USER MODELING meets the Web Alexandra I. Cristea Modern Information Systems seminar, 26th February 2002 Overview: UM 1. 2. 3. 4.
5. 6. 7. 8. UM: What is it? Why? & What for? & How? Early history Demands & traditional (academic) developments What can we adapt to? Generic User Modelling techniques Newer developments
The future? What is a user model? Elaine Rich: "Most systems that interact with human users contain, even if only implicitly, some sort of model of the creatures they will be interacting with."
What is a user model? Robert Kass: "... systems that tailor their behaviour to individual users' needs often have an explicit representation structure that contains information about their users, this structure is generally called a user model." What is a user model, here:
If a program can change its behaviour based on something related to the user, then the program does (implicit or explicit) user modelling. Why user modelling? pertinent information What is pertinent to me may not be pertinent to you information should flow within and between users users should control the level of information-push
large amounts of information too much information, too little time people often become aware of information when it is not immediately relevant to their needs Difficult to handle Etc. What for? In tutoring systems: To adapt to the students needs, so that better learning occurs
To adapt to the teachers needs, so better teaching takes place In commercial systems: To adapt to the customer, so that better(?) selling takes place Etc. TO ADAPT TO THE USER How? Simplest version: Include facts about the
user Adapt to known facts about the user Adapt to inferred properties of the user Has Eurocard ---> likes travelling Stereotypical user modelling Adaptivity example User: Could the student's mispronunciation errors be due to dialect? Response to parent: Yes, non-standard pronunciations may be due to dialect rather than poor decoding skills.
Response to psychologist: Yes, the student's background indicates the possibility of a dialectical difference. Stereotypes User modelling is always about guessing Early history Start: 1978, 79: Allen, Cohen & Perrault: Speech research for dialogue coherence
Elaine Rich: Building & Exploiting User Models (PhD thesis) 10 year period of developments UM performed by application system No clear distinction between UM components & system tasks mid 80s: Kobsa, Allgayer, etc. Distinction appears No reusability consideration
other Early systems GUMS (Finin & Drager, 1989; Kass 1991) General User Modelling System Stereotype hierarchies Stereotype members + rules about them
Consistency verification set framework for General UM systems Called UM shell systems (Kobsa) Academic developments Doppelgaenger [Orwant 1995] Hardware & software sensors Offers techniques of data generalization (linear prediction, Markov models, unsupervised clustering for stereotype formation)
TAGUS [Paiva & Self 1995] Stereotype hierarchy, inference mech., TMS, diagnostic system + misconception library Other UM shells um: UM Toolkit [Kay 1990, 1995, 1998] Represents assumptions on knowledge, beliefs, preferences (attribute value pairs) actually, library BGP-MS [Kobsa & Pohl 1995, Pohl 1998] Belief, Goal and Plan Maintenance System
Assumptions about users + user groups Allows multi-user, can work as network server LMS [Machado et al. 1999] Learner Modelling Server UM shell services (Kobsa 95) Representation of assumptions on user characteristic(s) E.g., knowledge, misconceptions, goals, plans, preferences, tasks, abilities
Representation of common characteristics of users Stereotypes, (sub-)groups, etc. Recording of user behaviour Past interaction w. system
formation of assumption based on interaction Generalization of interaction (histories) Stereotypes Inference engine Drawing new assumptions based on initial ones Current assumption + justification Evaluation of entries in current UM and comparison w. standards
Consistency maintenance UM shell services (Kobsa 95) UM shells Requirements Generality As many services as possible Concessions: student-adaptive tutoring systems Expressiveness
Able to express as many types of assumptions as possible (about U) Strong Inferential Capabilities AI, formal logic (predicate l., modal reasoning, reasoning w. uncertainty, conflict resolution) An Example of UM system David Benyon, 1993 Gerhard Fischer 1 HFA Lecture, OZCHI2000
Deep or shallow modelling? Deep models give more inferential power! Same knowledge can affect several parts of the functionality, or even several applications Better knowledge about how long an inference stays valid But deep models are more difficult to acquire Where do all the inference rules come from? How do we get information about the user? Aim of UM
Obtaining of U metal picture Vs. U behaviour modelled per se A Comparison between Adaptive and Adaptable Systems Gerhard Fischer 1 HFA Lecture, OZCHI2000 What can we adapt to?
