Users participate in multiple social networks for different services. User identity linkage aims to predict whether users across different social networks refer to the same person, and it has received significant attention for downstream tasks such as recommendation and user profiling. Recently, researchers proposed measuring the relevance of user-generated content to predict identity linkages of users. However, there are two challenging problems with existing content-based methods: first, barely considering the word similarities of texts is insufficient where the semantical correlations of named entities in the texts are ignored; second, most methods use time discretization technology, where the texts are divided into different time slices, resulting in failure of relevance modeling. To address these issues, we propose a user identity linkage model with the enhancement of a knowledge graph and continuous time decay functions that are designed for mitigating the influence of time discretization. Apart from modeling the correlations of the words, we extract the named entities in the texts and link them into the knowledge graph to capture the correlations of named entities. The semantics of texts are enhanced through the external knowledge of the named entities in the knowledge graph, and the similarity discrimination of the texts is also improved. Furthermore, we propose continuous time decay functions to capture the closeness of the posting time of texts instead of time discretization to avoid the matching error of texts. We conduct experiments on two real public datasets, and the experimental results show that the proposed method outperforms state-of-the-art methods.

Single-frame Infrared Small Target Detection (SIRSTD) identifies and isolates small objects in a single infrared image, playing a crucial role in fields such as search and rescue, particularly under extreme climatic conditions. The main challenge lies in segmenting the small target from the background. When the sampling frequency doesn’t meet the Nyquist criterion, aliasing occurs, which significantly reduces the accuracy of detection. Although deep learning has significantly improved the accuracy and efficiency of SIRSTD, it also brings issues such as high memory usage, redundant parameters, and decreased efficiency. To address these challenges, we propose a novel efficient anti-aliasing network (EAA-Net) for single-frame infrared small target detection. Our approach involves an anti-aliasing contextual aggregation (AACA) module and a partial differential channel pruning (PDCP) module. The AACA module suppresses aliasing effects by filtering out channels severely affected by aliasing. In addition, to tackle the problems of redundant parameters and low model efficiency, we design a PDCP module which represents the weight matrix with the partial differential equation (PDE) and formulate a partial differential channel pruning method. These structures not only significantly enhance the capacity of anti-aliasing of the network but also reduce the size. Extensive experiments on the public IRSTD-1k and SIRST datasets demonstrate that the proposed EAA-Net outperforms the state-of-the-art methods.