Workshop Outcomes

“Putting the Control back with the Humans”

Panel research outcome initially triggered in relation to the presented papers

How to realize user control
  • Whether the users can check correctness of understanding of their actions (creativity phases; engagement with video watching for learning; personality traits; the record held about them);
  • Ability to choose what data to provide, and for how long;
  • Users can change the pace of the interaction with the system by changing their behavior;
  • To give the user the choice of reverting the decision made by the adaptation;
  • User for car personal assistant can still drive/ passengers and their presence;
  • Can choose to switch off certain adaptation;
  • Can decline to give mandatory data by explaining why
What is missing
  • Providing a playback of users’ interactions;
  • The relationship between the interaction and the type of content;
  • Need an environment to collect constant feedback from user in the interactions (in the context of time and space);
  • Need a meta understanding of interactions (open model, glass box model);
  • The current technology is quite limited;
  • To identify what game applies in this maintenance of equilibrium states of interaction
Way forward (solutions)
  • Should be able to reveal the model underneath and give the power to users for influencing the model and decisions taken;
  • The decision behind the pedagogical interventions – providing the answer to why;
  • Better movement capture while interacting with the system and how it relates to the elements of the games;
  • More simulation studies (e.g. with health and safety in mind);
  • Need progressive enhancement for missing data (e.g. not provide income, then check if that can be deduced from other sources);
  • How to enable the users to trust the system by explaining the adaptation strategies (and also address the privacy issues);
  • To devise a simulation environment to study these challenges of ESI with a set of small but multidimensional data

Interactive session in working groups

Research outcome of brainstorming on current challenges

(a) Privacy in ethics and personalization
  • Research vs. general purpose ethics (as researchers we are, and have been for a while, covered by ethical approval issued by institutional ethics boards, which already cover most aspects of GDPR)
  • Issues with data from modalities like audio/video – how do we anonymise these?
  • Variety of data that falls under personally identifiable information (PII) and other ‘sensitive’ information
  • Misinformation using PII (e.g. targeting specific demographics during elections)
  • Under GDPR you can’t have algorithms making important decisions regarding people (e.g. should someone be approved for mortgage)
    • explainable AI (XAI) could help make this process both scalable, but still fair, transparent, and accountable
  • There is still a lot of uncertainty around the enforcement of GDPR; as such we are not sure how strictly the rules will be enforced, and what will this mean for UMAP research
  • Functional algorithms (~ functional cookies, i.e. the cookies which are necessary for the functioning of the website; if the user doesn’t consent to using functional cookies, then they can’t use the website), which obtain informed consent (but is it actually informed?)
  • Progressive enhancement (users can opt out of giving certain data, but the algorithms and system would still need to provide fair access and usage even with such missing data) – is this workable?
(b) Beyond recommendations: New adaptation methods, in mixed realities
  • Interacting with the same entity (e.g., service) across various realities (platforms and contexts)
    • Recommendations (+ data processing, handling) coming from central point of access
    • Intelligent digital assistance
    • o Models at higher levels of abstraction capturing the holistic user actions
  • Agent systems providing recommendations (more obvious control on the agent, e.g., automate the process and actions and then approve or not, e.g., automotive area)
  • Degree of control applicability highly dependent on the domain of activity
  • User models that consider users’ learning curve for recommending
  • How recommendations may differ in regard to more “physical” interactions e.g., augmented reality