User Models for Personalization

1
User Models for Personalization

• Josh Alspector
• Chief Technology Officer

2
One-to-One Marketing

• Peppers Rogers
• Customized products, services for individual
  customers
• Market knowledge from observations, dialogue and
  feedback with individuals
• Focus on customer loyalty
• Customer Relationship Management

3
Technical Heritage

• Customer databases remember this specific
  customer
• Interactivity customer talks to us or acts
• Mass customization make or do something for
  him

4
Loyalty A Learned Relationship

• Customer tells you what he wants
• You tailor your product, service or elements
  associated with it
• The more effort the customer invests, the greater
  their stake in product or service
• Now the customer finds it more convenient to
  remain loyal rather than re-teach a competitor

5
Market vs. Customer Share
Customer Needs Satisfied
Traditional Marketing
Customers Reached
6
E-Commerce Choices

• If you operate in the product dimension
• Then you must be the lowest cost producer
• Buy a new car at 25 over invoice
• Or, operate in the customer dimension
• Remember this customer when he comes back
• Make it easier and easier to do business

7
Personalization

• Deliver customized offerings
• Create products from components
• Configure and deliver to personal taste
• Generate recommendations
• Analyze user data
• Recognize patterns of behavior
• Develop adaptive models of users
• Retain customers
• Identify and understand individuals
• Match products with needs

8
User Model

• Ideally a model of the users mind
• allows perfect prediction of users needs for
  news and entertainment
• allows advertisers to create ads user will always
  click on
• allows vendors to present products a user will
  always buy
• Nothing is more valuable in the information age

9
Benefits for Customers

• Reduce search time effort
• Improve recommendations
• reduce cost, increase satisfaction
• Improve over time through learning
• Tailored content and advertising
• One-to-one marketing
• Build communities

10
Benefits for Providers

• Match customer needs
• Convert browsers to buyers
• 80 of orders come from 20 of audience
• Higher customer loyalty satisfaction
• Continuous improvement from learning
• Continuous high-quality market research

11
How to Study User Models

• Simulations
• Understand properties
• Controlled experiments
• Focus groups
• Friendly users
• Field studies
• Use actual marketplace

12
Group Models Fill-in Profiles

• Usually a registration procedure
• income, education, sex, age, zip code
• sports, hobbies, entertainment, news
• understanding demographics used by vendors in
  exchange for access to site
• basis for most targeted ads
• interests dont fall into categories, are hard to
  articulate, miss users richness

13
Group ModelsCliques Clicks

• Clique-based classifiers
• collaborative filtering looks at users with
  similar tastes to predict choices
• Amazon suggest books based on your order, richer
  than category romance
• Clickstream analysis high reach
• Polluted data from random clicking

of Audience with Clickstream Data
of Audience with Registration Data
of Audience with Transaction Data
14
Individual Models Features

• Feature-based classifiers
• multiple attributes considered
• compared both for movies
• Text-based classifiers
• information retrieval word vector space
• cluster documents with similar words
• NewSense displays precision of 75
• most internet information is text
• no need to fill in form or rate products

15
Individual Model for Movies
16
Group vs. Individual Movies
User ID Linearcomb. features Cliquerank distance
U21 .36 .67
U111 .53 .84
U39 .54 .75
U145 .31 .31
U77 .37 .34
Avg. Correlation .38 .58
17
Data Analysis NewSense

• Bag of words for visited headlines
• stemming, stop words
• Score recent words higher
• Similarity measure
• cosine (query, document) word vectors
• Query based on visited documents
• terms in relevant (visited) - factorterms in
  irrelevant (not visited) documents

18
Evaluation of Data

• Precision well-defined
• visitedrelevant/all visited
• Recall ill-defined here
• visitedrelevant/allrelevant
• Use average precision
• weighted by threshold of relevancy
• Rocchio, Bayes, SVM P0.75

19
Individual Model News
20
Simulation study (Ariely, MIT)

• Create people
• Create products
• Create decision rule
• Create markets with smart agents

21
Group Individual results I

• Constant taste

Recommendation Quality
Time
22
Group Individual results II

• Gradual taste Change

Recommendation Quality
Time
23
Group Individual results III

• Abrupt taste Change

Recommendation Quality
Time
24
Group Individual results IV

• New Product I

Adoption
Time
25
Group Individual results IV

• New Product II

Recommendation Quality
Time
26
Conclusion

• Wide variety of user models with different
  analyses, applicability effectiveness
• Group models can jump start from zero knowledge
• Individual adaptive models are better over the
  long-run and for new products