Multivariate time series forecasting has attracted wide attention in areas, such as system, traffic, and finance. The difficulty of the task lies that traditional methods fail to capture complicated non-linear dependencies between steps multiple series. Recently, recurrent neural network mechanism have been used model periodic temporal patterns across steps. However, these models fit not well for with dynamic-period or nonperiodic patterns. In this paper, we propose a dual self-attention...

While few-shot learning (FSL) aims for rapid generalization to new concepts with little supervision, self-supervised (SSL) constructs supervisory signals directly computed from unlabeled data. Exploiting the complementarity of these two manners, auxiliary has recently drawn much attention deal few labeled Previous works benefit sharing inductive bias between main task and tasks (SSL), where shared parameters are optimized by minimizing a linear combination losses. However, it is challenging...

Recently, despite the unprecedented success of large pre-trained visual-language models (VLMs) on a wide range downstream tasks, real-world unsupervised domain adaptation (UDA) problem is still not well explored. Therefore, in this paper, we first experimentally demonstrate that unsupervised-trained VLMs can significantly reduce distribution discrepancy between source and target domains, thereby improving performance UDA. However, major challenge for directly deploying such UDA tasks prompt...

This paper studies the problem of learning interactive recommender systems from logged feedbacks without any exploration in online environments. We address by proposing a general offline reinforcement framework for recommendation, which enables maximizing cumulative user rewards exploration. Specifically, we first introduce probabilistic generative model and then propose an effective inference algorithm discrete stochastic policy based on feedbacks. In order to perform more effectively, five...

This paper studies the problem of learning message propagation strategies for graph neural networks (GNNs). One challenges is that defining strategy. For instance, choices steps are often specialized to a single and not personalized different nodes. To compensate this, in this paper, we present propagate, general framework only learns GNN parameters prediction but more importantly, can explicitly learn interpretable propagate nodes various types graphs. We introduce optimal as latent...

Tensor decomposition is one of the most effective techniques for multi-criteria recommendations. However, it suffers from data sparsity when dealing with three-dimensional (3D) user-item-criterion ratings. To mitigate this issue, we consider effectively incorporating side information and cross-domain knowledge in tensor decomposition. A deep transfer (DTTD) method proposed by integrating structure Tucker decomposition, where an orthogonal constrained stacked denoising autoencoder (OC-SDAE)...

The purpose of few-shot recognition is to recognize novel categories with a limited number labeled examples in each class. To encourage learning from supplementary view, recent approaches have introduced auxiliary semantic modalities into effective metric-learning frameworks that aim learn feature similarity between training samples (support set) and test (query set). However, these only augment the representations available semantics while ignoring query set, which loses potential for...

In recent years, the application of multimodal large language models (MLLM) in various fields has achieved remarkable success. However, as foundation model for many downstream tasks, current MLLMs are composed well-known Transformer network, which a less efficient quadratic computation complexity. To improve efficiency such basic models, we propose Cobra, linear computational complexity MLLM. Specifically, Cobra integrates Mamba into visual modality. Moreover, explore and study modal fusion...

Due to the similar characteristics between event-based visual data and point clouds, recent studies have emerged that treat event as clouds learn based on cloud analysis. Additionally, some works approach from perspective of vision, employing Spiking Neural Network (SNN) due their asynchronous nature. However, these contributions are often domain-specific, making it difficult extend applicability other intersecting fields. Moreover, while SNN-based tasks seen significant growth, conventional...

Graph Neural Networks (GNNs) are state-of-the-art models for performing prediction tasks on graphs. While existing GNNs have shown great performance various related to graphs, little attention has been paid the scenario where out-of-distribution (OOD) nodes exist in graph during training and inference. Borrowing concept from CV NLP, we define OOD as with labels unseen set. Since a lot of networks automatically constructed by programs, real-world graphs often noisy may contain unknown...

Recently, meta learning providing multiple initializations has drawn much attention due to its capability of handling multi-modal tasks from diverse distributions. However, because the difference class distribution between meta-training and meta-test domain, domain shift occurs in meta-learning setting. To improve performance on tasks, we propose multi-initialization with adaptation (MIML-DA) tackle such shift. MIML-DA consists a modulation network novel separation (MSN), where is encode...

Matrix factorization is one of the most successful methods for single-criterion recommender systems but not multi-criteria that contain multiple criterion-specific ratings. Tensor have been developed to learn predictive models in by dealing with three-dimensional (3D) user-item-criterion However, they suffer from data sparsity issues real applications. In order alleviate this problem, we propose deep tensor (DTF) paper integrating representation learning and factorization, where side...

In few-shot image classification scenarios, meta-learning methods aim to learn transferable feature representations extracted from seen domains (base classes) in the meta-training phase and quickly adapt unseen (novel meta-testing phase. However, when have a large discrepancy, existing approaches do not perform well due incapability of generalizing domains. this paper, we investigate challenging domain generalized problem. We design an Meta Regularization Network (MRN) domain-invariant...

Offline reinforcement learning (RL) aims to learn a policy using only pre-collected and fixed data. Although avoiding the time-consuming online interactions in RL, it poses challenges for out-of-distribution (OOD) state actions often suffers from data inefficiency training. Despite many efforts being devoted addressing OOD actions, latter (data inefficiency) receives little attention offline RL. To address this, this paper proposes cross-domain which assumes incorporate additional...

It is of significance for an agent to autonomously explore the environment and learn a widely applicable general-purpose goal-conditioned policy that can achieve diverse goals including images text descriptions. Considering such perceptually-specific goals, one natural approach reward with prior non-parametric distance over embedding spaces states goals. However, this may be infeasible in some situations, either because it unclear how choose suitable measurement, or (heterogeneous)...

In recent years, graph contrastive learning has achieved promising node classification accuracy using neural networks (GNNs), which can learn representations in an unsupervised manner. However, such cannot be generalized to unseen novel classes with only few-shot labeled samples spite of exhibiting good performance on seen classes. order assign generalization capability learning, we propose multimodal meta (MGMC) this paper, integrates into learning. On one hand, MGMC accomplishes...

Compositional zero-shot learning (CZSL) refers to recognizing unseen compositions of known visual primitives, which is an essential ability for artificial intelligence systems learn and understand the world. While considerable progress has been made on existing benchmarks, we suspect whether popular CZSL methods can address challenges few-shot few referential compositions, common when in real-world environments. To this end, study challenging reference-limited compositional (RL-CZSL) problem...

In this paper, a novel nonlinear Radial Basis Function Neural Network (RBF-NN) ensemble model based on ν-Support Vector Machine (SVM) regression is presented for financial time series forecasting. the process of modeling, first stage initial data set divided into different training sets by used Bagging and Boosting technology. second stage, these are input to individual RBF-NN models, then various single predictors produced diversity principle. third Partial Least Square (PLS) technology...