The 2nd Workshop on Trustworthy Learning on Graphs (TrustLOG)

Colocated with the Web Conference 2024

About

Learning on graphs (LOG) is fundamental and essential to a wide range of web applications, including information retrieval, social network analysis, computational chemistry, and intelligent transportation. Researchers in the area have made significant contributions to theories, algorithms, and open-source systems tailored to address diverse learning tasks. State-of-the-art LOG algorithms have demonstrated superior empirical performance in answering what the optimal learning results would be to serve downstream web applications. For instance, what are the most relevant web pages based on user queries? What are the best groupings of online users to form online communities? What items should recommender systems offer to fit user preferences and to stimulate the growth of e-commerce platforms? Despite the substantial progress in developing high-utility LOG algorithms, recent research has raised concerns about the algorithmic trustworthiness in terms of several critical social aspects, including fairness, transparency, privacy, and security. Black-box LOG algorithms are also found to be vulnerable to malicious attacks, biased against individuals from certain demographic groups, or insecure to information leakage. The untrustworthiness would further limit the potential of LOG algorithm to be deployed in high-impact domains such as online banking and digital health. Therefore, it is essential to ask: why are LOG algorithms untrustworthy? And how can we develop and deploy trustworthy LOG algorithms? To answer these questions, it is essential to introduce a paradigm shift, from solely addressing what questions to understanding how and why. Such paradigm shift could benefit a wide range of web technologies, including fair and reliable anomalies, robust webpage ranking, fair information campaign on social media, as well as factual and private share of knowledge on web platform.

There are several key research challenges involved in trustworthy learning on graphs, including:

  • Understanding the theoretical implications of non-IID graph on the classic trustworthy machine learning;
  • Discovering graph-specific measurements and techniques for trustworthy learning;
  • Achieving trustworthy learning on graphs at scale;
  • Accommodating the heterogeneity of graph data;
  • Dealing with dynamically changing and/or temporal graphs.

Building upon the success of its previous edition, this one-day workshop (TrustLOG-WWW'24) aims to bring together researchers and practitioners from different backgrounds to study these key challenges and enhance the trustworthiness of learning on graphs. The workshop will consist of invited talks, discussion panels, contributed talks, contributed posters on a wide variety of methods and applications related to trustworthy learning on graphs. The proposed workshop welcomes contributions in various format, including research papers, benchmarks and datasets, vision papers, position papers, and white papers. The TrustLOG workshop intends to share visions of investigating new approaches at the intersection of trustworthy learning on graphs and real-world applications.

Agenda (Tentative)

Time Event
9:00 AM~9:10 AM Opening Remarks
9:10 AM~9:55 AM Invited Talk Speaker: Dr. Leman Akoglu
9:55 AM~10:40 AM Invited Talk Speaker: Dr. Tina Eliassi-Rad
10:40 AM~11:00 AM Coffee Break
11:00 AM~11:45 AM Oral Presentations for Accepted Papers
11:45 AM~12:30 PM Invited Talk Speaker: Dr. Vagelis Papalexakis
12:30 PM~13:30 PM Lunch Break
13:30 PM~14:15 PM Invited Talk Speaker: Dr. Xiang Wang
14:15 PM~15:00 PM Panel: The Next Frontiers of Trustworthy Learning on Graphs Panelists: Dr. Hanghang Tong, Dr. Bryan Hooi, Dr. Xiang Wang
15:00 PM~15:15 PM Award Ceremony and Closing Remarks

Keynote Speakers

Dr. Leman Akoglu

Dean's Associate Professor

Carnegie Mellon University

Title: Detecting Real-world Anomalies with Decorated Graphs

Abstract: Given a database of decorated graphs, with node- and edge-attributes, edge multiplicities and directions, and associated metadata for each graph, how can we spot the anomalous ones? Several real-world problems can be cast as graph inference tasks where the topology captures complex relational phenomena (e.g., transactions between financial accounts in a journal entry), along with metadata reflecting tabular features (e.g. approver, effective date, etc.), where the simultaneous handling of relational and tabular features poses a challenge.

In this talk I will present a novel graph neural network model that handles directed multigraphs, providing a unified end-to-end architecture that fuses metadata and graph-level representation learning through an unsupervised anomaly detection objective. Experiments in two domains, based on general-ledger journal entries from different firms (accounting) and human GPS trajectories from thousands of individuals (urban mobility), validate the generality and detection effectiveness on expert-guided ground-truth anomalies. We also find that a built-in synthesis of the two data modalities (graph and metadata) outperforms the baselines that handle these separately with post-hoc synthesis efforts.

Dr. Tina Eliassi-Rad

Joseph E. Aoun Professor

Northeastern University

Title: Trustworthy Network Science

Abstract: As the use of machine learning (ML) algorithms in network science increases, so do the problems related to explainability, transparency, fairness, privacy, and robustness, to name a few. In this talk, I will give a brief overview of the field and present recent work from my lab on the (in)stability and explainability of node embeddings, attacks on ML algorithms for graphs, and equality in complex networks.

Dr. Vagelis Papalexakis

Associate Professor

University of California Riverside

Title: Low-rank Approximation and Robustness

Abstract: In this talk we are going to explore the connections between low-rank approximation and robustness. In particular, focusing on graph mining and learning as our motivating application, we are going to discuss recent results where we demonstrate the power of low-rank methods in (a) consolidating and improving the performance of potentially noisy node embeddings, and (b) defending against adversarial attacks. We will conclude with results that point to the generality of our observations and discuss future directions.

Dr. Xiang Wang

Professor

University of Science and Technology of China

Title: Making Large Language Models Aligned with Graph

Abstract: The burgeoning field of AI has seen significant advancements in the capabilities of large language models (LLMs). However, applying LLMs to graph-centric domains requires novel approaches that bridge the language space with complex, structured graph data. I will share our recent work in AI4Science, which focuses on integrating LLMs with graph-based data representations (e.g., 2D molecular graphs, 3D molecular point clouds, and graph-graph interactions). These works collectively propel forward the frontier of LLM-driven graph understanding and generation, offering new tools for scientists to harness the power of AI in their research. This talk will detail the paradigm, challenges, and potential impacts of aligning LLMs with scientific graphs, setting the stage for future innovations in AI-driven scientific exploration.

Panel

Topic: The Next Frontiers of Trustworthy Learning on Graphs

Panelists

Contributions

Best Paper

  • Smooth Anonymity for Sparse Graphs
    Hossein Esfandiari, Alessandro Epasto, Vahab Mirrokni and Andres Munoz Medina

Best Paper Runner-up

  • Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks
    Xu Zheng, Farhad Shirani, Tianchun Wang, Wei Cheng, Zhuomin Chen, Haifeng Chen, Hua Wei and Dongsheng Luo

Accepted Papers

  • Smooth Anonymity for Sparse Graphs
    Hossein Esfandiari, Alessandro Epasto, Vahab Mirrokni and Andres Munoz Medina
  • Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks
    Xu Zheng, Farhad Shirani, Tianchun Wang, Wei Cheng, Zhuomin Chen, Haifeng Chen, Hua Wei and Dongsheng Luo
  • Interpreting Graph Neural Networks with In-Distributed Proxies
    Zhuomin Chen, Jiaxing Zhang, Jingchao Ni, Xiaoting Li, Yuchen Bian, Md Mezbahul Isam, Ananda Mohan Mondal, Hua Wei and Dongsheng Luo
  • Structural Group Unfairness: Measurement and Mitigation by means of the Effective Resistance
    Adrian Arnaiz-Rodriguez, Georgina Curto and Nuria Oliver
  • ToDA: Target-oriented Diffusion Attacker against Recommendation System
    Xiaohao Liu, Zhulin Tao, Ting Jiang, He Chang, Yunshan Ma, Xianglin Huang and Xiang Wang

Call for Papers

We invite contributions to the Trustworthy Learning on Graphs (TrustLOG) co-located with The Web Conference 2024 (formerly known as WWW). The workshop will take place in Singapore, May 13 - 14, 2024.

