The GraphNEx team gathered at the Idiap Research Institute in Martigny, Switzerland, surrounded by the beautiful Alps for an intense 2-day workshop with warm and sunny Spring weather of early May. The workshop was an opportunity for the team to meet up again in person, bond and strengthen previous relationships, and discuss the next steps of the project.

The focus was on progressing with a common software architecture for the two use-case: privacy protection and system genetics. The software should be flexible and modular in order to enable different users to train explainable predictive models on their own data.

During the workshop, the team updated the overall workflow that was agreed in the first in-person GraphNEx workshop and that consists of four modules: data, model, explainability and interface. Both model and explainability are complemented by their own evaluation toolkit, whereas the interface allows users and experts to visualise explanations on the browser.

From general discussions and presentations to focused sub-groups activities, the GraphNEx team defined the steps to create a repository for the development of the common architecture and interface. Here is a summary of the main decisions regarding the modules to be developed.

• Datasets commonly used by each research community, such as the Breast Cancer dataset from The Cancer Genome Atlas and the Image Privacy dataset were selected for prototyping and testing. The team will curate the datasets in the same format to be used by the models.
• Several models, such as graph convolutional networks and graph attention networks, were identified to be integrated in the common architecture. Additionally, the team discussed integrating simpler models as reference for comparing learning performance and explainability. Challenges exist to train and test the models on the previously mentioned datasets due to the different nature of the problems. For instance, separate pre-processing steps should be included to avoid end-to-end pipelines that would become specific for a use case, breaking the common and modular principles behind the workshop.
• For the interface, the team members agreed on the need to visualise the predictions and explanations of the models at different resolutions. In addition, discussions emerged on an interactive component that would allow the user to update the model / explanations in the back-end.

Stay tuned to know more about our upcoming repository with explainable models for privacy and system genetics!

View of the mountains in Martigny the day of the workshop

This year the CHIST-ERA annual projects seminar was back in-person and held in Bratislava, Slovakia on 4-5 April. The event brought together researchers involved in the ongoing CHIST-ERA funded projects and representatives of their research funding organisations. CHIST-ERA Projects Seminar aims at strengthening the scientific communities working in the topics of the CHIST-ERA calls and supporting effort toward new collaborative and interdisciplinary research. The GraphNEx team members contributed to the event with three posters, a project presentation, flyers and a video demo.

The GraphNEx poster covered the objectives and challenges, the two use cases (System Genetics and Privacy Protection), the achievements, and the events organised to disseminate our works and to engage with the general public over the last year.

Denis and Basile (EPFL) presented the Explainable Value Proposition Canvas (xVPC) that originated from the GraphNEx workshop in 2022 and finalised in a peer-reviewed paper published in the IEEE TALE. xVPC borrows from the value proposition canvas broadly used by entrepreneurs and intrapreneurs to create new services and products. The GraphNEx canvas is applied for collaborative activities with blended design thinking in education. The canvas also supports the user-centred design of interfaces and helps the average citizen understand what private information they disclosed when sharing pictures online. To know more about our xVPC check out the poster here.

Myriam (ENSL) presented the limits of current explainability techniques for classification tasks on gene expression data. This is joint work with Anaïs and Maria (EPFL). Check out the poster for more details.

The CHIST-ERA Projects Seminar was a great opportunity for us to interact in person, to discuss current achievements, and plan collaborations. All of this without losing enjoyable walks and meals with traditional Slovakian food around the beautiful city centre of Bratislava and the stunning castle overlooking the Danube!

A GraphNEx demo on Privacy Awareness was presented at the Victoria and Albert museum during the Friday Late event on 23 September 2022 and the Digital Design Weekend on 24-25 September 2022. In both events, Prof Andrea Cavallaro and Dr Riccardo Mazzon (QMUL) in collaboartion with Prof Laura Ferrarello and Rute Fiadeiro (Royal College of Art, London) showcased “You and Your Images” an interactive activity designed to increase the knowledge and literacy about the consequences of disclosing private information in images shared on Social Media.

Both events were well attended by an estimation of over 2000 participants in total. Over 100 participants took part in the You and Your Images activity that involved going through a web experience where, after answering a set of questions, the participant had the possibility to upload an image, receive back a set of keywords describing the image and grasp how these information may be exploited by a service provider.

From the young professionals to the experts on data privacy, we received multiple positive feedback from the participants as the activity helped people to better understand how much information images are revealing and how, possibily, this information can be used.

You and Your Images website   Friday Late   Digital Design Weekend 2022

Can we extrapolate semantic concepts and extract meaningful relationships that promote human interpretability from the current advances in explainable artificial intelligence (XAI)?

The GraphNEx team gathered on a three intense days' workshop (11-13 May 2022) near the beautiful lake Léman in Lausanne, Switzerland, to generate and share ideas around this question. The team focussed on i) defining explainability in terms of user expectations, ii) outlining a value proposition canvas (adapted from business terminology), iii) designing an explainability protocol to quantify the relevance of the explanations, and iv) reflecting on the added value of a-priori information.

The elements of an explainability protocol to quantify the relevance of the explanations were then defined. Explanations should be simple, stable and robust to missing data. They should also be domain appropriate and subject dependent (user-centred). A thorough analysis on the added value of a priori information, used to impose rules on the model and induce more constraints that are present in the data, was properly undertaken. The team noticeably enriched by the experience, used the final workshop session to construct the workflow of the project and set the main milestones to progress towards fulfilling users' expectations of XAI while planning for the next in-person meeting.