NEXT:
2 July 2024 | h 15:30 | @Aula Magistrale 3rd floor

Talk 1

Assistive technologies supporting people with visual impairment during navigation

Gabriele Galimberti

People with visual disabilities face difficulties in moving around due to the lack of sight. To overcome this, some visually impaired individuals use mobile applications that provide useful directions for navigating unfamiliar environments. Gabriele Galimberti's doctoral research is focused on studying position calculation systems based on computer vision applied to assistive technologies for people with visual impairments. Additionally, his work explores sustainable mobility, addressing the challenges faced by people with visual disabilities during mobility. In the last year, the research activity has focused mainly on the design and testing of a navigation system that uses an advanced space representation model and navigation instruction processing to guide a person with blindness along a path. The proposed system addresses some common challenges encountered during navigation focusing on the modeling of the space domain and on the navigation instruction processing. Primarily, the navigation system seeks to reduce "parsimonious guidance," (i.e., the number of instructions provided to the participant during navigation) through a layered division of the system and targeted design choices. An innovative feature introduced in the system is the ability to avoid reaching an area around a current target point where a turn must be made along a path. This can happen using a graph space representation that allows moving from the current target point to the next target point without having to pass through the current target point. Additionally, the system is designed to offer a more flexible direction of movement, preventing the user from being limited by a precise direction. This is possible by providing a range of directions that do not require continuous adjustments. A preliminary evaluation of data collected with 8 visually impaired participants in a navigation test reveals that the proposed system, compared with a state-of-the-art navigation system, gets the user to their destination in less time, reduces parsimonious guidance, and it is more appreciated by users than the state-of-the-art navigation system. This work was conducted in collaboration with Smith-Kettlewell Eye Research Institute of San Francisco. Concurrently, the doctoral candidate also conducted a study on the sustainable mobility of people with visual impairments through semi-structured interviews with 17 individuals with visual impairments. The themes emerging from the interviews primarily concern the use of apps in mobility, behavior in daily situations during mobility, and how sustainable mobility impacts the lives of people with visual impairments.

Talk 2

Graph Machine Learning for temporal heterogeneous networks

Manuel Dileo

In my research activity talk, I will present some of our recent works in the field of graph machine learning for temporal heterogeneous networks (THNs). Specifically, we will focus on Steemit, one of the most well-known and used Web3 social platforms, which can be quite naturally represented as a THN. I will describe a framework based on discrete-time graph neural networks and BERT text encoders to answer the following research questions: How do we handle dynamic textual and structural information to predict links in online social networks? How do textual features impact social relationships and economic transactions over time? The results show the effectiveness of our framework and offer valuable insights into the characteristics of these platforms, highlighting an interplay between users' social and economic relationships, and showing that users' interests play a role in financial transactions. In the second part of the talk, I will briefly discuss my latest works about understanding graph machine learning models through the lens of network science, highlighting the importance of considering network science research when dealing with temporal graph learning, to push the research community toward better models for link prediction.