Web-Technology Lecture 14

Pre-quiz What does it mean - “lost in hyperspace”? What are the three stages of Adaptive Search? What are the two main categories of Adaptive Hypermedia technologies? Why Amazon wants you to rate products?

Adaptive Web

Size of the Web - Content # of Web-hosts: > 1 billion # of Web-pages: > 55 billion 80% of Web-content generated by users: Daily: 500 Million Tweets 4 Million Hours of YouTube videos 3.6 Billion Instagram Likes 4.3 BILLION Facebook messages Many modern Web systems suffer from an inability to be “all things to all people”. Web courses present the same static learning material to students with widely differing knowledge of the subject. Web stores offer the same selection of "featured items" to customers with different needs and preferences. Health information sites present the same information to readers with different health problems. The systems that can’t meet the needs of their heterogeneous users are often compared to a department store that offers one size of clothing to all customers. A remedy for the negative effects of this "one-size-fits-all'' approach is to develop systems that are able to adapt their behavior to the goals, tasks, interests, and other features of individual users and groups of users. Enter the Adaptive Web – a new research area on the crossroads of human- computer interaction and artificial intelligence. Starting with a few pioneering works on adaptive hypertext in early 1990, it now attracts many researchers from different communities such as hypertext, user modeling, machine learning, natural language generation, information retrieval, intelligent tutoring systems, and cognitive science. Currently, the established application areas of adaptive Web systems are education, information retrieval, and kiosk-style information systems. A number of more recent projects are also exploring new application areas such as e-commerce, medicine, and tourism.

Size of the Web - Users Many modern Web systems suffer from an inability to be “all things to all people”. Web courses present the same static learning material to students with widely differing knowledge of the subject. Web stores offer the same selection of "featured items" to customers with different needs and preferences. Health information sites present the same information to readers with different health problems. The systems that can’t meet the needs of their heterogeneous users are often compared to a department store that offers one size of clothing to all customers. A remedy for the negative effects of this "one-size-fits-all'' approach is to develop systems that are able to adapt their behavior to the goals, tasks, interests, and other features of individual users and groups of users. Enter the Adaptive Web – a new research area on the crossroads of human- computer interaction and artificial intelligence. Starting with a few pioneering works on adaptive hypertext in early 1990, it now attracts many researchers from different communities such as hypertext, user modeling, machine learning, natural language generation, information retrieval, intelligent tutoring systems, and cognitive science. Currently, the established application areas of adaptive Web systems are education, information retrieval, and kiosk-style information systems. A number of more recent projects are also exploring new application areas such as e-commerce, medicine, and tourism.

Size of Web - Activities Searching and Browsing Reading, listening, watching Shopping and banking Communicating and exchanging information Working and having fun

The big question Does one size fit all? Many information systems suffer from an inability to tailor for “all people”. Web courses present the same static material to all students with differing knowledge. Web stores offer the same selection of "featured items" to customers with different needs. Health information sites present the same information to readers with different health problems. The systems that can’t meet the needs of their heterogeneous users are often compared to a department store that offers one size of clothing to all customers. A remedy for the negative effects of this "one-size-fits-all'' approach is to develop systems that are able to adapt their behavior to the goals, tasks, interests, and other features of individual users and groups of users. Enter the Adaptive Web – a new research area on the crossroads of human-computer interaction and artificial intelligence. Starting with a few pioneering works on adaptive hypertext in early 1990, it now attracts many researchers from different communities such as hypertext, user modeling, machine learning, natural language generation, information retrieval, intelligent tutoring systems, and cognitive science. Currently, the established application areas of adaptive Web systems are education, information retrieval, and kiosk-style information systems. A number of more recent projects are also exploring new application areas such as e-commerce, medicine, and tourism.

