Tensor Convolutional Network for Higher-Order Interaction Prediction in Sparse Tensors
Jun-Gi Jang, Jingrui He, Andrew Margenot, Hanghang Tong · 2025
Read Tensor Convolutional Network for Higher-Order Interaction Prediction in Sparse Tensors by Jun-Gi Jang, Jingrui He, Andrew Margenot, Hanghang Tong (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-2503-11786.txt.
Introduction
Lots of real world data form sparse grids, where most cells sit empty and the task is to guess which links will form. Jun-Gi Jang, Jingrui He, Andrew Margenot, and Hanghang Tong take on that guesswork in a 2025 study. They build a network tool that learns richer marks for each item in the grid.
Their tool first turns the sparse grid into a graph, then learns marks for items from their links in that graph. Those marks plug into older grid methods and lift their scores for picking the most likely new links. Tests show the paired tool beats both older grid methods and a graph based rival. The design works with many hosts, which widens its use.
Catalog record
| CREATOR | Jun-Gi Jang, Jingrui He, Andrew Margenot, Hanghang Tong |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 63 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-11786
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Tensor Convolutional Network for Higher-Order Interaction Prediction in Sparse Tensors Jun-Gi Jang, Jingrui He, Andrew Margenot, Hanghang Tong Abstract . Many real-world data, such as recommendation data and temporal graphs, can be represented as incomplete sparse tensors where most entries are unobserved. For such sparse tensors, identifying the top- k higher-order interactions that are most likely to occur among unobserved ones is crucial. Tensor factorization (TF) has gained significant attention in various tensor-based applications, serving as an effective method for finding these top- k potential interactions. However, existing TF methods primarily focus on effectively fusing latent vectors of entities, which limits their expressiveness. Since most entities in sparse tensors have only a few interactions, their latent representations are often insufficiently trained. In this paper, we propose TCN , an accurate and compatible tensor convolutional network that integrates seamlessly with existing TF methods for predicting higher-order interactions. We design a highly effective encoder to generate expressive latent vectors of entities. To achieve this, we propose to (1) construct a graph structure derived from a sparse tensor and (2) develop a relation-aware encoder, TCN , that learns latent representations of entities by leveraging the graph structure. Since TCN complements traditional TF methods, we seamlessly integrate TCN with existing TF methods, enhancing the performance of predicting top- k k interactions. Extensive experiments show that TCN integrated with a TF method outperforms competitors, including TF methods and a hyperedge prediction method. Moreover, TCN is broadly compatible with various TF methods and GNNs (Graph Neural Networks), making it a versatile solution. DOI: XXXXXXX.XXXXXXX Conference: The 48th International ACM SIGIR Conference on Research and Development in Information Retrieval; July, 2025; Padova, Italy ISBN: 978-1-4503-XXXX-X/18/06 1. Introduction Tensor, a multi-dimensional array, is a natural representation of real-world datasets that contain higher-order relationships. An N -th order tensor of the size I_{1}\times\cdots\times I_{N}=T has T entries each of which has a value indexed by indices (i_{1},...i_{n},...,i_{N}) where i_{n} denotes i_{n} -th entity in n -th dimension. Specifically, many real-world data such as recommendation data, knowledge graphs, social networks, and web logs can be represented as sparse tensors whose most entries are unobserved. In other words, there are only a few observed interactions between entities. For example, given higher-order recommendation data where entities of each dimension correspond to users, location, and timestamps, respectively, an (i_{1},i_{2},i_{3}) -th entry value indicates an implicit (or explicit) feedback that the i_{1} -th user is interested in the i_{2} -th location at the i_{3} -th timestamp. Most entries are unobserved since most users provide feedback for only a few items at specific timestamps: as summarized in Table 1 , 99.9984% entries are unobserved on SG dataset ( Li et al., 2015 ) which is POI (Point-of-Interest) recommendation data. Given an N -th order sparse tensor, it is crucial to infer potential interactions between