SV Adaptive Interactive Systems
Lecture 1 – Introduction
➢ This course is about the design and evaluation of interactive systems that automatically
adapt to users and their context.
Adaptive systems
➢ On the Web, we encounter adaptive systems on a daily basis (search engines, apps, job
platforms, e-commerce sites, social media feeds)
➢ In the plethora of information, we have to find the right, relevant item.
➢ Information and choice overload— Two interrelated problems
o originates in information theory exposure to or provision of too much
information (or data) humans have fairly limited cognitive processing capacity
o thus, with information overload, it is likely that decision quality decreases
o → We need to filter information
➢ Information curation (also called content curation):
o carried out (manually) by specially designated curators
▪ e.g., museums and galleries have curators to select items for collection
and display / DJs of radio stations select songs to be played on air
➢ Algorithmic curation: the automatic selection, organizing, and presenting of content.
➢ Addressing consumers is economically important – but its effectiveness is
suffering.
o It is increasingly difficult to attract consumers’ attention
o Consumers are overwhelmed by the quantity of advertising messages
o Consumers suffer from information overload in general
➢ Digital signage
o Information overload does not only happen on the Web but also in physical
space (Times Square, NY)
➢ Attention
o A majority of consumers do not look at displays. Even less pay attention to the
contents
o We need the ‘right’ information, at the ‘right’ time, in the ‘right’ place, in the
‘right’ way, to the ‘right’ person.
Users
➢ One size does not fit all.
➢ Everybody is different and has different preferences and demands.
➢ Personalisation: tailoring a service or a product to meet someone’s individual
requirements → leads to higher relevance
➢ Segmentation
o discovering and addressing groups of individuals with a
common, yet broad, set of characteristics
o e.g., geographic location, interests, time of visit, etc.
➢ Personalization
o it is segmentation stripped to its roots—the individual
o tailoring at the most personal individual level
➢ Terms are not always used as defined
o Where are the boundaries between segmentation and personalization?—It tends
to be blurry.
1
, o Recommendations are often referred to as being “personalized in any case”,
which is not necessarily true. e.g.:
▪ recommendations solely based on gender or country of residence
▪ recommendation of what is currently popular or trendy on a platform
o Is so-called “targeted advertising” a personalization or a segmentation
approach?—It depends.
Context
➢ It is not all about the person—(also) the situation matters.
➢ Depends on the person <> Depends on the situation
➢ An ideal intelligent system is aware of its context —both the person as well as the
situation— and adapts to it.
➢ Music may adapt in real-time to the situation (e.g., activity levels)
Lecture 2 – User Modelling
User models
➢ Why?
o Every user is different
o Has different needs, preferences, and goals
➢ A user model is an internal representation of user characteristics used by a system
(as a basis for adaptation).
➢ A user model is a specification of user characteristics aiming to facilitate reasoning
about the user’s needs, preferences, and behavior.
➢ Frequent user characteristics:
o Age, gender, location, level of education, job title, ...
o Knowledge, experience, expertise, competence, ...
o Personality, cognitive/learning style, ...
o Interests, preferences, habits, ...
o Needs, goals, tasks, …
o Mental states,...
o Emotional state, mood, tired, stressed, ...
o Cultural background, ....
o Interaction patterns
➢ Which characteristics are really relevant?
o Depends on the system’s goals
o Depends on the time frame
o Depends on the available data
o Depends on the required accuracy
User modelling
➢ User Modeling is..
o The process of creating and updating a user model
o by deriving user characteristics from user data
o —which is data that is explicitly provided by the user
o or data that stems from indirect events and observations
➢ Explicit and Implicit User Modelling
o Explicitly provided/stated characteristics by the users themselves (explicit
information) → representations of these characteristics
o Characteristics inferred from the (raw) user data of U (implicit information) →
estimates from the system S
2
, ➢ User modelling process
1. Acquisition of user data
2. Inference of knowledge from the data
3. Representation of the user model
4. Updating a user model
▪ Not always done, but is recommended as user changes over time
Acquisition of user data
1. Asking the user (direct input)
2. More complex way is by observing the user’s actions
3. Using stereotypes
o Stereotypes: any kind of group or class we can identify of users. Adapt to groups
instead of individual persons. E.g. teenage user
➢ You can mix and match the acquisitions of user data
➢ How can the system ask users?
o Yes / no question
o Ratings
o Multiple choice, multiple correct
o Restricted open question (e.g., age, post code)
o Open question with grammar (e.g., search query)
➢ What can the system observe to get user characteristics? What can the system see?
o Clicks, mouse over, movement, time elapsed, ...
o Selection of page (book, news item, etc), time spent, items bought, text written,…
o Frequency of actions
o Physiological responses (through sensors)
➢ Challenges of stereotyping
o More than one stereotype can be active
o Putting the user in a box, a simplification → Adaptation to a group of similar
users
o Stereotypes are often a result of market research → Data is needed
o Based on answers to an initial set of (simple) questions
▪ → Often oversimplification
▪ → Used to initialize user model (to cover cold start period)
▪ → Revised quickly based on user interaction
3
, •
once you have a user model, you can move away from
stereotypes.
