web objects

Session
A sequence of pageviews by a single user

All sessions belonging to a user can be aggregated to create the profile for that user

Data Sources for Web Usage Mining

Content data
Collection of objects and relationships conveyed to the user
Text Images

Also, semantic or structual meta-data embedded within the site

Domain ontology
Could use an ontology language such as RDF Or a database schema

Data Sources for Web Usage Mining

Also, operational databases for the site may include additional information about user and items
Geographic information User ratings

Primary Tasks in Data Preprocessing for Web Usage Mining

Data Preprocessing for Web Usage Mining

Goal:
Transform click-stream data into a set of user profiles

This sessionized data can be used as the input for a variety of data mining algorithms or further abstracted

Data Preprocessing for Web Usage Mining

Tasks in usage data preprocessing:
Data Fusion Data Cleaning Pageview Identification Sessionization Episode Identification

Data Preprocessing for Web Usage Mining

Data Fusion:
Merging of log files from web and application servers

Data Cleaning:
Tasks such as:
Removing extraneous references to embedded objects Removing references due to spider navigations

Data Preprocessing for Web Usage Mining

Pageview Identification:
Aggregation of collection of objects or pages, which should be considered a unit This process is dependent on the linkage structure of the site In the simplets case, each HTML file has a one-toone correlation with a pageview Must distinguish between users
Authentication system or cookies

Data Preprocessing for Web Usage Mining

Sessionization:
Process of segmenting the user activity log of each user into sessions, each representing a single visit to the site

Episode Identification:
Episode is a subsequence of a session comprised of related pageviews

Data Preprocessing for Web Usage Mining

These tasks ultimately result in a set of n pageviews
P = {p1, p2, , pn}

A set of v user transactions

T = {t1, t2, , tv}

A user transaction captures the activity of a user during a particular session

Data Preprocessing for Web Usage Mining

Finally, one or more transactions or sessions associated with a given user can be aggregated to form the final profile for that user
If the profile is generated from a single session, it represents short-term interests Aggregation of multiple sessions results in profiles that capture long-term interests

Data Preprocessing for Web Usage Mining

The collection of these profiles comprises the m x n matrix UP which can be used to perform various data mining tasks After basic clickstream preprocessing steps, data from other sources is integrated:
Content, structure and user data

James Silver Pattern Discovery Predictive Web User Modelling Part 1

Model-Based Collaborative Techniques

Two-stage recommendation process:
(A) offline model-building (B) Real-time scoring
(Explicit & Implicit user behavioural data used)

Offline model-building algorithms:

(1) Clustering, (2) Association Rule Discovery, (3) Sequential Pattern Discovery, (4) Latent Variable Models (part 2)

We also look at hybrid models (part 2)

(1) Clustering
Clustering divides data into groups where:
Inter-cluster similarities are minimised Intra-cluster similarities are maximised

Generalization to Web usage mining User-based vs. Item-based clustering Efficiency and scalability improvements

(1) Clustering: User-based

User profiles Partitions Matrix UP
Clusters represent user segments based on common navigational behaviour

Recommendations (target user u, target item i)

Centroid vector vk computed for each cluster Ck Neighbourhood: All user segments that have a score for i and whose vk is most similar to u

(1) Clustering: Other

Fuzzy Clustering
Desirable to group users into many categories

Distance issues
Consider web-transactions as sequences

Association Rule Hypergraph Partitioning

(ARHP)

(2) Association Rule Discovery

Finding groups of pages or items that are commonly accessed or purchased together

Originally for mining supermarket basket data Discovering Association Rules involves:
1)Discovering frequent itemsets
Satisfying a minimum support threshold

2)Discovering association rules

Satisfying a minimum confidence threshold

(2) Association Rules: Concepts

Transactions set T Itemsets I = {I1,I2,...,Ik} over T Association rule r has the form X => Y (sr, cr)
sr = the support of X U Y
(i.e. probability that X and Y occur together in a transaction) cr = the confidence of the rule r (i.e. the conditional probability that Y occurs in a transaction, given that X has occurred in that transaction)

(2) Recommendations
Matching rule antecedents with target user profiles
Sliding window solution Naive approach Frequent Itemset Graph

Finding Candidate pages:

Match current user session window with previously discovered frequent itemsets

Recommendation Value
Confidence of corresponding association rule

(2) Recommendations

(3) Sequential Models

Now we consider the order when discovering frequently occurring itemsets.
So: given the user transaction {i1,i2,i3} Association rules (i1=>i2) and (i2=>i1) are fine But sequential pattern (i2=>i1) not supported

Two types of sequences: i3

Contiguous (closed) sequence Open Sequence {i1,i2,i4,i3}

i1,i2 =>
{i1,i2,i3}

Frequent Navigational Paths

(3) Recommendations
Trie-structure (aggregate tree)
Each node is an item, root is the empty sequence