User knowledge Cognitive properties (learning style, personality, etc.) User goals and plans User mood and emotions User preferences
Adaptation to User Knowledge U option knowledge about possible actions via an interface Conceptual knowledge that can be explained by the system Problem solving knowledge how knowledge can be applied to solve particular problems
misconceptions erroneous knowledge How can we infer user knowledge? Its in general hard to infer something about the users knowledge. Techniques used: Query the user (common in tutoring systems) Infer from user history (if youve seen an explanation, you understand the term) Rule-based generalisation based on domain structure (if you understand a specific term, you understand its generalisation)
Rule-based generalisation based on user role (if youre a technician, you should understand these terms) Bug libraries (recognise common errors) Generalization based on other similar users past The Model overlay technique Advantage: Simple and cheap Cannot model misconceptions, or new knowledge What can we adapt to?
User knowledge Cognitive properties (learning style, personality, etc.) User goals and plans
User mood and emotions User preferences Why model cognitive properties? Navigation in hypermedia: Very large differences (20:1) in performance, partially related to spatial ability. New tools needed! Learning: Different people require different learning styles (example / theory / group based)
Van der Veer et al. Kolb (1984) 2-D learning styles scale; 4 extreme cases: 1.converger (abstract, active):
2.diverger (concrete, reflective): concrete experience and reflective observation; great advantage in imaginative abilities, awareness of meanings and values, generating alternative hypotheses and ideas; question: "Why?" 3. assimilator (abstract, reflective):
abstract conceptualization and active experimentation; great advantage in traditional IQ tests, decision making, problem solving, practical applications of theories; knowledge organizing: hypothetical-deductive; question: "How?". abstract conceptualization and reflective observation; great advantage in inductive reasoning, creating theoretical models; focus more on logical soundness and preciseness of ideas; question: "What?". 4. accomodator (concrete, active):
concrete experience and active experimentation; focus on risk taking, opportunity seeking, action; solve problems in trial-and-error manner; question: "What if?". Inferring user cognitive characteristics The user does not know - not possible to ask! Stable properties - use lots of small signs over time. Studies required to establish correlation between indications and properties. A better solution may be to use these aspects of
user models only at design time (offer different interaction alternatives)? What can we adapt to? User knowledge Cognitive properties (learning style, personality, etc.)
User goals and plans User mood and emotions User preferences User Goals and Plans What is meant by this? A user goal is a situation that a user wants to achieve. A plan is a sequence of actions or event that the user expects will lead to the goal.
System can: Infer the users goal and suggest a plan Evaluate the users plan and suggest a better one Infer the users goal and automatically fulfil it (partially) Select information or options to user goal(s) (shortcut menus)
What information is available? Intended Plan Recognition: Limit the problem to recognizing plans that the user intends the system to recognize User does something that is characteristic for the plan Keyhole Plan Recognition: Search for plans that the user is not aware of that the system searches for. Obstructed Plan Recognition:
Search for plans while user is aware and obstructing Keyhole Plan Recognition Kautz & Allen 1990: Generalized plan recognition Hierarchical plan structures Method for inferring top-level actions from lower level observations. Axioms Abstraction:
Bottom down Cook-spaghetti Cook-pasta Decomposition: Top down Make-pasta-dish Preconditions, Effects, internal constraints, Make Noodles, Make Sauce, Boil
Intended Plan Recognition Used in Natural Language Interpretation. I want to take the eight oclock train to London. How do I get to platform four? Speaker intends to do that by taking the eight oclock train. Speaker believes that there is an eight oclock train to London. Speaker wants to get to London. Speaker believes that going to platform four will
help in taking the eight oclock train. Are these models useful? The keyhole case suffers from: Very little actual information from users Users that change their plans and goals The intended case suffers from: need of complex models of intentionality Multiple levels of plans plans for interaction, domain plans, plans for forming plans
Differences in knowledge between user and system Local plan recognition Make no difference between system and user plans (the keyhole case is limited to recognising plans that belong to the plan library anyway). Only domain (or one-level) plans. Be forgetful -- inferences based on latest actions. Let the user inspect and correct plans. Works best with probabilistic or heuristic methods.
