User models and profiles (representation)

9 minute read

Updated:

What can be modelled in a user model/ user profile

  1. Knowledge
  2. Interests – came from recommender system, profile of the user
  3. Goals and tasks
  4. Background
  5. Individual traits
  6. Context of work more context adaptive

User models – three broad questions

  • image
    1. *What is being modelled?
    • The nature of information
      1. *How is this information represented?
    • In what structure
      1. How to construct and maintain these user models?

1. Knowledge

  • Most popular in learner modelling in adaptive education system, and could vary how we represent it.
  • We need to have a domain of area we want to map. Domain should be supplied to the system; as a list or scalar models

Scalar models

  • Scalar models – categories, numbers which is associated
  • e.g. in evaluating language proficiency
    • Categorical - Unix commands (novice - beginner - intermediate-expert) qualitative
    • Numeric – No.of years doing x (0,1,2,3,4) quantitative
  • Based on these values you can assume the proficiency / knowledge

Structural models – graphs
image

  • In graphs, we have:
    • Nodes – key entities and concepts in the world
    • Links – relevant nodes, or can name the links
  • and create a graph based upon it
  • Have graphs from outside and check if user knows these concepts, and we overlay the graph model on top each other
    • We overlay user/ learner knowledge with the domain knowledge (or expected expertise)
  • Counting the relationship and checking it and based on how many we know we overlay on top of entities, and we count how many concepts each person know

  • E.g.
    • User knowledge - indicating something about the user and the info that user may know. So the fact that she’s German
    • Expected knowledge – Since she is German, so she’d know BMW

Exercise – User model of UoL

  • Wikipedia has a vast amt of information. And wiki takes the knowledge and plots wikidata (as a model)
  • Steps
    • Build a graph model for domain knowledge (what are the nodes and how they are linked) structure
    • Your own model will be an overlay model over the domain model
  • Types of knowledge – declarative vs procedural
    • * Declarative (what we cover) – represented by networks of concepts (i.e. facts and their relationships)
    • Procedural – represented as a set of problem-solving rules
  • Knowledge model of UoL
    • image
      • Binary (do they know? Yes or no)
      • Categories (how well do you know (+++ know well, ++ some knowledge, - no)

2. Interests

  • Most popular in information-oriented systems (e.g. recommendation system)
    • News rec, item rec, interesting travel guide
    • Could have shallow or deep level like user model
  • Interest is just a list of keywords. Can aggregate the list of keywords that indicates user’s interest
    • E.g. news - What kind of news user is reading; can look at title, metadata and gain list of words. user interest could be most frequent word!! key word based
  • Pull it further to concept based
    • Those keywords to be a concept, you need a knowledge beyond/ behind, which comes from knowledge source. (wikidata, google)
    • Have external knowledge graph, you mark that word or phrase and map it to a concept; you have a lot around that concept gives richness around that word
    • It gives richness. So next time you read about oil crisis you can approximate and pull another news coz you know that user is interested
  • Concept level is more powerful in reasoning hence adaptation– if I have concept and relationship you can jump from to another
  • Can stay in keyword and look for synonyms, and could be integrated within that business
    • Either manually create the list of synonyms or use existing source of synonyms

Interests II

  • If we have concepts, we will overlay on top of concepts;
    • We don’t know if they know it or not, we know that these interests them
  • image concepts and knowledge
  • Using overlay model over a structure
    • Knowledge - Can have all the keywords of the system and overlay those with user’s interest
    • Concepts – have all the concepts and their relationships in the system and overlay
      • Smartness of having knowledge behind it is that, if I know user has interest in certain nodes, you can go up and approximate and if the users are interested
  • Challenge – you usually have a limited interaction with the user
    • Have short session w the system and very little detail about their interests sparsity
  • You may not have the exact items user are interested; hence you need some sort of similarity to provide to user
    • To overcome So in concept, connected node; for list of knowledge, similar words
    • Some form of categorization of the list

Exercise – user model on interests in Leeds

  • Taxonomy limits the knowledge to those that has inheritance (instances); but makes it easier to process
  • image

  • Problem with the model
    • Graphs are complex coz it has relationship + goes up. When you pull external taxonomy, it could be very rich. So you need to know how far up you’ll go
    • External source could be noisy and has limited data
    • Users don’t talk in terms of groups, and human settlement so if system starts using this it can be an understanding challenging

3. Goals and tasks

  • image
  • Most challenging/ changeable as goal/ task recognition is difficult
  • Usually modelled with a ‘goal catalogue’ approach (predefined list/ structure of user goals and tasks)
  • To mark users goal somewhere in the catalogue then activate appropriate adaption (e.g. a learning system, help system)
  • Have goals and sub goals, with a tree structure
  • Challenge is to find where the user is in the goal map and how far they’ve accomplished

