Although augmentations (e.g., perturbation of graph edges, image crops) boost the efficiency Contrastive Learning (CL), feature level augmentation is another plausible, complementary yet not well researched strategy. Thus, we present a novel spectral argumentation for contrastive learning on graphs (and images). To this end, each data view, estimate low-rank approximation per map and subtract that from to obtain its complement. This achieved by proposed herein incomplete power iteration,...

Xiaocui Yang, Shi Feng, Yifei Zhang, Daling Wang. Proceedings of the 59th Annual Meeting Association for Computational Linguistics and 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Graph contrastive learning (GCL) improves graph representation learning, leading to SOTA on various downstream tasks. The augmentation step is a vital but scarcely studied of GCL. In this paper, we show that the node embedding obtained via augmentations highly biased, somewhat limiting models from discriminative features for tasks.Thus, instead investigating in input space, alternatively propose perform hidden (feature augmentation). Inspired by so-called matrix sketching, COSTA, novel...

Trustworthy artificial intelligence (AI) technology has revolutionized daily life and greatly benefited human society. Among various AI technologies, Federated Learning (FL) stands out as a promising solution for diverse real-world scenarios, ranging from risk evaluation systems in finance to cutting-edge technologies like drug discovery sciences. However, challenges around data isolation privacy threaten the trustworthiness of FL systems. Adversarial attacks against privacy, learning...

Trustworthy artificial intelligence (TAI) has proven invaluable in curbing potential negative repercussions tied to AI applications. Within the TAI spectrum, federated learning (FL) emerges as a promising solution safeguard personal information distributed settings across multitude of practical contexts. However, realm FL is not without its challenges. Especially worrisome are adversarial attacks targeting algorithmic robustness and systemic confidentiality. Moreover, presence biases opacity...

Serious concerns have been raised about the role of `socialbots' in manipulating public opinion and influencing outcome elections by retweeting partisan content to increase its reach. Here we analyze influence socialbots on Twitter determining how they contribute retweet diffusions. We collect a large dataset tweets during 1st U.S. presidential debate 2016 1.5 million users from three perspectives: user influence, political behavior (partisanship engagement) botness. First, define measure...

The incompleteness of knowledge graphs triggers considerable research interest in relation prediction. As the key to predicting relations among entities, many efforts have been devoted learning embeddings entities and by incorporating a variety neighbors' information which includes not only from direct outgoing incoming neighbors but also ones indirect on multihop paths. However, previous models usually consider entity paths limited length or ignore sequential Either simplification will make...

Graph Neural Networks (GNNs) have demonstrated great power for the semi-supervised node classification task. However, most GNN methods are sensitive to noise of graph structures. structure learning (GSL) is then introduced robustification, which contains two major parts: recovering optimal and fine-tuning parameters on this generated downstream Nonetheless, existing GSL solutions merely focus features during first module generation exploit label information only by back-propagation second...

Searching on bipartite graphs is basal and versatile to many real-world Web applications, e.g., online recommendation, database retrieval, query-document searching. Given a query node, the conventional approaches rely similarity matching with vectorized node embeddings in continuous Euclidean space. To efficiently manage intensive computation, developing hashing techniques for graph structured data has recently become an emerging research direction. Despite retrieval efficiency Hamming...

Graph Neural Networks (GNNs) have achieved remarkable success by extending traditional convolution to learning on non-Euclidean data. The key the GNNs is adopting neural message-passing paradigm with two stages: aggregation and update. current design of considers topology information in stage. However, updating stage, all nodes share same function. identical function treats each node embedding as i.i.d. random variables thus ignores implicit relationships between neighborhoods, which limits...

Maximizing the user-item engagement based on vectorized embeddings is a standard procedure of recent recommender models. Despite superior performance for item recommendations, these methods however implicitly deprioritize modeling user-wise similarity in embedding space; consequently, identifying similar users underperforming, and additional processing schemes are usually required otherwise. To avoid thorough model re-training, we propose WSFE, model-agnostic training-free representation...

Abstract Differential privacy can effectively help federated learning resist attacks from various parties. However, existing approaches that use differential for protection greatly decrease the model performance of learning, especially in scenarios with complex structures and large parameters. In this paper, we propose a novel preservation scheme combines automatic gradient clipping transformation perturbation. Our approach primarily reduces impact on two aspects. Firstly, efficiently...

Purpose The paper aims to investigate current library instruction programs help business students make better use of resources and improve their information. However, students’ information acquisition ability, usage perception toward user education are inevitably changing along with the rapidly evolving landscape as well socio-cultural environment driven by technologies. Design/methodology/approach For this study, 90 from three different majors at Faculty Business Economics, University Hong...

The financial sector presents many opportunities to apply various machine learning techniques. Centralized creates a constraint which limits further applications in finance sectors. Data privacy is fundamental challenge for variety of and insurance that account on model across different sections. In this paper, we define new practical scheme collaborative one party owns data, but another labels only, term \textbf{Asymmetrically Collaborative Machine Learning}. For scheme, propose novel...

Graph representation learning via Contrastive Learning (GCL) has drawn considerable attention recently. Efforts are mainly focused on gathering more global information contrasting a single high-level graph view, which, however, underestimates the inherent complex and hierarchical properties in many real-world networks, leading to sub-optimal embeddings. To incorporate these of graph, we propose Cross-Scale Knowledge Synergy (CGKS), generic feature framework, advance contrastive with enhanced...

Contrastive learning, a form of Self-Supervised Learning (SSL), typically consists an alignment term and regularization term. The minimizes the distance between embeddings positive pair, while prevents trivial solutions expresses prior beliefs about embeddings. As widely used technique, soft decorrelation has been employed by several non-contrastive SSL methods to avoid solutions. While is designed address issue dimensional collapse, we find that it fails achieve this goal theoretically...

The semi-supervised multi-label classification problem primarily deals with Euclidean data, such as text a 1D grid of tokens and images 2D pixels. However, the non-Euclidean graph-structured data naturally constantly appears in learning tasks from various domains like social networks, citation protein-protein interaction (PPI) networks. Moreover, existing popular node embedding methods, Graph Neural Networks (GNN), focus on graphs simplex labels tend to neglect label correlations setting, so...

Serious concerns have been raised about the role of 'socialbots' in manipulating public opinion and influencing outcome elections by retweeting partisan content to increase its reach. Here we analyze influence socialbots on Twitter determining how they contribute retweet diffusions. We collect a large dataset tweets during 1st U.S. Presidential Debate 2016 (#DebateNight) 1.5 million users from three perspectives: user influence, political behavior (partisanship engagement) botness. First,...

Concept relatedness estimation (CRE) aims to determine whether two given concepts are related. Existing methods only consider the pairwise relationship between concepts, while overlooking higher-order that could be encoded in a concept-level graph structure. We discover this underlying satisfies set of intrinsic properties CRE, including reflexivity, commutativity, and transitivity. In paper, we formalize CRE introduce structure named ConcreteGraph. To address data scarcity issue novel...

With the evolution of Knowledge Graphs (KGs), new entities emerge which are not seen before. