Difference between revisions of "Publications/veyrin-forrer.22.dke"
From LRDE
(Created page with "{{Publication | published = true | date = 2022-10-26 | title = In pursuit of the hidden features of GNN's internal representations | journal = Data & Knowledge Engineering | p...") |
|||
(2 intermediate revisions by the same user not shown) | |||
Line 2: | Line 2: | ||
| published = true |
| published = true |
||
| date = 2022-10-26 |
| date = 2022-10-26 |
||
− | | title = In |
+ | | title = In Pursuit of the Hidden Features of GNN's Internal Representations |
| journal = Data & Knowledge Engineering |
| journal = Data & Knowledge Engineering |
||
− | | |
+ | | volume = 142 |
+ | | pages = 102097 |
||
⚫ | |||
| authors = Luca Veyrin-Forrer, Ataollah Kamal, Stefan Duffner, Marc Plantevit, Céline Robardet |
| authors = Luca Veyrin-Forrer, Ataollah Kamal, Stefan Duffner, Marc Plantevit, Céline Robardet |
||
⚫ | |||
| lrdekeywords = IA |
| lrdekeywords = IA |
||
| lrdenewsdate = 2022-10-26 |
| lrdenewsdate = 2022-10-26 |
||
+ | | publisher = Elsevier |
||
| abstract = We consider the problem of explaining Graph Neural Networks (GNNs). While most attempts aim at explaining the final decision of the model, we focus on the hidden layers to examine what the GNN actually captures and shed light on the hidden features built by the GNN. To that end, we first extract activation rules that identify sets of exceptionally co-activated neurons when classifying graphs in the same category. These rules define internal representations having a strong impact in the classification process. Then - this is the goal of the current paper - we interpret these rules by identifying a graph that is fully embedded in the related subspace identified by the rule. The graph search is based on a Monte Carlo Tree Search directed by a proximity measure between the graph embedding and the internal representation of the rule, as well as a realism factor that constrains the distribution of the labels of the graph to be similar to that observed on the dataset. Experiments including 6 real-world datasets and 3 baselines demonstrate that our method DISCERN generates realistic graphs of high quality which allows providing new insights into the respective GNN models. |
| abstract = We consider the problem of explaining Graph Neural Networks (GNNs). While most attempts aim at explaining the final decision of the model, we focus on the hidden layers to examine what the GNN actually captures and shed light on the hidden features built by the GNN. To that end, we first extract activation rules that identify sets of exceptionally co-activated neurons when classifying graphs in the same category. These rules define internal representations having a strong impact in the classification process. Then - this is the goal of the current paper - we interpret these rules by identifying a graph that is fully embedded in the related subspace identified by the rule. The graph search is based on a Monte Carlo Tree Search directed by a proximity measure between the graph embedding and the internal representation of the rule, as well as a realism factor that constrains the distribution of the labels of the graph to be similar to that observed on the dataset. Experiments including 6 real-world datasets and 3 baselines demonstrate that our method DISCERN generates realistic graphs of high quality which allows providing new insights into the respective GNN models. |
||
| type = article |
| type = article |
||
| id = veyrin-forrer.22.dke |
| id = veyrin-forrer.22.dke |
||
− | | identifier = doi: |
+ | | identifier = doi:10.1016/j.datak.2022.102097 |
| bibtex = |
| bibtex = |
||
@Article<nowiki>{</nowiki> veyrin-forrer.22.dke, |
@Article<nowiki>{</nowiki> veyrin-forrer.22.dke, |
||
− | title = <nowiki>{</nowiki>In |
+ | title = <nowiki>{</nowiki>In Pursuit of the Hidden Features of <nowiki>{</nowiki>GNN<nowiki>}</nowiki>'s Internal |
− | + | Representations<nowiki>}</nowiki>, |
|
journal = <nowiki>{</nowiki>Data \& Knowledge Engineering<nowiki>}</nowiki>, |
journal = <nowiki>{</nowiki>Data \& Knowledge Engineering<nowiki>}</nowiki>, |
||
− | + | volume = <nowiki>{</nowiki>142<nowiki>}</nowiki>, |
|
+ | pages = <nowiki>{</nowiki>102097<nowiki>}</nowiki>, |
||
year = <nowiki>{</nowiki>2022<nowiki>}</nowiki>, |
year = <nowiki>{</nowiki>2022<nowiki>}</nowiki>, |
||
+ | month = nov, |
||
issn = <nowiki>{</nowiki>0169-023X<nowiki>}</nowiki>, |