Recommendation Generation
Found in O(s) by traversing the tree
s = the length of the current user transaction deemed to be useful in recommending the next set of items

Sliding window w Maximum depth of tree therefore is |w|+1 Controlling the size of the tree

(3) Sequential Models: Contiguous

Contiguous sequence patterns are particularly restrictive
Valuable in page pre-fetching applications Rather than in general context of recommendation generation

(3) Sequential Models: Markov

Another approach for sequential modelling
Based on Stochastic methods

Modelling the navigational activity in the website as a Markov chain

(3) Sequential Models: Markov

A Markov model is represented by the 3-tuple <A,S,T>
A: set of possible actions (items) S: set of n states for which the model is built (visitors navigation history) T=[pi,j]nxn: Transition Probability Matrix
pi,j: probability of a transition from state si to state sj

Order : Number of prior events used in predicting each future event

(3) Markov for Web-mining

Designed to predict the next user action based on the users previous surfing behaviour Also used to discover high-probability user navigational paths in a website
User-prefered trails

Various optimization methods Apart from Markov: Mixture Models

Aaron John-Baptiste Pattern Discovery Predictive Web User Modelling Part 2

(4) Latent Variable Models (LVMs)

Latent Variables are variables that haven't been directly observed but have rather been inferred.
E.g. Morale is not measured directly but inferred

Have more recently become popular as a modelling approach in web usage mining Two commonly used LVMs
Finite Mixture Models (FMM) Factor Analysis (FA)

(4) FA and FMM

Factor Analysis
Aims to summarise and find relationships within observed data (all data) Used in pattern recognition, collaborative filtering and personalization based web usage mining

Finite Mixture Models (FMM)

Use a finite number of components to model (a page view, or user rating)

(4) Drawbacks to pure usage based models

Pure usage based models have drawbacks
Process relies on user transactions or rating data New items or pages are therefore never recommended (new item problem) Also do not use knowledge from underlying domain and so cannot make more complex recommendations

(5) Hybrid models

Uses a combination of user-based and contentbased modelling. Three main types used in web mining
Integrating content features Integrating semantic knowledge Using Linkage structure

(5) Integrating content features with usage-based models

Solves new item problem
Use content characteristics of pages with userbased data Extract keywords from content to be used to discover patterns Not just using user data means new pages with relevant content can be recommended Users interests can be mapped to content, (concepts or topics)

(5) Integrating structured semantic knowledge with usage-based models

Content feature integration is useful when pages are rich in text and keywords However cannot capture more complex relationships where items have underlying properties Idea is to take the underlying meanings of objects and add them to the user-based data. Recommendations can then be made to pages or items with similar semantic meanings

(5) Using Linkage structure for model learning and selection

Other semantic data can be used such as relational databases and the hyperlink structure on a web page Mobasher proposes a hybrid recommendation system that switches between different algorithms based on the degree of connectivity in the site and user E.g. in a highly connected website, with short paths, non sequential models performed better

Asad Qazi Evaluating Personalized Models and Conclusion

Evaluating Personalization models

The Primary Goal of this section is to evaluate the accuracy and effectiveness of web personalization models

Why Evaluate?
More complex web-based applications and more complex user interaction requires the selection of more sophisticated models Need to further explore the impact of recommended model on user behaviour There are several different modelling approaches to web personalization Evaluating personalized models is an inherently challenging task firstly, because different models require different evaluation metrics, secondly, the required personalization actions may be quite different depending on the underlying domain, relevant data and intended application Finally, there is also a lack of consensus among researchers as to what factors affect quality of service in personalized systems and of what elements contribute to user satisfaction

Common evaluation approaches

A number of metrics have been proposed in literature for evaluating the robustness and predictive accuracy of a recommender system: this includes Mean Absolute Error (MAE) Classification Metrics (Precision and Recall) Receiver Operating Characteristic (ROC) The use of business metrics to measure the customer loyalty and satisfaction such as Recency Frequency Monetary (RFM) The use of other key dimensions along with metrics such as: Accuracy, Coverage, Utility, Explainability, Robustness, Scalability and User Satisfaction

Conclusions
Web personalisation is viewed as an application of data mining which dynamically serves customized content (pages, products, recommendations, etc.) to users based on their profiles, preferences, or expected interests of data available to personalization systems, the modelling approaches employed and the current approaches to evaluating these systems We have also discussed the various sources of data available to personalization systems, the modelling approaches employed and the current approaches to evaluating these systems Recent user studies have found that a number of issues can affect the perceived usefulness of personalization systems including, trust in the system, transparency of the recommendation logic, ability for a user to refine the system generated profile and diversity of recommendations Most personalization systems tend to use a static profile of the user. However user interests are not static, changing with time and context. Few systems have attempted to handle the dynamics within the user profile.

Any Questions?