BEGIN:VCALENDAR
VERSION:2.0
PRODID:IEEE vTools.Events//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:Asia/Shanghai
BEGIN:STANDARD
DTSTART:19910915T010000
TZOFFSETFROM:+0900
TZOFFSETTO:+0800
TZNAME:CST
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20261001T025345Z
UID:66794A91-4365-470B-97ED-647378116E16
DTSTART;TZID=Asia/Shanghai:20261012T150000
DTEND;TZID=Asia/Shanghai:20261012T160000
DESCRIPTION:Graph Neural Networks rely on graph structure as a fundamental 
 inductive bias\, yet in many real-world applications the underlying relati
 onal structure is unknown\, incomplete\, or potentially inaccurate. Graph 
 Structure Learning (GSL) addresses this challenge by inferring relational 
 structures directly from data and optimizing them for the downstream task.
  This seminar provides an overview of the main approaches to GSL\, focusin
 g on structure initialization\, edge scoring\, graph sparsification\, and 
 loss design. We discuss several strategies for constructing an initial gra
 ph\, including Pearson correlation\, Granger causality\, pairwise input si
 milarity\, Dirichlet energy minimization\, and graph rewiring. Different e
 dge-scoring approaches are then considered\, ranging from input-independen
 t parameterizations to input-dependent models based on neural networks and
  attention mechanisms.\nParticular attention is given to the challenges of
  learning sparse and discrete graph structures\, including thresholding an
 d Bernoulli sampling\, as well as the gradient-estimation techniques requi
 red for their optimization. Finally\, we discuss how downstream objectives
  and regularization terms can guide the learned graph toward structures th
 at are both task-relevant and consistent with desired properties. Overall\
 , the seminar highlights the potential of GSL as an alternative to predefi
 ned graph topologies\, while emphasizing the trade-offs between expressive
 ness\, computational efficiency\, and structural constraints.\n\nSpeaker(s
 ): Cesare Alippi\n\nRoom: 418室\, Bldg: 工二\, 大学城广东工业大
 学\, Guangzhou\, Guangdong\, China
LOCATION:Room: 418室\, Bldg: 工二\, 大学城广东工业大学\, Guangz
 hou\, Guangdong\, China
ORGANIZER:cschenwn@scut.edu.cn
SEQUENCE:11
SUMMARY:Graph Structure Learning
URL;VALUE=URI:https://events.vtools.ieee.org/m/581240
X-ALT-DESC:Description: &lt;br /&gt;&lt;p class=&quot;MsoNormal&quot; style=&quot;text-align: justi
 fy\; text-justify: inter-ideograph\;&quot;&gt;&lt;span lang=&quot;EN-US&quot; style=&quot;mso-ansi-l
 anguage: EN-US\;&quot;&gt;Graph Neural Networks rely on graph structure as a funda
 mental inductive bias\, yet in many real-world applications the underlying
  relational structure is unknown\, incomplete\, or potentially inaccurate.
  Graph Structure Learning (GSL) addresses this challenge by inferring rela
 tional structures directly from data and optimizing them for the downstrea
 m task. This seminar provides an overview of the main approaches to GSL\, 
 focusing on structure initialization\, edge scoring\, graph sparsification
 \, and loss design. We discuss several strategies for constructing an init
 ial graph\, including Pearson correlation\, Granger causality\, pairwise i
 nput similarity\, Dirichlet energy minimization\, and graph rewiring. Diff
 erent edge-scoring approaches are then considered\, ranging from input-ind
 ependent parameterizations to input-dependent models based on neural netwo
 rks and attention mechanisms.&lt;br&gt;Particular attention is given to the chal
 lenges of learning sparse and discrete graph structures\, including thresh
 olding and Bernoulli sampling\, as well as the gradient-estimation techniq
 ues required for their optimization. Finally\, we discuss how downstream o
 bjectives and regularization terms can guide the learned graph toward stru
 ctures that are both task-relevant and consistent with desired properties.
  Overall\, the seminar highlights the potential of GSL as an alternative t
 o predefined graph topologies\, while emphasizing the trade-offs between e
 xpressiveness\, computational efficiency\, and structural constraints.&lt;/sp
 an&gt;&lt;/p&gt;
END:VEVENT
END:VCALENDAR