What can we adapt to? User knowledge Cognitive properties (learning style, personality, etc.)
User goals and plans User mood and emotions User preferences Moods and emotions? New, relatively unexplored area! Unconscious level difficult to recognise, but it is possible to look at type speed, error rates / facial expressions, sweat, heartbeat rate... Conscious level can be guessed from task fulfilment (e.g. failures) Emotions affect the users cognitive capabilities
it can be important to affect the users emotions (e.g. reduce stress) Conscious and unconscious emotions Conscious Unconscious Emotional Modelling We address how emotions arise from an evaluation of the relationship between environmental events &
an agents plans and goals, as well as the impact of emotions on behaviour, in particular the impact on the physical expressions of emotional state through suitable choice of gestures & body language. Gratch, 5th Int. Conf. on Autonomous Agents, Montreal, Canada, 2001 Sample model of emotion assessment Conati, AAAI, North Falmouth, Massachusetts 2001
The layers in student modeling knowledge & cognitive model layer learning profile layer believability and emotional layer Abou-Jaoude & Frasson, AI-ED99, Le Mans, France, 1999 What can we adapt to?
User knowledge Cognitive properties (learning style, personality, etc.) User goals and plans
User mood and emotions User preferences Adaptation to user preferences So far, the most successful type of adaptation. Preferences can in turn be related to knowledge / goals / cognitive traits, but one needs not care about that. Examples:
Firefly www.amazon.com Mail filters Grundy (Rich: personalized book recommendation expert system) Inferring preferences Explicitly stated preferences (CNN News)
Matching the users behaviour towards the user group (Amazon) Matching the users behaviour towards rule base, and modify the rule base based on groups of users
(Grundy) Combining values from several stereotypes high value + high value high value + low value low value + low value
Adaptation model in Grundy The characteristic properties are those that have high or low value and high confidence. Choose a book that fits these. Describe those properties of the books that fit the users interests. Can the stereotypes be learned? Positive feedback -->
Increase certainty on key and property in all triggered stereotypes. Negative feedback --> Decrease certainty on key and property in all triggered stereotypes. No method to learn totally new stereotypes Preference models in general
Advantages: Simple models Users can inspect and modify the model Methods exist to learn stereotypes from groups of users (clustering)
Disadvantages: The Grundy model for stereotypes does not work in practice ==> machine learning! What can we adapt to? User knowledge
Cognitive properties (learning style, personality, etc.) User goals and plans User mood and emotions User preferences Generic User Modelling
Probability-based frameworks A decision theoretic framework Sub-symbolic techniques Example-based frameworks Rule-based frameworks Declarative Representation : BGP-MS(Kobsa): A User Modelling Shell A Hybrid Representation: SB-ONE Pure Logic Based
Rule-based adaptations Quantification (levels of expertise) Stereotypes (U classified) Overlay (actual use compared to ideal) Knowledge representation The system knowledge is partitioned into different
parts, System beliefs User beliefs Joint beliefs and more User goals
Stereotypes: can be activated if certain information is present. User Model Partitions Pros and Cons Very general and empty - difficult to use Truth Maintenance required (expensive) There are weights and thresholds, but not much theory behind those Learning from feedback not included
Frame-based frameworks E.g., semantic network Knowledge stored in structures w. slots to be filled Useful for small domain Network-based framework Knowledge represented in relationships between facts Can be used to link frames Statistical models, pros and cons
A theory exist for the calculations Usually requires training before usage (no learning from feedback) Weak representation of true knowledge Example: The MS Office assistant (the Lumire project) UM in Bayesian Networks Normally, relates observations to explanations Plan Inference, Error Diagnosis In Lumire, models the whole chain from observations to adaptation
The BN approach allows for a combination of declarative knowledge about structure with empirical knowledge about probabilities Lumire: Network Bayesian Nodes Observations Explanations as parameters in the user model Selection of adaptation help message
Selection of adaptation strategy active / passive help Lumire & Office helper High level problem structure Partial BN structure from Lumire Problems of BN in UM
Dealing with previous wrong guesses Dealing with changes over time Providing end-user inspection and control Advantages and Disadvantages
