Graph-based models are a powerful technique for recommendation systems, which seek to leverage the graph structure created by user-item interactions for elevated performance. The majority of them are designed for unsigned graphs, which fail to exploit negative interactions (e.g., dislikes, returns) from users, and hence, incur compromised effectiveness. To tap into such negative signals, in recent years, a number of efforts have been invested towards extending graph neural networks (GNNs) and Transformer models to signed graphs. Unfortunately, the former approaches produce sub-par results due to the lack of access to global information, whereas the latter achieve superior performance for recommendation but suffer from severe over-globalizing problems and substantial computational overhead. To bridge this gap, this paper presents SPGNN, which significantly unleashes the capabilities of GNNs and advances its performance for top-K recommendation in signed graphs through two non-trivial technical contributions. Firstly, we propose to upgrade the neighborhood aggregation scheme in GNNs with two novel notions of signed local proximity (SLP) and signed global proximity (SGP) based on weak balance theory, which can accurately capture sign-aware multi-scale relations between nodes in signed graphs. On top of that, SPGNN includes a theoretically-grounded module for effective feature initialization, which carefully crafts sign-aware structure embeddings via fast spectral decomposition. Extensive experiments show that SPGNN significantly outperforms other unsigned and sign-aware models on six benchmark datasets with up to a gain of 19.42% in Recall and 28.18% in NDCG, which indicates the traditional GNN architecture also holds great potential for signed graph recommendation with appropriate modification. Our code is available at https://github.com/yfsong00/SPGNN.

Incomplete node features are ubiquitous in real-world scenarios such as user profiling and cold-start recommendation, which severely hinders the practical deployment of graph learning systems (e.g., GNNs). Existing solutions typically rely on diffusion-based structural smoothing (e.g., feature propagation) to impute missing values. However, we find that these approaches suffer from structural overfitting, leading to three progressive challenges: 1) performance degradation on disjoint graphs, 2) loss of semantic diversity due to over-smoothing, and 3) feature distribution shift when generalizing to unseen graph structures (inductive tasks). To address these challenges, we introduce the DART framework. It begins by employing Global Structural Augmentation (GSA), which establishes global correlations to bridge disjoint components and extend diffusion coverage. Building upon this, we design a semantic rectifier based on masked autoencoding. This module learns the latent feature manifold to recover natural semantic details. Crucially, we introduce a test-time distribution rectification mechanism that projects structurally biased features back onto the learned manifold during inference, effectively bridging the inductive distribution gap. Furthermore, considering that synthetic masking fails to reflect realworld sparsity, we present a new dataset Sailing collected from voyage records with naturally missing attributes. Extensive experiments on six public datasets and Sailing demonstrate that DART significantly outperforms state-of-the-art methods in both transductive and inductive settings. Our code and dataset are available at https://github.com/yfsong00/DART.

While diffusion distillation has enabled one-step generation through methods like Variational Score Distillation, adapting distilled models to emerging new controls -- such as novel structural constraints or latest user preferences -- remains challenging. Conventional approaches typically requires modifying the base diffusion model and redistilling it -- a process that is both computationally intensive and time-consuming. To address these challenges, we introduce Joint Distribution Matching (JDM), a novel approach that minimizes the reverse KL divergence between image-condition joint distributions. By deriving a tractable upper bound, JDM decouples fidelity learning from condition learning. This asymmetric distillation scheme enables our one-step student to handle controls unknown to the teacher model and facilitates improved classifier-free guidance (CFG) usage and seamless integration of human feedback learning (HFL). Experimental results demonstrate that JDM surpasses baseline methods such as multi-step ControlNet by mere one-step in most cases, while achieving state-of-the-art performance in one-step text-to-image synthesis through improved usage of CFG or HFL integration.