N entities i_{1},...,i_{N} on different dimensions. We need to identify top- k interactions carefully among unobserved interactions since it is not the desired situation that all interactions in a sparse tensor are observed. For example, given a 3rd-order tensor of recommendation data (user, item, time), it is not realistic for every user to interact with every item at every time slot. Each user has distinct preferences only for a few specific items, and there are dedicated browsing time slots. Therefore, observed interactions in a sparse tensor represent meaningful higher-order relationships between entities (e.g., users, items, and times). Identifying top- k interactions has various real-world applications such as recommendation systems, temporal graph analysis, and knowledge graph completion. In addition, one such application is an enhanced tensor completion where only significant entries among all unobserved entries are filled carefully; (1) we identify top-k interactions using an identification method and then (2) fill in the values for the top-k interactions using a completion method. Then, how can we find top- k potential interactions between entities accurately? For predicting top- k potential interactions, we can use tensor factorization methods such as traditional methods (e.g., CP ( Carroll and Chang, 1970 ; Harshman et al., 1970 ; Kiers, 2000 ) and Tucker ( Tucker, 1966 ) decomposition) and deep learning-based methods (e.g., CostCo ( Liu et al., 2019 ) and NeAT ( Ahn et al., 2024 ) ). Tensor factorization has garnered significant attention across various applications such as tensor completion ( Tomasi and Bro, 2005 ; Acar et al., 2011 ; Smith et al., 2016 ; Liu et al., 2014 ) , anomaly detection ( Jang and Kang, 2021 ; Jang et al., 2023 ) , recommendation ( Choi et al., 2019 ; Chen and Li, 2020 ) , knowledge graph completion ( Balazevic et al., 2019 ; Lacroix et al., 2020 ) , and so on ( Perros et al., 2017 ; Oh et al., 2018 ; Chen and Li, 2018 ) . A commonly considered approach is to (1) design a score function using a tensor factorization method, (2) train its parameters to make observed interactions high scores, (3) compute scores of unobserved interactions using the score function, and (4) pick top- k highest-scored ones. However, traditional TF methods face a significant challenge for incomplete sparse tensors where the latent vectors of most entities with a limited number of interactions remain undertrained due to high sparsity. This is because traditional TF methods typically focus on how to fuse multiple latent vectors effectively. In this paper, we propose TCN , an accurate T ensor C onvolutional N etwork (TCN) that is fully compatible with a variety of tensor factorization methods for higher-order interaction prediction. To address the challenge of traditional TF methods, our proposition is to (1) extract a graph structure from a given sparse tensor and (2) generate enriched latent vectors of entities by aggregating information over the graph. Specifically, we construct a hypergraph structure from a given N -th order sparse tensor where its entities and interactions correspond to nodes and hyperedges of the size N , respectively. Then we obtain a clique expansion graph from the hypergraph where each interaction (i.e., hyperedge) of an N -th order tensor can be viewed as a N -clique. Next, we design TCN that extracts relation-aware latent vectors of entities by neighborhood aggregation through the graph. TCN effectively addresses the issue of poorly learning the latent vectors for entities with limited interactions. Since the output of TCN can be readily utilized as the input of a tensor factorization method, we construct an end-to-end model that consists of TCN and a tensor factorization method. TCN is highly compatible with a variety of TF methods since TCN and TF methods serve different but complementary roles. The main contributions of this paper are summarized as follows: • A New Insight. We provide a novel insight in a tensor-based problem: (1) a construction of hypergraph and its clique-expanded graph derived from a sparse tensor, and (2) a development for a relation-aware encoder that learns latent vectors of entities using the graph. • Accurate Method. We propose TCN , which is highly accurate in achieving the state-of-the-art method for top- k higher-order interaction prediction. • Plug-and-Play. The plug-in capability for various tensor factorization (See Table 2 ) is the strength of TCN . In addition, our tensor-originated clique expansion graph