Dense video captioning aims to detect and describe all events in untrimmed videos. This paper presents a dense video captioning network called Multi-Concept Cyclic Learning (MCCL), which aims to: (1) detect multiple concepts at the frame level and leverage these concepts to provide temporal event cues; and (2) establish cyclic co-learning between the generator and the localizer within the captioning network to promote semantic perception and event localization. Specifically, weakly supervised concept detection is performed for each frame, and the detected concept embeddings are integrated into the video features to provide event cues. Additionally, video-level concept contrastive learning is introduced to produce more discriminative concept embeddings. In the captioning network, a cyclic co-learning strategy is proposed, where the generator guides the localizer for event localization through semantic matching, while the localizer enhances the generator’s event semantic perception through location matching, making semantic perception and event localization mutually beneficial. MCCL achieves state-of-the-art performance on the ActivityNet Captions and YouCook2 datasets. Extensive experiments demonstrate its effectiveness and interpretability.

The ground-state conformation of a molecule is often decisive for its properties. However, experimental or computational methods, such as density functional theory (DFT), are time-consuming and labor-intensive for obtaining this conformation. Deep learning (DL) based molecular representation learning (MRL) has made significant advancements in molecular modeling and has achieved remarkable results in various tasks. Consequently, it has emerged as a promising approach for directly predicting the ground-state conformation of molecules. In this regard, we introduce GTMGC, a novel network based on Graph-Transformer (GT) that seamlessly predicts the spatial configuration of molecules in a 3D space from their 2D topological architecture in an end-to-end manner. Moreover, we propose a novel self-attention mechanism called Molecule Structural Residual Self-Attention (MSRSA) for molecular structure modeling. This mechanism not only guarantees high model performance and easy implementation but also lends itself well to other molecular modeling tasks. Our method has been evaluated on the Molecule3D benchmark dataset and the QM9 dataset. Experimental results demonstrate that our approach achieves remarkable performance and outperforms current state-of-the-art methods as well as the widely used open-source software RDkit.

Predicting the metabolic pathway classes of compounds in the human body is an important problem in drug research and development. For this purpose, we propose a Multi-Scale Graph Neural Network framework, named MSGNN. The framework includes a subgraph encoder, a feature encoder and a global feature processor, and a graph augmentation strategy is adopted. The subgraph encoder is responsible for extracting the local structural features of the compound, the feature encoder learns the characteristics of the atoms, and the global feature processor processes the information from the pre-training model and the two molecular fingerprints, while the graph augmentation strategy is to expand the train set through a scientific and reasonable method. The experiment result illustrates that the accuracy, precision, recall and F1 metrics of MSGNN reach 98.17%, 94.18%, 94.43% and 94.30%, respectively, which is superior to the similar models we have known. In addition, the ablation experiment demonstrates the indispensability of MSGNN modules.

Protein sequences accumulate in large quantities, and the traditional method of annotating protein function by experiment has been unable to bridge the gap between annotated proteins and unannotated proteins. Machine learning-based protein function prediction is an effective approach to solve this problem. Most of the existing methods only use the protein sequence but ignore the three-dimensional structure which is closely related to the protein function. And the hierarchy of protein functions is not adequately considered. To solve this problem, we propose a graph neural network (GNNGO3D) that combines the three-dimensional structure and functional hierarchy learning. GNNGO3D simultaneously uses three kinds of information: protein sequence, tertiary structure, and hierarchical relationship of protein function to predict protein function. The novelty of GNNGO3D lies in that it integrates the learning of functional level information into the method of predicting protein function by using tertiary structure information, fully learning the relationship between protein functions, and helping to better predict protein function. Experimental results show that our method is superior to existing methods for predicting protein function based on sequence and structure.

Recommendation retrievers commonly retrieve user potentially preferred items from numerous items, where the query and item representation are learned according to the dual encoders with the log-softmax loss. Under real scenarios, the number of items becomes considerably large, making it exceedingly difficult to calculate the partition function with the whole item corpus. Negative sampling, which samples a subset from the item corpus, is widely used to accelerate the model training. Among different samplers, the in-batch sampling is commonly adopted for online recommendation retrievers, which regards the other items within the mini-batch as the negative samples for the given query, owing to its time and memory efficiency. However, the sample selection bias occurs due to the skewed feedback, harming the retrieval quality. In this paper, we propose a negative sampling approach named Batch-Mix Negative Sampling (BMNS), which adopts batch mixing operation to generate additional negatives for model training. Concretely, BMNS first generates new negative items with the sampled mix coefficient from the Beta distribution, after which a tailored correct strategy guided by frequency is designed to match the sampled softmax loss. In this way, the effort of re-encoding items out of the mini-batch is reduced while also improving the representation space of the negative set. The empirical experiments on four real-world datasets demonstrate BMNS is superior to the competitive negative inbatch sampling method.

Retrosynthesis and forward synthesis prediction are fundamental challenges in organic synthesis, computer-aided synthesis planning (CASP), and computer-aided drug design (CADD). The objective is to predict plausible reactants for a given target product and its corresponding inverse task. With the rapid development of deep learning, numerous approaches have been proposed to solve this problem from various perspectives. The methods based on molecular graphs benefit from their rich features embedded inside but face difficulties in applying existing sequence-based data augmentations due to the permutation invariance of graph structures. In this work, we propose SeqAGraph, a template-free approach that annotates input graphs with its root atom index to ensure compatibility with sequence-based data augmentation. The matrix product for global attention in graph encoders is implemented by indexing, elementwise product, and aggregation to fuse global attention with local message passing without graph padding. Experiments demonstrate that SeqAGraph fully benefits from molecular graphs and sequence-based data augmentation and achieves state-of-the-art accuracy in template-free approaches.

