TL;DR
DURENDAL is a flexible graph deep learning framework designed for temporal heterogeneous networks, enabling better modeling of evolving complex systems and introducing new datasets for benchmarking.
Contribution
It proposes a novel framework that adapts heterogeneous graph models to evolving networks and introduces two new high-resolution temporal datasets for evaluation.
Findings
DURENDAL outperforms existing methods in future link prediction tasks.
The framework effectively captures the dynamics of evolving heterogeneous networks.
New datasets enhance benchmarking for temporal heterogeneous graph learning.
Abstract
Temporal heterogeneous networks (THNs) are evolving networks that characterize many real-world applications such as citation and events networks, recommender systems, and knowledge graphs. Although different Graph Neural Networks (GNNs) have been successfully applied to dynamic graphs, most of them only support homogeneous graphs or suffer from model design heavily influenced by specific THNs prediction tasks. Furthermore, there is a lack of temporal heterogeneous networked data in current standard graph benchmark datasets. Hence, in this work, we propose DURENDAL, a graph deep learning framework for THNs. DURENDAL can help to easily repurpose any heterogeneous graph learning model to evolving networks by combining design principles from snapshot-based and multirelational message-passing graph learning models. We introduce two different schemes to update embedding representations for…
Peer Reviews
Decision·ICLR 2024 Conference Withdrawn Submission
1. Benchmarking dynamic heterogeneous graphs is important. 2. Two datasets are introduced by transforming the original open datasets. 3. Experiments on performed.
1. The mechanism of why DURENDAL outperforms baselines is unclear. 2. The comparison of UTA and ATU is not clear. System-level (e.g. run time, memory usage) evaluation might be helpful. 3. More commonly used datasets are needed if the paper wants to be a benchmark paper (e.g. Open Academic Graph).
1. The designed framework is generic and can be integrated with any static heterogeneous GNNs. Given its simplicity and wide adaptivity, it can facilitates future research on dynamic heterogeneous graph learning. 2. This work introduces two new benchmark datasets of dynamic heterogeneous graphs, including one dataset from e-commerce recommendation and one dataset from blockchain-based online social network. Specifically, the TaobaoTH is of a relatively large size with ~360k nodes. 3. The designe
1. Based on my understanding of the differences between dynamic graphs and temporal graphs, I think it would be better if this work is positioned for dynamic heterogeneous networks instead of temporal heterogeneous networks. Dynamic networks are snapshot-based networks, i.e., aggregating edges and nodes within certain time windows, which is exactly what this paper considers. In contrast, temporal networks are more dynamically changing where each edge is associated with a timestamp (not a snapsho
The paper is overall well written, easy to follow and with good references for people that might be approaching the field of dynamic graphs for the first time. While the approach appears rather straightforward (especially when compared to ROLAND), experimental results look promising on the considered dataset. The introduction of new datasets is also something that the community will most likely benefit from.
As it might have emerged from my comments in the “Strengths” section, the approach appears to be a not particularly original improvement over ROLAND (unless my understanding is wrong, the main addition is the introduction of an aggregation mechanism across multiple relations and the use of heterogeneous GNNs for feature extrapolation). On top of this, while yes the method appears to show good results on the considered datasets compared to the baselines, I’ve some doubts about the experimental ev
1. The authors provide a large number of experiments to analyze the effectiveness of the model. 2. THGs are worth exploring.
1. The shortcomings of other THNs aren't clarified clearly. For instance, what does *easily incorporate state-of-the-art designs from static GNNs* mean? And what are the specific drawbacks of these methods? The current presentation lacks clarity, diminishing the paper's motivation when compared to other THNs. Besides, related work should be cited in the introduction section. 2. This paper's contribution is limited for ICLR standard. The authors primarily employ the ROLAND framework and convent
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
