1\ell_1 Latent Distance based Continuous-time Graph Representation

ICLR 2026 Conference SubmissionAnonymous Authors
$\ell_1$ distancegraph representationsequential survival processultra-low-dimensional embedding
Abstract:

Continuous-time graph representation (CTGR) is a widely-used methodology in machine learning, physics, bioinformatics, and social networks. The sequential survival process in a latent space with the squared 2\ell_2 distance is an important ultra-low-dimensional embedding for CTGR. However, the squared 2\ell_2 distance violates the triangle inequality, which may cause distortion of the relative node positions in the latent space and thus deteriorates in social, contact, and collaboration networks. Reverting to the 2\ell_2 distance is infeasible because the corresponding integral computation is intractable. To solve these problems, we propose a theoretically-sound 1\ell_1 latent distance based continuous-time graph representation (1\ell_1LD-CTGR). It facilitates a true latent metric space for the sequential survival process. Moreover, the integral of the hazard function is found to be a closed-form piece-wise exponential integral, which well fits the ultra-low-dimensional embedding. To handle the non-differentiable 1\ell_1 norm, we successfully find a descent direction of the hazard function to replace the gradient, enabling mainstream learning architectures to learn the parameters. Extensive experiments using both synthetic and real-world data show the competitive performance of 1\ell_1LD-CTGR.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper proposes an ℓ₁ latent distance model for continuous-time graph representation, addressing triangle inequality violations in prior squared-ℓ₂ approaches. It resides in the 'Distance-Based Survival Process Models' leaf, which contains only two papers total (including this one). This leaf sits within 'Latent Space Models for Continuous-Time Networks', a branch with three sub-topics and six papers overall. The sparse population suggests this is a relatively focused research direction rather than a crowded area, though the broader latent-space modeling branch has moderate activity across multiple geometric and process-based approaches.

The taxonomy reveals neighboring work in 'Hawkes Process Latent Space Models' (one paper) and 'Trajectory-Based Latent Dynamics' (three papers), both exploring continuous-time embeddings but through different mechanisms—mutually exciting processes versus velocity-driven trajectories. The sibling paper in the same leaf (Sequential Survival Process) shares the survival-process foundation but does not explicitly address metric properties or ℓ₁ distances. Adjacent branches like 'Hyperbolic and Non-Euclidean Temporal Embeddings' explore alternative geometries for discrete-time snapshots, highlighting a broader field interest in moving beyond standard Euclidean metrics, though in different temporal modeling regimes.

Among 22 candidates examined, the ℓ₁ latent distance framework (Contribution A) showed no clear refutations across two candidates, suggesting limited direct prior work on ℓ₁-based survival models. The closed-form integral derivation (Contribution B) examined ten candidates with no refutations, indicating novelty in the mathematical treatment. However, the descent direction optimization method (Contribution C) found one refutable candidate among ten examined, pointing to existing subgradient or proximal techniques for non-differentiable norms. The limited search scope (22 papers, not exhaustive) means these findings reflect top semantic matches rather than comprehensive coverage.

Based on the top-22 semantic matches and taxonomy structure, the work appears to occupy a relatively sparse niche within continuous-time latent-space modeling. The ℓ₁ metric choice and closed-form integral derivation show novelty signals, though the optimization approach has partial overlap with known non-smooth methods. The analysis does not cover broader optimization literature or domain-specific survival modeling outside the examined candidates, so conclusions remain provisional pending deeper review.

Taxonomy

Core-task Taxonomy Papers
28
3
Claimed Contributions
22
Contribution Candidate Papers Compared
1
Refutable Paper

Research Landscape Overview

Core task: continuous-time graph representation learning with latent distance metrics. The field addresses how to embed evolving networks into latent spaces where distances capture the likelihood or timing of interactions. The taxonomy reveals several complementary perspectives: Latent Space Models for Continuous-Time Networks focus on probabilistic frameworks that treat node positions as evolving continuously and model edge formation via distance-based survival or point processes; Discrete-Time Temporal Network Embeddings discretize time into snapshots and learn embeddings per interval; Temporal Knowledge Graph Embeddings extend relational reasoning to time-stamped facts; Link Prediction and Optimization branches emphasize algorithmic efficiency and predictive accuracy; Temporal Graph Distance and Similarity Metrics develop measures to compare dynamic graphs; Domain-Specific Applications tailor methods to areas like social networks or epidemiology; and Theoretical Foundations provide cross-domain principles. Representative works such as Latent Functions Networks[3] and Tensor Factorization Networks[2] illustrate how different branches balance expressiveness, scalability, and interpretability. Within the latent-space modeling branch, a particularly active line explores distance-based survival process models, which treat the hazard of edge formation as a function of latent distances that evolve smoothly over time. L1 Latent Distance[0] sits squarely in this cluster, proposing an L1-norm metric to capture anisotropic node dynamics and improve interpretability compared to Euclidean alternatives. Its closest neighbor, Sequential Survival Process[1], similarly models continuous-time interactions via survival analysis but emphasizes sequential event dependencies. Both contrast with earlier static latent-space approaches like Social Network Latent[16] and more recent hybrid methods such as Hyperbolic Temporal Embedding[4], which leverage non-Euclidean geometries. A central open question across these works is how to balance model complexity—richer distance functions or time-varying embeddings—against computational tractability and the risk of overfitting sparse temporal data.

Claimed Contributions

l1 latent distance based continuous-time graph representation (l1LD-CTGR)

The authors introduce l1LD-CTGR, a new method that uses l1 distance instead of squared l2 distance in the sequential survival process for continuous-time graph representation. This approach establishes a valid metric space that satisfies the triangle inequality, addressing distortion issues in the latent space that affect social, contact, and collaboration networks.

2 retrieved papers
Closed-form piece-wise exponential integral for l1 distance hazard function

The authors derive a tractable closed-form solution for computing the integral of the hazard function when using l1 distance. This piece-wise exponential integral differs from the Gaussian integral used in squared l2 distance methods and enables efficient computation in ultra-low-dimensional embeddings.

10 retrieved papers
Descent direction method for non-differentiable l1 norm optimization

The authors develop a method to handle the non-differentiability of the l1 norm by identifying a descent direction that replaces the gradient in optimization. This enables the use of standard learning frameworks like PyTorch for parameter learning despite the non-smooth objective function.

10 retrieved papers
Can Refute

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

l1 latent distance based continuous-time graph representation (l1LD-CTGR)

The authors introduce l1LD-CTGR, a new method that uses l1 distance instead of squared l2 distance in the sequential survival process for continuous-time graph representation. This approach establishes a valid metric space that satisfies the triangle inequality, addressing distortion issues in the latent space that affect social, contact, and collaboration networks.

Contribution

Closed-form piece-wise exponential integral for l1 distance hazard function

The authors derive a tractable closed-form solution for computing the integral of the hazard function when using l1 distance. This piece-wise exponential integral differs from the Gaussian integral used in squared l2 distance methods and enables efficient computation in ultra-low-dimensional embeddings.

Contribution

Descent direction method for non-differentiable l1 norm optimization

The authors develop a method to handle the non-differentiability of the l1 norm by identifying a descent direction that replaces the gradient in optimization. This enables the use of standard learning frameworks like PyTorch for parameter learning despite the non-smooth objective function.

$\ell_1$ Latent Distance based Continuous-time Graph Representation | Novelty Validation