Steve Azzolin's Avatar

Steve Azzolin

@steveazzolin.bsky.social

ELLIS PhD student @ UNITN/UniCambridge || Prev. Visiting Research Student at UniCambridge || Prev. Research intern at SISLab

185 Followers  |  122 Following  |  14 Posts  |  Joined: 02.12.2024  |  1.8914

Latest posts by steveazzolin.bsky.social on Bluesky

Beyond Topological Self-Explainable GNNs: A Formal Explainability... Self-Explainable Graph Neural Networks (SE-GNNs) are popular explainable-by-design GNNs, but their explanations' properties and limitations are not well understood. Our first contribution fills...

Paper: openreview.net/forum?id=mkq...

13.07.2025 17:29 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0

๐Ÿ’กWhat can we do to make self-explanations less ambiguous?

-> We propose to automatically adapt explanations to the task by stitching together SE-GNNs with white-box models and combining their explanations.

13.07.2025 17:29 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

- Self-explanations can be "unfaithful" by design

13.07.2025 17:29 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

- Models encoding different tasks can produce the same self-explanations, limiting the usefulness of explanations

13.07.2025 17:29 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

Studying some popular models, we found that:

- The information that self-explanations convey can radically change based on the underlying task to be explained, which is, however, generally unknown

13.07.2025 17:29 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0
Post image

๐Ÿค” What are the properties of self-explanations in GNNs? What can we expect from them?

We investigate this in our #ICML25 paper.

Come to have a chat at poster session 5, Thu 17 11 am.

w. Sagar Malhotra @andreapasspr.bsky.social @looselycorrect.bsky.social

13.07.2025 17:29 โ€” ๐Ÿ‘ 1    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

Happening tomorrow!

Poster number 508
Saturday's session 10-12:30

25.04.2025 14:27 โ€” ๐Ÿ‘ 0    ๐Ÿ” 1    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0

3. ITS ROLE IN OOD GENERALISATION

Domain-Invariant GNNs make predictions over a domain-invariant subgraph to achieve OOD generalisation. We show that unless this subgraph is also *sufficient*, DIGNNs are not domain-invariant.

5/5

17.04.2025 13:43 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0

2. HOW GNNs AIM TO ACHIEVE IT

We highlight several architectural design choices of Self-Explainable GNNs favoring information leakage from nodes outside the explanation, and propose mitigations.

4/5

17.04.2025 13:43 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

We propose rethinking faithfulness from three essential angles:

1. HOW TO COMPUTE IT

Many ways to compute faithfulness exists, but we show:

- they are not interchangeable
- some of them do not have the desired semantics

3/5

17.04.2025 13:43 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

Paper: "Reconsidering Faithfulness in Regular, Self-Explainable and Domain Invariant GNNs"
Link: openreview.net/forum?id=kiO...
Poster session: 26 April 10am

2/5

17.04.2025 13:43 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0
Post image

Faithfulness of GNN explanations isnโ€™t one-size-fits-all๐Ÿงข
Our last @iclr-conf.bsky.social paper breaks it down across:

1. Evaluation metrics
2. Model implementations
3. OOD generalisation

w: Antonio L. @looselycorrect.bsky.social @andreapasserini.bsky.social

1/5

17.04.2025 13:43 โ€” ๐Ÿ‘ 5    ๐Ÿ” 3    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 1

Kudos to the organisers for setting up the poster session in the fanciest room I've ever seen๐Ÿ‘€

08.12.2024 21:26 โ€” ๐Ÿ‘ 3    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

Hello World!

02.12.2024 09:45 โ€” ๐Ÿ‘ 7    ๐Ÿ” 0    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 0

@steveazzolin is following 20 prominent accounts