An alternative: Adaptive systems Adaptive information systems attempt to treat different users differently Adaptive Web: combines principles of AI (formal models of user, task, domain, …) and HCI (interactive interfaces with dynamically personalized components) builds systems that provide personalized experience by supporting users in their online information tasks Web Collects information about individual user User Modeling side Adaptive User Model System Adaptation side Provides adaptation effect Classic loop “user modeling - adaptation” in adaptive systems

What can be modelled by a user model? Knowledge about the content and the system Short-term and long-term goals Interests and needs Navigation / action history User category, background, profession, language, capabilities Platform, bandwidth, context… … the exact content of user model is task-dependent

What Can be Adapted? Adaptive Search Systems tailor search results Adaptive recommender systems Suggest new information items Adaptive Hypermedia Systems adaptive presentation adaptive navigation support Adaptive news systems Filter out irrelevant news or reorder news feeds Adaptive GUI menu adaptation dialog form adaptation Intelligent Tutoring Systems adaptive course sequencing adaptive group formation ...

What personalization can help solving Information overload occurs when the amount of input to a system exceeds its processing capacity. Decision makers have limited cognitive processing capacity. Consequently, information overload causes reduction in decision quality. Lost in a hyperspace A phenomenon of disorientation experienced by users reading and navigating hypermedia documents Getting right information at the right time Different users are different Same users are different at different times There is no silver bullet adaptive technology Adaptation itself is not a silver bullet technology

Personalized Information Access Adaptive search (IR, from 1980) Use word-level profile of interests and remedial feedback to adapt search and result presentation Adaptive hypermedia (HT, ITS, from 1990) Use explicit domain models and manual indexing to deliver a range of adaptation effects to different aspects of user models Web recommenders (AI, ML, from 1995) Use explicit and implicit interest indicators, apply clickstream analysis/log mining to recommend best resources for detected user interests Content-based recommenders Collaborative recommenders Hybrid recommenders

Adaptive search

Why Search Personalization? Different users need different documents in response to the same query Relevance is not enough if the volume of data is high With the growth of information even a good query can return thousands of "relevant" documents Personalization is an attempt to find most relevant documents using information about user's goals, knowledge, preferences, navigation history, etc.

Adaptive Search How the search process can be adapted to the user? How we can model the user in adaptive search? Which adaptation technologies can be applied?

How can Search be Adapted? Results Query Search Engine User profile User profile User profile Before search During search After search

Modeling Users in Adaptive Search Most essential feature: user interests Observing user document selection, adaptive IR systems build profile of user interests Keyword-level modeling A long list of keywords (terms) in place of a domain model User interests are modeled as weighted vector or terms More advanced systems use several profiles for different domains or timeframes

Keyword-based User Profiles

Before: Query Expansion User profile is applied to add terms to the query Popular terms could be added to introduce context Similar terms could be added to resolve indexer- user mismatch Related terms could be added to resolve ambiguity Works with any IR model or search engine

During The user profile is used to organize the results of the retrieval process Retrieve the most interesting documents Filter out irrelevant documents In this case the use of the profile adds an extra step to processing Extended profile can be used effectively Demographics Location Social circle

After Re-ranking of search results is a typical approach for post-filtering Each document is rated according to its relevance (similarity) to the user or group profile This rating is fused with the relevance rating returned by the search engine The results are ranked by fused rating Annotation of search results Results are provided with visual cues encoding useful information Adaptive navigation appraoch

Google adaptive search Milestones: March, 2004 – beta April, 2005 – non-beta extra service November, 2011 – part of normal search October, 2009 – social search added Authenticated users – Google user profile Non-authenticated users – User profile contains: Location Search History (queries and clicks) Web History (Google tracking scripts from adSense and adWords, Chrome, Android,….) Social Networks History from a Multitude of Google services gender, age, languages Topics

Personalized search experiment

Adaptive Hypermedia

Adaptive Hypermedia How hypertext and hypermedia can become adaptive? Which adaptation technologies can be applied? How can we model the user in adaptive hypermedia?

Why Adaptive Hypermedia? Different people are different Individuals are different at different times "Lost in hyperspace” We may need to make hypermedia adaptive where .. There us a large variety of users Same user may need a different treatment The hyperspace is relatively large

What Can Be Adapted? Web-based systems = Pages + Links Adaptive presentation content adaptation Adaptive navigation support link adaptation

Classification of Adaptive Hypermedia techniques

Adaptive Stretchtext (PUSH)

Adaptive annotation in InterBook 3 2 √ 1 1. State of concepts (unknown, known, ..., learned) 2. State of current section (ready, not ready, nothing new) 3. States of sections behind the links (as above + visited)