Explicit model of adaptation rules Not possible to learn new rules Rules could be taken from HCI literature BUT - there exist no such rules for adaptive behaviour! Possible to tune the adaptations based on feedback What should be tuned? User modelling or adaptation modelling? Example-based framework
Knowledge represented implicitly within decision structure Trained to classify rather than programmed w. rules Requires little knowledge aquisition Some Challenging Research Problems for User Modeling
identify user goals from low-level interactions - active help systems, data detectors - every wrong answer is the right answer to some other question integrate different modeling techniques - domain-orientation - explicit and implicit - give a user specific problems to solve capture the larger (often unarticulated) context and what users are doing (especially beyond the direct interaction with the computer system) - embedded communication - ubiquitous computing
reduce information overload by making information relevant - to the task at hand - to the assumed background knowledge of the users support differential descriptions (relate new information to information and concepts assumed to be known by the user) Gerhard Fischer 1 HFA Lecture, OZCHI2000 Commercial Boom (late 90s) E-commerce:
Product offering Sales promotion Product news Banners targeted to individual U Commercial Systems (2000) Group Lens (Net Perceptions)
Collaborative filtering alg. Explicit/implicit rating (navigational data) Transaction history LikeMinds (Andromedia) More modular architecture, load distribution Personalization Server (ATG) Rules to assign U to U groups (demographic data: gender, age) stereotype approach Frontmind (Manna)
Bayesian networks Learn Sesame (Open Sesame) Domain model: objects + attributes + events Clustering algorithms Characteristics of CS Client-server architecture for the WEB !!! Advantages: Central repository w. U info for 1/more applic. Info sharing between applications Complementary info from client DB integrated easily
Info stored non-redundant Consistency & coherence check possible Info on user groups maintained w. low redundancy (stereotypes, apriori or computed) Security, id, authentication, access control, encryption can be applied for protecting UM UM server UM server Services Comparison of U selective actions
Amazon: Customers who bought this book also bought:  Import of external U info ODBC(Open Database Connectivity) interfaces, or support for a variety of DB Privacy support Company privacy policies, industry, law UM server Requirements Quick adaptation
Preferably, at first interaction, to attract customers levels of adaptation, depending on data amount Extensibility To add own methods, other tools API for U info exchange Load balancing Reaction to increased load: e.g., CORBA based components, distributed on the Web Failover strategies (in case of breakdown)
Transaction Consistency Avoidance of inconsistencies, abnormal termination Conclusion: New UM server trends More recommender systems than real UM Based on network environments Less sophisticated UM, other issues (such as response time, privacy) are more important Separation of tasks is essential, to give flexibility: Not only system functions separately from UM functions, but also
UM functions separation: domain modelling, knowledge, cognitive modelling, goals and plans modelling, moods and emotion modelling, preferences modelling, and finally, interface related modelling In this way, the different levels of modelling can be added at different times, and by different people
Ch. 11 - Chemical Bonds I. Why Atoms Combine (p.298-302) Chemical Formula Chemical Bond Stability H2O 2 hydrogen atoms 1 oxygen atom A. Chemical Formula Shows: 1) elements in the compound 2) ratio of their atoms B. Chemical Bond Strong...
Clearance fit: The largest permissible diameter of the shaft is smaller thanthe diameter of the smallesthole. This type of fit always provides clearan Small clearances are provided for a precise fit that can easily be assembled withoutthe assistance of tools....
Candidates may progress from a simple narrative of causation/consequence with developed sequencing and reasoning supported by a range of accurate factual knowledge and understanding which might be related, for example to an analysis of how / why tension increased at...
Vitamin E Vitamin E and Selenium Selenium Biochemical role through glutathione peroxide Enhance neutrophil function Vitamin E and Selenium Both vitamin E and Se influences the function of immune cells especially in mammary gland immunity Cows are immuno suppressed when...
85% of flowers pollinated by insects (bees, butterflies, beetles, flies, ants); they get pollen and nectar. Birds and a few small mammals especially bats. Less pollen needs to be produced than in wind pollination. Flowers are bisexual—increases efficiency
Ready to download the document? Go ahead and hit continue!