issn = <nowiki>{</nowiki>0169-023X<nowiki>}</nowiki>, |
||
− | doi = <nowiki>{</nowiki> |
+ | doi = <nowiki>{</nowiki>10.1016/j.datak.2022.102097<nowiki>}</nowiki>, |
− | url = <nowiki>{</nowiki>https://www.sciencedirect.com/science/article/pii/S0169023X2200088X<nowiki>}</nowiki>, |
||
author = <nowiki>{</nowiki>Luca Veyrin-Forrer and Ataollah Kamal and Stefan Duffner |
author = <nowiki>{</nowiki>Luca Veyrin-Forrer and Ataollah Kamal and Stefan Duffner |
||
and Marc Plantevit and C\'<nowiki>{</nowiki>e<nowiki>}</nowiki>line Robardet<nowiki>}</nowiki>, |
and Marc Plantevit and C\'<nowiki>{</nowiki>e<nowiki>}</nowiki>line Robardet<nowiki>}</nowiki>, |
||
keywords = <nowiki>{</nowiki>Graph Neural Networks, Explainable artificial |
keywords = <nowiki>{</nowiki>Graph Neural Networks, Explainable artificial |
||
intelligence, Monte Carlo Tree Search<nowiki>}</nowiki>, |
intelligence, Monte Carlo Tree Search<nowiki>}</nowiki>, |
||
+ | publisher = <nowiki>{</nowiki>Elsevier<nowiki>}</nowiki>, |
||
abstract = <nowiki>{</nowiki>We consider the problem of explaining Graph Neural |
abstract = <nowiki>{</nowiki>We consider the problem of explaining Graph Neural |
||
Networks (GNNs). While most attempts aim at explaining the |
Networks (GNNs). While most attempts aim at explaining the |
Latest revision as of 17:19, 14 December 2022
- Authors
- Luca Veyrin-Forrer, Ataollah Kamal, Stefan Duffner, Marc Plantevit, Céline Robardet
- Journal
- Data & Knowledge Engineering
- Type
- article
- Publisher
- Elsevier
- Keywords
- IA
- Date
- 2022-10-26
Abstract
We consider the problem of explaining Graph Neural Networks (GNNs). While most attempts aim at explaining the final decision of the model, we focus on the hidden layers to examine what the GNN actually captures and shed light on the hidden features built by the GNN. To that end, we first extract activation rules that identify sets of exceptionally co-activated neurons when classifying graphs in the same category. These rules define internal representations having a strong impact in the classification process. Then - this is the goal of the current paper - we interpret these rules by identifying a graph that is fully embedded in the related subspace identified by the rule. The graph search is based on a Monte Carlo Tree Search directed by a proximity measure between the graph embedding and the internal representation of the rule, as well as a realism factor that constrains the distribution of the labels of the graph to be similar to that observed on the dataset. Experiments including 6 real-world datasets and 3 baselines demonstrate that our method DISCERN generates realistic graphs of high quality which allows providing new insights into the respective GNN models.
Bibtex (lrde.bib)
@Article{ veyrin-forrer.22.dke, title = {In Pursuit of the Hidden Features of {GNN}'s Internal Representations}, journal = {Data \& Knowledge Engineering}, volume = {142}, pages = {102097}, year = {2022}, month = nov, issn = {0169-023X}, doi = {10.1016/j.datak.2022.102097}, author = {Luca Veyrin-Forrer and Ataollah Kamal and Stefan Duffner and Marc Plantevit and C\'{e}line Robardet}, keywords = {Graph Neural Networks, Explainable artificial intelligence, Monte Carlo Tree Search}, publisher = {Elsevier}, abstract = {We consider the problem of explaining Graph Neural Networks (GNNs). While most attempts aim at explaining the final decision of the model, we focus on the hidden layers to examine what the GNN actually captures and shed light on the hidden features built by the GNN. To that end, we first extract activation rules that identify sets of exceptionally co-activated neurons when classifying graphs in the same category. These rules define internal representations having a strong impact in the classification process. Then - this is the goal of the current paper - we interpret these rules by identifying a graph that is fully embedded in the related subspace identified by the rule. The graph search is based on a Monte Carlo Tree Search directed by a proximity measure between the graph embedding and the internal representation of the rule, as well as a realism factor that constrains the distribution of the labels of the graph to be similar to that observed on the dataset. Experiments including 6 real-world datasets and 3 baselines demonstrate that our method DISCERN generates realistic graphs of high quality which allows providing new insights into the respective GNN models. } }