  • e.g. SQL tutor
    • There’s a simple goal on top and based on tasks users have accomplished, identify the user goal

Goals and tasks II

  • image
  • Using an overlay model with predefined goals or tasks
  • How to approximate that they’ve accomplished the goal?
    • Graphs or lists
  • Goal system with Bayesian network – and use some probabilities to except to what extent users have gone over the goals and sub goals

Exercise – transport guide

  • Q - What will the user model look like if we want a ‘transport guide’ software to adapt to the user when advising on how to get around in Yorkshire?
    • Goal – types of transport
      • Sub goal – tasks to use the transport
    • Goal – locations you wanna visit
      • Sub goal – how to get there
  • image

## 4. Background

  • Relatively stable information about their previous experiences
    • Profession, experience of work in related areas, opinions
    • Demographic information – name, age, sex, nationality
    • To put it in some form of category
  • Common to use user’s background for stereotype modelling (not overlay)
  • Challenge? Privacy !! hence not popular

5. Individual traits

  • Relatively stable characteristics of the user info which together define a user as an individual
    • Characteristics such as cognitive styles, learning styles, personality etc
    • Can be based on psychometric tests
  • E.g. If someone is intrinsically or extrinsically motivated
    • Intrinsically – turn on values, e.g. if I do well my fam will do well
    • Extrinsically – just wanting awards
  • Personality model
    • E.g. Introvert, extrovert, fast thinker, future thinker
    • Personality is useful when: you are trying to convince someone. If they are future/ present thinker you recommend it but tweak it.

6. Context of work

  • Context awareness is paramount for mobile devices
  • Is more like a holistic model
  • User model is more individual, but to personalise we need to know the individual AND the world
    • Hence contextual awareness
  • Early work started from platform adaptation issues, then has grown to mobile and ubiquitous adaptive systems
  • Now extending to other dimensions of the context such as:

    • User location – where is the user and what do I know about that. i.e noise level
      • E.g. Delivering video-based lecture I a mobile device and how long students are gonna consume the video. In a café, room, etc. It’s not a desktop app so we need to change sth. One of the things is to break it down to a smaller entity
    • Physical environment - Is it a noisy environment? Lighting?

    • Social context – in Facebook recommendations based on what your friends are doing. Recommendation based on your social cloud. Could be used for movement, who’s nearby and who should it be directed to

    • Affective state – can we detect the mental state of the user. We can model with some approximation from i.e. user’s speech – positive and negative sentiment
      • Problem – what do we do after that. After we’ve found out if the user is happy or not, back to the application and how are we gonna change it?
  • You only bring one parameter of context each time
    • No social context and physical environment together, bc we don’t know how to deal with so many things
    • E.g. medicine – we don’t give handful of medicine we give pill by poll
    • We isolate parameter by parameter to see what works
  • + While not totally about the user, the addition of the context makes adaptation more effective

Dimensions of context

  • image
  • Can take it from the context of the user (right) and see how far you can stretch further
    • Task/ goal
    • Personal context – interesting coz it can bring issues like disability, mobility and vision
    • Social context – anything about the time, location
    • Overall platform/ ecosystem
      • Time/ location - when, where, how long they’ve got
      • Physical env – noise? Dim?
  • OR take it from the context of the device (left) and see how far you can stretch further
    • Capability of the device – we use the performance of the device
      • E.g. smart watches – very limited screen, so need to be aware of that
    • Environment context – noise, pollution, movement, crowd
    • Human context – what are the influences that may affect the person?
  • Can start with something that is easy to capture and handle (task/goal, device) and you stretch further

  • Q - Personalised travel assistant – which context dimensions will you use? Past exam Q
    • From the user model – some demographic, INTEREST about places! main thing
    • Personal context – mobility problems? (imp in this case)
      • If they have respiratory problem, it can affect it Rec becomes useless if they’ve got problems and it could rather remind them
    • Social context – who they are with, as they exhibit diff behaviours if they are going museum alone and when they are with someone else
    • Time location – when, where, how long they’ve got

An example of User model/ profile

  • Demographic - Name, age, nationality, education level
  • User interest - List of key words, list of topics
  • User preferences - Disabilities, preferred interaction style, preferred media

Summary on user model representation (about 2 lectures)

  • User model can include a range of parameters:
    • Knowledge, interests, goals and tasks, background, individual traits, context of work
  • In class
    • Overlaying with existing / external knowledge base
    • Different ways of representing a model – taxonomy, graph, table
    • Easy to model goals but is hard to find out whereabouts the user is in achieving that goal
  • User model is for ADAPTATION!

Leave a comment