TAMI: Taming Heterogeneity in Temporal Interactions for Temporal Graph Link Prediction

Zhongyi Yu, Jianqiu Wu, Zhenghao Wu, Shuhan Zhong, Weifeng Su, Chul-Ho Lee, Weipeng Zhuo · 2025

Read TAMI: Taming Heterogeneity in Temporal Interactions for Temporal Graph Link Prediction by Zhongyi Yu, Jianqiu Wu, Zhenghao Wu, Shuhan Zhong, Weifeng Su, Chul-Ho Lee, Weipeng Zhuo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-23577.txt.

Introduction

Zhongyi Yu and six co-authors tackle temporal graph link prediction: guessing which nodes in a changing network will interact next, given their history. Real interaction streams are lopsided — a few pairs account for most events, and gaps between events vary wildly. Standard models encode time poorly under that lopsidedness and tend to forget pairs that interact only now and then, exactly where predictions are hardest.

Their framework, TAMI, attacks both problems. A log time encoding squashes long and short gaps into a more balanced scale, and a link history aggregation step keeps each target pair's past from fading. Existing top temporal graph networks can plug into TAMI without redesign. Across thirteen classic datasets and three newer benchmark sets, TAMI steadily improved the models it wrapped, in both settings where test nodes were seen before and where they were new. The code is public.

Catalog record

CREATORZhongyi Yu, Jianqiu Wu, Zhenghao Wu, Shuhan Zhong, Weifeng Su, Chul-Ho Lee, Weipeng Zhuo
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE113 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-23577