• But when there is no/ little data, stereotypes can be useful
➢ Explicit representation
o exam: implicit/ explicit for certain case, and why
o Advantages
▪ process is intuitive
▪ models are interpretable and reproducible
o Disadvantages
▪ limitations in extendibility: users reluctant to put effort in answering
▪ users may be reluctant to disclose certain information
▪ characteristics of user change over time
▪ users are not reliable
▪ users have difficulty knowing/specifying what they want
➢ Implicit estimates
o Advantages
▪ no additional user effort needed
▪ continuously extendable
▪ multiple events over time can strengthen the estimate
o Disadvantages
▪ cold start problem: takes time to learn about the user
▪ possibly conflicting or faulty information
▪ high data volume
▪ observation is indirect
▪ incorrect inference possible
▪ sensors can be intrusive
Inference of knowledge from the data
➢ Knowledge inference is the process of interpreting events and observations on a user U,
making use of conditions, rules or other forms of reasoning, and the storage of the inferred
knowledge in the user model
➢ Three approaches:
1. Detecting patterns in user behavior
▪ Useful when the system should respond to recurrent behavior or infer items
that may be of the user’s interest.
▪ You need to know in advance what you’re searching for and then have the
system detect patterns
▪ Not always positive/ negative (e.g. skipping a song because you have a
different mood, not because you don’t like the song)
▪ Look for patterns and make it mean something (looking at paintings from a
specific artist/ looking at a painting, not because you like it but because there
is a seat in front of it and you want to sit)
2. Matching user behavior with the behavior of other users
▪ Useful when a user behaves in a similar way to other users. Often used for
recommending “unseen” items.
3. Classifying users or content based on user behavior
▪ Used, e.g., for stereotyping and modeling of user interests.
▪ Having some model that classifies users based on user behaviour
▪ clustering or classification that leads to a certain class/ stereotype
4
Lecture 1 – Introduction
➢ This course is about the design and evaluation of interactive systems that automatically
adapt to users and their context.
Adaptive systems
➢ On the Web, we encounter adaptive systems on a daily basis (search engines, apps, job
platforms, e-commerce sites, social media feeds)
➢ In the plethora of information, we have to find the right, relevant item.
➢ Information and choice overload— Two interrelated problems
o originates in information theory exposure to or provision of too much
information (or data) humans have fairly limited cognitive processing capacity
o thus, with information overload, it is likely that decision quality decreases
o → We need to filter information
➢ Information curation (also called content curation):
o carried out (manually) by specially designated curators
▪ e.g., museums and galleries have curators to select items for collection
and display / DJs of radio stations select songs to be played on air
➢ Algorithmic curation: the automatic selection, organizing, and presenting of content.
➢ Addressing consumers is economically important – but its effectiveness is
suffering.
o It is increasingly difficult to attract consumers’ attention
o Consumers are overwhelmed by the quantity of advertising messages
o Consumers suffer from information overload in general
➢ Digital signage
o Information overload does not only happen on the Web but also in physical
space (Times Square, NY)
➢ Attention
o A majority of consumers do not look at displays. Even less pay attention to the
contents
o We need the ‘right’ information, at the ‘right’ time, in the ‘right’ place, in the
‘right’ way, to the ‘right’ person.
Users
➢ One size does not fit all.
➢ Everybody is different and has different preferences and demands.
➢ Personalisation: tailoring a service or a product to meet someone’s individual
requirements → leads to higher relevance
➢ Segmentation
o discovering and addressing groups of individuals with a
common, yet broad, set of characteristics
o e.g., geographic location, interests, time of visit, etc.
➢ Personalization
o it is segmentation stripped to its roots—the individual
o tailoring at the most personal individual level
➢ Terms are not always used as defined
o Where are the boundaries between segmentation and personalization?—It tends
to be blurry.
1
, o Recommendations are often referred to as being “personalized in any case”,
which is not necessarily true. e.g.:
▪ recommendations solely based on gender or country of residence
▪ recommendation of what is currently popular or trendy on a platform
o Is so-called “targeted advertising” a personalization or a segmentation
approach?—It depends.
Context
➢ It is not all about the person—(also) the situation matters.
