Social networks exhibit a Gravitational Field phenomenon, where key nodes, called Gravitational Sources, generate influential zones in their neighborhoods. Traditional methods mainly rely on centrality metrics or heuristic approaches that estimate influence. But they fail to capture the dual characteristics of topology and dynamics required for accurately identifying Gravitational Sources. Identifying Gravitational Sources faces three main challenges: 1) how to construct an effective evaluation mechanism to calculate gravitational source scores; 2) how to achieve effective model training in large-scale network environments with label scarcity; and 3) how to quantify the uncertainty of prediction results to enhance identification reliability. To address these challenges, we propose uncertainty-aware Bayesian graph convolutional network with sequential transfer learning (UBGCN-STL). Our approach integrates an evaluation method that combines centrality and influence metrics, employs a sequential transfer learning strategy to leverage knowledge from label-rich small networks for large networks, and incorporates a Bayesian mechanism for uncertainty-aware predictions. Experiments on seven real-world networks (social, protein interaction, and animal networks) show that UBGCN-STL outperforms representative baseline methods based on centrality, network structure, or learning models (LCNN, NDM, IpGCN), achieving superior predictive performance and reliable uncertainty quantification even with limited labels.

Domain generalization (DG) tasks aim to learn cross-domain models from source domains and apply them to unknown target domains. Recent research has demonstrated that diverse and rich source domain samples can enhance domain generalization capability. This work argues that the impact of each sample on the model's generalization ability varies. Even a small-scale but high-quality dataset can achieve a notable level of generalization. Motivated by this, we propose a domain-adversarial active learning (DAAL) algorithm for classification tasks in DG. First, we analyze that the objective of DG tasks is to maximize the inter-class distance within the same domain and minimize the intra-class distance across different domains. We design a domain adversarial selection method that prioritizes challenging samples in an active learning (AL) framework. Second, we hypothesize that even in a converged model, some feature subsets lack discriminatory power within each domain. We develop a method to identify and optimize these feature subsets, thereby maximizing inter-class distance of features. Lastly, We experimentally compare our DAAL algorithm with various DG and AL algorithms across four datasets. The results demonstrate that the DAAL algorithm can achieve strong generalization ability with fewer data resources, thereby significantly reducing data annotation costs in DG tasks.

The widespread adoption of location-based applications has created a growing demand for point-of-interest (POI) recommendation, which aims to predict a user’s next POI based on their historical check-in data and current location. However, existing methods often struggle to capture the intricate relationships within check-in data. This is largely due to their limitations in representing temporal and spatial information and underutilizing rich semantic features. While large language models (LLMs) offer powerful semantic comprehension to solve them, they are limited by hallucination and the inability to incorporate global collaborative information. To address these issues, we propose a novel method SeCor, which treats POI recommendation as a multi-modal task and integrates semantic and collaborative representations to form an efficient hybrid encoding. SeCor first employs a basic collaborative filtering model to mine interaction features. These embeddings, as one modal information, are fed into LLM to align with semantic representation, leading to efficient hybrid embeddings. To mitigate the hallucination, SeCor recommends based on the hybrid embeddings rather than directly using the LLM’s output text. Extensive experiments on three public real-world datasets show that SeCor outperforms all baselines, achieving improved recommendation performance by effectively integrating collaborative and semantic information through LLMs.

Event extraction is an essential but challenging task in information extraction. This task has considerably benefited from pre-trained language models, such as BERT. However, when it comes to the trigger-word mismatch problem in languages without natural delimiters, existing methods ignore the complement of lexical information to BERT. In addition, the inherent multi-role noise problem could limit the performance of methods when one sentence contains multiple events. In this article, we propose a Mask-Attention-based BERT (MABERT) framework for Chinese event extraction to address the above problems. Firstly, in order to avoid trigger-word mismatch and integrate lexical features into BERT layers directly, a mask-attention-based transformer augmented with two mask matrices is devised to replace the original one in BERT. By the mask-attention-based transformer, the character sequence interacts with external lexical semantics sufficiently and keeps its structure information at the same time. Moreover, against the multi-role noise problem, we make use of event type information from representation and classification, two aspects to enrich entity features, where type markers and event-schema-based mask matrix are proposed. Experimental results on the widely used ACE2005 dataset show the effectiveness of our proposed MABERT on Chinese event extraction task compared with other state-of-the-art methods.