Important Dates

  • Paper submission: February 13, 2024
  • Reviews period: February 15 - February 26, 2024
  • Final notification: March 4, 2024
  • Camera-ready submission: March 11, 2024
  • Workshop dates: May 13 - 14, 2024

Submission Site

We use EasyChair to manage the submission and review. Abstracts and papers can be submitted through the following link: https://easychair.org/conferences/?conf=thewebconf2024_workshops. Please select TrustLOG: The Second Workshop on Trustworthy Learning on Graphs to submit to our workshop.

Scope

We invite submissions on a broad range of trustworthy learning on graphs. The topics of interest include (but are not limited to):

  • Robustness, explainability, fairness, reliability, safety, and social norm of graph learning
  • Environmental well-being of graph learning methods
  • Risks and limitations of graph learning methods and foundation models (e.g., LLMs, LMMs) on graphs
  • Applications of trustworthy learning on graphs (e.g., recommender system, knowledge graph, social network analysis, drug discovery, material design, etc.)
  • Datasets and benchmarks for trustworthy learning on graphs

Various types of contributions are welcomed, such as (but are not limited to):

  • Extended abstract
  • Research paper
  • Work-in-progress paper
  • Demo paper
  • Visionary papers/white paper
  • Appraisal papers of existing methods or toolboxes
  • Evaluatory papers on assumptions, methods or toolboxes
  • Relevant work that will be or have been published

Submission Guidelines

Anonymity. The review process will be double-blind. The submitted document should omit any author names, affiliations, or other identifying information. This may include, but is not restricted to acknowledgments, self-citations, references to prior work by the author(s), and so on. Please use the third person to identify your own prior work. You may explicitly refer in the paper to organizations that provided datasets, hosted experiments, or deployed solutions and tools.

Formatting Requirements. Submissions must be a single PDF file: 8 (eight) pages as the main paper, with unlimited pages for references and an optional Appendix (that can contain details on reproducibility, proofs, pseudo-code, etc).

Submissions must be in English, in double-column format, and must adhere to the ACM template and format (also available in Overleaf). Word users may use the Word Interim Template and the recommended setting for LaTeX is:

\documentclass\[sigconf, anonymous, review\]{acmart}.

Originality and Concurrent Submissions. Accepted papers at the workshop are optional to be included in the Companion Proceedings of The Web Conference 2024.

  • Opt-in for the Companion Proceedings if accepted: Submissions that are under review or published/accepted to any peer-reviewed conference/journal with published proceedings cannot be submitted. Submissions that have been previously presented orally, as posters or abstracts-only, or in non-archival venues with no formal proceedings, including workshops or PhD symposia without proceedings, are allowed.
  • Opt-out for the Companion Proceedings if accepted: We allow submissions that are under review at or published/accepted to any preprint servers (e.g., arXiv) and/or peer-reviewed conference/journal with published proceedings.

Authors may submit anonymized work that is already available as a preprint (e.g., on arXiv or SSRN) without citing it. The ACM has a strict policy against plagiarism, misrepresentation, and falsification that applies to all publications.

Ethical Use of Data and Informed Consent. Authors are encouraged to include a section on the ethical use of data and/or informed consent of research subjects in their paper, when appropriate. You and your co-authors are subject to all ACM Publications Policies, including ACM's Publications Policy on Research Involving Human Participants and Subjects (posted in 2021). Please ensure all authors are familiar with these policies.

Please consult the regulations of your institution(s) indicating when a review by an Institutional Ethics Review Board (IRB) is needed. Note that submitting your research for approval by such may not always be sufficient. Even if such research has been approved by your IRB, the program committee might raise additional concerns about the ethical implications of the work and include these concerns in its review.

Organization

Organzing Chairs

Publicity Chair

Contact

For questions, please contact us at trustlogworkshoporganizers@gmail.com.