Self-supervised learning advances audio representation for multimedia analysis. However, prevailing data-centric approaches rely on massive real-world corpora, increasing training costs, curation burdens, and privacy barriers. To address this, we present AudioPG, a procedural synthesis framework eliminating real audio recordings during pre-training. AudioPG trains a Transformer-based masked autoencoder on waveforms generated on-the-fly from basic acoustic primitives and composition rules. The encoder transfers effectively to real audio benchmarks, achieving 90.60% accuracy on ESC-50, 0.546 mAP on FSD50K, 88.17% on UrbanSound8K, and 97.03% on Speech Commands V2. Notably, pre-training completes in under 20 minutes on a single GPU. Latent space analysis reveals physical factors, including fundamental frequency and relative intensity, emerge in orthogonal subspaces, making representations linearly decodable. These results establish procedural synthesis as an efficient, interpretable pre-training signal when large-scale corpora are unavailable. Our code is available at: https://github.com/Freyliu0516/audioPG.

Inspired by the success of contrastive learning in graphs, researchers have begun exploring the potential of hypergraph contrastive learning. However, existing methods exhibit the following limitations when modeling higher-order relationships in unlabeled data: (1) a failure to fully consider the reliability of different nodes and their critical role in contrastive learning; (2) most methods overlook the importance of temperature coefficients in the contrastive learning process. To address these issues, we propose a hypergraph contrastive learning method based on high-reliability node awareness and adaptive temperature coefficients (ADT-HCL). First, we design a high-reliability node awareness technique based on perturbation invariance. This technique can quantify node reliability and filter out nodes with high noise robustness from both ordinary graph and hypergraph perspectives. Additionally, we propose a temperature coefficient adaptation strategy based on sample pair similarity. This strategy dynamically adjusts the temperature coefficients in contrastive learning, thereby enhancing the model’s ability to distinguish between sample pairs. We conducted extensive experiments on multiple mainstream datasets, and the results demonstrate that the proposed ADT-HCL method achieves outstanding performance in hypergraph node classification tasks.

Graph convolutional networks (GCN) have recently emerged as powerful node embedding methods in network analysis tasks. Particularly, GCNs have been successfully leveraged to tackle the challenging link prediction problem, aiming at predicting missing links that exist yet were not found. However, most of these models are oriented to undirected graphs, which are limited to certain real-life applications. Therefore, based on the social ranking theory, we extend the GCN to address the directed link prediction problem. Firstly, motivated by the reciprocated and unreciprocated nature of social ties, we separate nodes in the neighbor subgraph of the missing link into the same, a higher-ranked and a lower-ranked set. Then, based on the three kinds of node sets, we propose a method to correctly aggregate and propagate the directional information across layers of a GCN model. Empirical study on 8 real-world datasets shows that our proposed method is capable of reserving rich information related to directed link direction and consistently performs well on graphs from numerous domains.

The classification of sentences is very challenging, since sentences contain the limited contextual information. In this paper, we proposed an Attention-Gated Convolutional Neural Network (AGCNN) for sentence classification, which generates attention weights from the feature’s context windows of di fferent sizes by using specialized convolution encoders. It makes full use of limited contextual information to extract and enhance the influence of important features in predicting the sentence’s category. Experimental results demonstrated that our model can achieve up to 3.1% higher accuracy than standard CNN models, and gain competitive results over the baselines on four out of the six tasks. Besides, we designed an activation function, namely, Natural Logarithm rescaled Rectified Linear Unit (NLReLU). Experiments showed that NLReLU can outperform ReLU and is comparable to other well-known activation functions on AGCNN.

Modularity Evaluation (ME) is usually used in community detection for evaluating the disjoint and overlapping communities. In this paper, two obvious defects of ME are revealed and proved, including the non-decreasing contribution of isolated nodes to modularity and lacking of appropriate measures on overlapping community. We also propose a new evaluation criterion, the USI-AUC, which is the Area Under the Curve (AUC), originated from link-prediction of Uniform-Structure-Information (USI) model. We test the new criterion on various datasets, and find that such criterion can avoid the issues exposed in ME.