Traffic prediction is the cornerstone of the intelligent transportation system (ITS), and accurate prediction is essential for planning route, alleviating traffic pressure, and optimizing public transportation resource allocation. Although many methods have been proposed, they still have deficiencies in capturing the spatial-temporal dependence of traffic data. In specific, their network structures are usually discrete, making it is challenging to continuously model the dynamic spatial-temporal patterns of the road network. Besides, static graph structure of the road network is not sufficient to express dynamic traffic patterns. In this paper, we propose a diffusion graph neural ordinary differential equation network (DGODE) to address the above challenges for traffic prediction. Firstly, DGODE represents the node relationships of the road network as a bidirectional spatial graph and the node relationships of the time as a unidirectional temporal graph, and then generates an adaptive diffusion matrix to explore potential node relationships and capture spatial and temporal dependencies. Next, the neural ordinary differential equation (NODE) is introduced to contin-uously model the dynamical change of traffic network, making it possible to capture global dependence at a deeper network structure. Since the solution of the ordinary differential equation is determined by the initial state, we design the structure with a jump link to supplement the spatial-temporal information in the historical data. Finally, the experimental evaluation on four real-world datasets shows that DGODE is significantly superior to several baseline methods.

Identifying protein-protein interaction (PPI) site is an important step in understanding biological activity, apprehending pathological mechanism and designing novel drugs. Developing reliable computational methods for predicting PPI site as screening tools contributes to reduce lots of time and expensive costs for conventional experiments, but how to improve the accuracy is still challenging. We propose a PPI site predictor, called Augmented Graph Attention Network Protein-Protein Interacting Site (AGAT-PPIS), based on AGAT with initial residual and identity mapping, in which eight AGAT layers are connected to mine node embedding representation deeply. AGAT is our augmented version of graph attention network, with added edge features. Besides, extra node features and edge features are introduced to provide more structural information and increase the translation and rotation invariance of the model. On the benchmark test set, AGAT-PPIS significantly surpasses the state-of-the-art method by 8% in Accuracy, 17.1% in Precision, 11.8% in F1-score, 15.1% in Matthews Correlation Coefficient (MCC), 8.1% in Area Under the Receiver Operating Characteristic curve (AUROC), 14.5% in Area Under the Precision-Recall curve (AUPRC), respectively.

ABSTRACT We develop an efficient hard detector for multiple‐input multiple‐output (MIMO) channels, which adaptively combines maximum‐likelihood detection (MLD) and minimum‐mean‐square error with a successive interference canceler together. Unlike the conventional joint combination scheme, which may suffer from considerable degradation in bit‐error‐rate (BER) performance over correlated channels and where only one data stream is detected by MLD, our proposed scheme adaptively controls the number of data streams to be detected by MLD based on an analytical characterization of reliability for the detection. Simulation results illustrate that near‐optimal BER performance can be obtained at much lower computational complexity by the proposed method as compared with existing techniques, regardless of the spatial correlation of the MIMO channels. Copyright © 2012 John Wiley & Sons, Ltd.

We develop an efficient hard-detector for multiple-input multiple-output (MIMO) channels, which adaptively combines maximum-likelihood detection (MLD) and minimum-mean-square-error (MMSE) with a successive interference canceler (SIC) together. Unlike the conventional joint combination scheme which may suffer from considerable degradation in bit-error-rate (BER) performance over correlated channels and where only one data stream is detected by MLD, our proposed scheme adaptively controls the number of data streams to be detected by MLD based on the analytical characterization of reliability for the detection. Simulation results illustrate that near-optimal BER performance can be obtained at much lower computational complexity by the proposed method as compared to existing techniques, regardless of the spatial correlation of the MIMO channels.

This paper develops a joint maximum a posteriori (MAP) spatial filtering receiver by employing a joint parameter estimation approach for multiple-input multiple-output (MIMO) orthogonal frequency-division multiplexing (OFDM) modulation systems, which can achieve promising bit-error-rate (BER) performance in the presence of co-channel interference (CCI). For a MAP receiver to suppress the CCI, conventionally, this can be done in two ways: 1) spatio-temporal filtering (STF) the received signals prior to the fast Fourier transform (FFT), which we refer to it as preFFT-STF; 2) spatial filtering (SF) the received signals posterior to FFT, which is referred to as postFFT-SF. Simulation results show that the proposed scheme outperforms significantly both the preFFT-STF and postFFT-SF systems.

Alpha stable distribution has gained much attention due to its generality to represent heavy-tailed and impulsive interference. In such non-Gaussian interference, the detection key is to evaluate the zero-memory nonlinearity (ZMNL) function of locally optimal (LO) detector. Unfortunately, there is no closed form expression for the probability density function (PDF) of alpha-stable distributions. Hereby, sub-optimum ZMNL function is adopted as an unavoidable approximation, such as classical Cauchy and Gaussian-tailed ZMNL (GZMNL). In this paper, an algebraic-tailed ZMNL (AZMNL) with a concise form is proposed. Based on such ZMNL, derived detector has near optimal performance in various impulsive noise environments. Furthermore, using Bi-parameter CGM (BCGM), a concise approximate expression for PDF of symmetric α-stable (SαS) distribution, the test threshold can be evaluated according with preset false alarm ratio easily.