QuizGuide: Dual Adaptive Annotations

User Modeling in Classic AH Classic AH use external models (besides user model) Domain models, adaptation (pedagogical) modes, stereotype hierarchy, etc. Users are modeled in relation to these models User knowledge of loops is high User is interested in 19th century architecture styles Resources are connected (indexed) with elements of these models (aka knowledge behind pages) This section presents while loop and increment This page is for field-independent learners This church is built in 1876

Domain Model Concept 4 Concept 1 Concept N Concept 2 Concept 5

Indexing of Nodes Concept 4 Concept 1 Concept n Concept 2 Concept m External (domain) model Hyperspace Concept 4 Concept 1 Concept n Concept 2 Concept m Concept 3

Indexing of Fragments Fragment 1 Concept 4 Concept 1 Concept N Concepts Node Fragment 1 Concept 4 Concept 1 Concept N Concept 2 Fragment 2 Concept 5 Concept 3 Fragment K

Concept-Level User Model 3 10 Concept N Concept 2 7 2 4 Concept 5 Concept 3

AH: Known effects Adaptive presentation helps users understand content faster and better Adaptive navigation support reduces navigation efforts and brings the users to the right place at the right time Altogether AH techniques can significantly improve the effectiveness of hypertext and hypermedia systems

Recommender systems

Recommender Systems “Native” adaptive information access approach How can we model the user in recommender systems? Which adaptation technologies can be applied?

Recommender Systems Started as extension of work on adaptive information filtering What is filtering? Search without explicit query Started as library notification services – user provided profiles Later considered user feedback (yes/no or ratings) to automatically improve profile Modern Recommenders can start without user profile, constructing it by observation and user feedback Clicking, rating, bookmarking, downloading, purchasing

Amazon.com Recommendations

Types of Recommender Systems Collaborative Filtering Recommender System Recommendations are generated based on the user history and the community of likeminded users Content-based Recommender System Recommendation generated from the content features associated with products and the ratings of the user Case-based Recommender System Similar to content-based recommendation. Information about items is represented as cases. The system recommends the cases that are most similar to user’s preference. Hybrid Recommender System Combination of two or more recommendation techniques to gain better performance with fewer of the drawbacks of any individual one

Recommendation Procedure Understand and model users Collect candidate items to recommend Based on the recommendation method, predict target users’ preferences for each candidate item Sort candidate items according to the prediction probability and recommend them

What is Collaborative Filtering? Traced back to the Information Tapestry project at Xerox PARC users annotate documents that they have read and the system recommends them new documents to read Expanded to “automatic” CF in the works of Resnick, Riedl, Maes the process of filtering items using the opinions of other people like you Key idea: people who agreed with me in the past, will also agree in the future Compare to Content-based recommendation: items with features similar to those I liked before will be also liked by me

User-based CF Item 1 Item 2 Item 3 Item 4 Item 5 Alice 5 3 4 ??? 16/4 The input for the CF prediction algorithms is a matrix of users’ ratings on items, referred as the ratings matrix. Item 1 Item 2 Item 3 Item 4 Item 5 Average Alice 5 3 4 ??? 16/4 User1 1 2 9/4 User2 14/4 User3 12/4 User4 13/4 Target User

User-based CF

User-Based NN Recommendation 1.Select like-minded peer group for a target user 2. Choose candidate items which are not in the list of the target user but in the list of peer group 3.Score the items by producing a weighted score and predict the ratings for the given items 4.Select the best candidate items and recommend them to a target user Redo all the procedures through 1 ~ 4 on a timely basis

User-based NN: User Similarity Pearson’s Correlation Coefficient for User a and User b for all Products P rated by both users: Pearson correlation takes values from +1 (Perfectly positive correlation) to -1 (Perfectly negative correlation) Average rating of user b

User-based NN: Rating Prediction All users with high enough sim(a,b) form the likeminded neighborhood of the user a: b ∈ neighbors(n) To compute the predictions for the user a to like a candidate product p ∉ P The items with highest prediction scores get recommended

Post-quiz What does it mean - “lost in hyperspace”? What are the three stages of Adaptive Search? What are the two main categories of Adaptive Hypermedia technologies? Why Amazon wants you to rate products?