Despite extensive research efforts focused on Out-of-Distribution (OOD) detection on images, OOD detection on nodes in graph learning remains underexplored. The dependence among graph nodes hinders the trivial adaptation of existing approaches on images that assume inputs to be i.i.d. sampled, since many unique features and challenges specific to graphs are not considered, such as the heterophily issue. Recently, GNNSafe, which considers node dependence, adapted energy-based detection to the graph domain with state-of-the-art performance, however, it has two serious issues: 1) it derives node energy from classification logits without specifically tailored training for modeling data distribution, making it less effective at recognizing OOD data; 2) it highly relies on energy propagation, which is based on homophily assumption and will cause significant performance degradation on heterophilic graphs, where the node tends to have dissimilar distribution with its neighbors. To address the above issues, we suggest training Energy-based Models (EBMs) by Maximum Likelihood Estimation (MLE) to enhance data distribution modeling and removing energy propagation to overcome the heterophily issues. However, training EBMs via MLE requires performing Markov Chain Monte Carlo (MCMC) sampling on both node feature and node neighbors, which is challenging due to the node interdependence and discrete graph topology. To tackle the sampling challenge, we introduce Decoupled Graph Energy-based Model (DeGEM), which decomposes the learning process into two parts—a graph encoder that leverages topology information for node representations and an energy head that operates in latent space. Additionally, we propose a Multi-Hop Graph encoder (MH) and Energy Readout (ERo) to enhance node representation learning, Conditional Energy (CE) for improved EBM training, and Recurrent Update for the graph encoder and energy head to promote each other. This approach avoids sampling adjacency matrices and removes the need for energy propagation to extract graph topology information. Extensive experiments validate that DeGEM, without OOD exposure during training, surpasses previous state-of-the-art methods, achieving an average AUROC improvement of 6.71% on *homophilic* graphs and 20.29% on *heterophilic* graphs, and even outperform methods trained with OOD exposure. Our code is available at: [https://github.com/draym28/DeGEM](https://github.com/draym28/DeGEM).

Link recommendation systems in online social networks (OSNs), such as Facebook's "People You May Know", Twitter's "Who to Follow", and Instagram's "Suggested Accounts", facilitate the formation of new connections among users. This paper addresses the challenge of link recommendation for the purpose of social influence maximization. In particular, given a graph G and the seed set S, our objective is to select k edges that connect seed nodes and ordinary nodes to optimize the influence dissemination of the seed set. This problem, referred to as influence maximization with augmentation (IMA), has been proven to be NP-hard.

Graph embedding aims at mapping each node to a low-dimensional vector, beneficial for various applications like pattern matching, retrieval augmented generation and recommendation. In this paper, we study the large-scale temporal graph embedding problem. Different from simple graphs, each edge has a timestamp in temporal graphs, which requires the embeddings to encode the temporal biases. Factorizing similarity matrix is a common approach for generating simple graph embeddings where similarity can be well characterized by some conventional metrics like personalized PageRank. However, how to construct a similarity that can encode interactions with temporal biases is a critical problem for large scale temporal graphs. To address this, we introduce the concept of temporal-based bipartite graph (TBG) and develop the temporal preferential attachment similarity (TPASim) that reflects concurrent node activity over time. Directly factorizing the TPASim matrix, which contains nearly n 2 non-zeros, is not feasible for large graphs with n nodes. Instead, we present LTGE, which constructs and factorizes a temporal matrix with at most 2 m non-zeros, where m is the number of edges. Our theoretical analysis shows that LTGE achieves the same embeddings as factorizing the TPASim matrix but significantly reduces complexity by a factor of n 2 / m. On the other hand, when graphs evolve over time, to avoid recomputing, we further propose LTGEInc that utilizes a novel incremental singular value decomposition (SVD) algorithm with provable guarantee for updating the embeddings. Extensive experiments on several datasets with up to 17 million nodes and 1.3 billion edges demonstrate that LTGE outperforms the state of the art significantly and is orders of magnitude faster than the baselines specially designed for temporal graphs. For embeddings update, LTGEInc retains the performance with small computational overhead.