➢ Depends on the person <> Depends on the situation
➢ An ideal intelligent system is aware of its context —both the person as well as the
situation— and adapts to it.
➢ Music may adapt in real-time to the situation (e.g., activity levels)
Lecture 2 – User Modelling
User models
➢ Why?
o Every user is different
o Has different needs, preferences, and goals
➢ A user model is an internal representation of user characteristics used by a system
(as a basis for adaptation).
➢ A user model is a specification of user characteristics aiming to facilitate reasoning
about the user’s needs, preferences, and behavior.
➢ Frequent user characteristics:
o Age, gender, location, level of education, job title, ...
o Knowledge, experience, expertise, competence, ...
o Personality, cognitive/learning style, ...
o Interests, preferences, habits, ...
o Needs, goals, tasks, …
o Mental states,...
o Emotional state, mood, tired, stressed, ...
o Cultural background, ....
o Interaction patterns
➢ Which characteristics are really relevant?
o Depends on the system’s goals
o Depends on the time frame
o Depends on the available data
o Depends on the required accuracy
User modelling
➢ User Modeling is..
o The process of creating and updating a user model
o by deriving user characteristics from user data
o —which is data that is explicitly provided by the user
o or data that stems from indirect events and observations
➢ Explicit and Implicit User Modelling
o Explicitly provided/stated characteristics by the users themselves (explicit
information) → representations of these characteristics
o Characteristics inferred from the (raw) user data of U (implicit information) →
estimates from the system S
2
, ➢ User modelling process
1. Acquisition of user data
2. Inference of knowledge from the data
3. Representation of the user model
4. Updating a user model
▪ Not always done, but is recommended as user changes over time
Acquisition of user data
1. Asking the user (direct input)
2. More complex way is by observing the user’s actions
3. Using stereotypes
o Stereotypes: any kind of group or class we can identify of users. Adapt to groups
instead of individual persons. E.g. teenage user
➢ You can mix and match the acquisitions of user data
➢ How can the system ask users?
o Yes / no question
o Ratings
o Multiple choice, multiple correct
o Restricted open question (e.g., age, post code)
o Open question with grammar (e.g., search query)
➢ What can the system observe to get user characteristics? What can the system see?
o Clicks, mouse over, movement, time elapsed, ...
o Selection of page (book, news item, etc), time spent, items bought, text written,…
o Frequency of actions
o Physiological responses (through sensors)
➢ Challenges of stereotyping
o More than one stereotype can be active
o Putting the user in a box, a simplification → Adaptation to a group of similar
users
o Stereotypes are often a result of market research → Data is needed
o Based on answers to an initial set of (simple) questions
▪ → Often oversimplification
▪ → Used to initialize user model (to cover cold start period)
▪ → Revised quickly based on user interaction
3
, •
once you have a user model, you can move away from
stereotypes.
• But when there is no/ little data, stereotypes can be useful
➢ Explicit representation
o exam: implicit/ explicit for certain case, and why
o Advantages
▪ process is intuitive
▪ models are interpretable and reproducible
o Disadvantages
▪ limitations in extendibility: users reluctant to put effort in answering
▪ users may be reluctant to disclose certain information
▪ characteristics of user change over time
▪ users are not reliable
▪ users have difficulty knowing/specifying what they want
➢ Implicit estimates
o Advantages
▪ no additional user effort needed
▪ continuously extendable
▪ multiple events over time can strengthen the estimate
o Disadvantages
▪ cold start problem: takes time to learn about the user
▪ possibly conflicting or faulty information
▪ high data volume
▪ observation is indirect
▪ incorrect inference possible
▪ sensors can be intrusive
Inference of knowledge from the data
➢ Knowledge inference is the process of interpreting events and observations on a user U,
making use of conditions, rules or other forms of reasoning, and the storage of the inferred
knowledge in the user model
➢ Three approaches:
1. Detecting patterns in user behavior
▪ Useful when the system should respond to recurrent behavior or infer items
that may be of the user’s interest.
▪ You need to know in advance what you’re searching for and then have the
system detect patterns
▪ Not always positive/ negative (e.g. skipping a song because you have a
different mood, not because you don’t like the song)
▪ Look for patterns and make it mean something (looking at paintings from a
specific artist/ looking at a painting, not because you like it but because there
is a seat in front of it and you want to sit)
2. Matching user behavior with the behavior of other users
▪ Useful when a user behaves in a similar way to other users. Often used for
recommending “unseen” items.
3. Classifying users or content based on user behavior
▪ Used, e.g., for stereotyping and modeling of user interests.
▪ Having some model that classifies users based on user behaviour
▪ clustering or classification that leads to a certain class/ stereotype
4