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

TAMI: Taming Heterogeneity in Temporal Interactions for Temporal Graph Link Prediction Zhongyi Yu, Jianqiu Wu, Zhenghao Wu, Shuhan Zhong, Weifeng Su, Chul-Ho Lee, Weipeng Zhuo Abstract Temporal graph link prediction aims to predict future interactions between nodes in a graph based on their historical interactions, which are encoded in node embeddings. We observe that heterogeneity naturally appears in temporal interactions, e.g., a few node pairs can make most interaction events, and interaction events happen at varying intervals. This leads to the problems of ineffective temporal information encoding and forgetting of past interactions for a pair of nodes that interact intermittently for their link prediction. Existing methods, however, do not consider such heterogeneity in their learning process, and thus their learned temporal node embeddings are less effective, especially when predicting the links for infrequently interacting node pairs. To cope with the heterogeneity, we propose a novel framework called TAMI, which contains two effective components, namely log time encoding function (LTE) and link history aggregation (LHA). LTE better encodes the temporal information through transforming interaction intervals into more balanced ones, and LHA prevents the historical interactions for each target node pair from being forgotten. State-of-the-art temporal graph neural networks can be seamlessly and readily integrated into TAMI to improve their effectiveness. Experiment results on 13 classic datasets and three newest temporal graph benchmark (TGB) datasets show that TAMI consistently improves the link prediction performance of the underlying models in both transductive and inductive settings. Our code is available at https://github.com/Alleinx/TAMI_temporal_graph . 1 Introduction Temporal link prediction is a fundamental task that forecasts future interactions between two nodes on continuous-time temporal graphs (CTTGs). This is important in a variety of real-world scenarios with dynamic graph topologies changing over time, such as social networks social_2 ; social_1 , user-item interaction systems JODIE ; user_item2 ; new_user_item_1 ; new_user_item_2 , traffic networks traffic1 ; new_traffic_1 ; new_traffic_2 , and physical systems physics ; new_physics_1 . The temporal link prediction task involves two major steps. The first step is to compute the temporal embedding of each node by aggregating information from its own historical interactions with neighboring nodes. For each target node pair for link prediction, the second step is to take their embeddings into a link predictor (typically an MLP) to estimate the probability of having a link between them. Heterogeneity arises in temporal interactions. For example, the number of interactions for each node pair varies substantially. The time between two consecutive interaction events can also be quite different. To see this, in Figure 1 , we plot the distribution of interaction intervals between any pair of nodes on the UCI dataset, showing that it follows a power-law distribution. In other words, most interactions are frequent ones, while the infrequent ones (very large interaction intervals) still exist with a non-negligible probability. We also compute and report the Fisher’s skewness in Figure 1 , indicating that the distribution is positively skewed or right-skewed. Here the skewness \Gamma of a random variable X is defined as \Gamma=\mathbb{E}\left[(X-\mu)^{3}/\sigma^{3}\right] , where \mu and \sigma denote the mean and standard deviation of X , respectively. The heterogeneity in temporal interactions poses two primary challenges to existing methods for temporal link prediction tgat ; graphmixer ; DyGFormer . First, they use sinusoidal functions as their time encoding functions to encode the temporal difference between the target time \tau and each interaction time t_{i} . We observe that the frequency parameters of the time encoding functions are harder to learn when the distribution of the temporal difference is highly skewed, which is the common case as seen from the distribution of interaction intervals. Second, they learn the embedding of each node based only on its recent interaction events. While the embedding encodes timely information, it can be predominantly influenced by a few neighbors having frequent interactions. As a result, the temporal link prediction can be inaccurate for a target node pair having infrequent interactions over time. For example, it is natural to predict that a couple would eat hamburgers if it is one of their recent favorites (frequent interactions). However, it makes more sense to predict that turkey will be on the table (infrequent interactions) if Thanksgiving is coming. Figure 1 : The interaction intervals between any pair of nodes on the UCI dataset follow a power-law distribution and have a high (positive) skewness value, meaning that the intervals are highly right-skewed. To address the above challenges, we propose TAMI, a novel framework that tam es the heterogeneity in temporal i nteractions to improve the performance of temporal link prediction. TAMI mainly contains two novel components, namely log time encoding function (LTE) and link history aggregation (LHA). We propose to use a logarithmic transformation in LTE to rescale the temporal difference before it is taken into the time encoding function, so that the frequency parameters of the time encoding function can be easier (or quicker) to learn. In addition, we develop LHA to preserve the information of the most recent k interactions for each target node pair for link prediction, regardless of their interaction frequency. In other words, the historical interactions of each node pair are explicitly used for their link prediction, no matter when they happened. Our key contributions can be summarized as follows: • First study on the heterogeneity in temporal interactions. To the best of our knowledge, this is the first work to identify the presence of heterogeneity in temporal interactions of CTTGs and investigate its impact on the performance of temporal link prediction. • TAMI: a novel framework to cope with the heterogeneity in temporal interactions. We propose TAMI, which effectively handles the heterogeneity in temporal interactions. TAMI contains two novel modules, namely LTE and LHA. LTE rescales the temporal differences using a logarithmic transformation while LHA is specifically designed to preserve historical interactions for each target node pair for link prediction. Existing temporal graph neural networks can be seamlessly and readily integrated into TAMI. • Extensive evaluation on open datasets . We validate the effectiveness of TAMI on 13 classic temporal graph datasets covering different fields with comprehensive temporal scales as well as three newest ones from temporal graph benchmark (TGB). Results show that TAMI consistently and substantially improves the link prediction accuracy and training efficiency of the underlying graph neural networks, with up to 87.05% improvement in link prediction accuracy and 76.7% reduction in total training time. 2 Related Work Time Encoding Functions in Link Prediction. Most prior methods tgat ; graphmixer ; DyGFormer ; CAWN ; TNCN ; TCL ; TGN ; rebuttal_linear_tpe use a common approach for time encoding, which is based on either periodic time encoding functions introduced in original_tpe or their simplified ones. To predict the existence of a link between two nodes at the target time, the temporal difference between the target time and the time of each historical interaction needs to be calculated. The time differences are then mapped to time encoding vectors using a periodic function such as sinusoidal function. However, this time encoding pipeline overlooks the effect of the skewness in the time differences. We observe that it takes longer than needed to train a model, and it can also degrade the performance of link prediction, especially when

Read the complete text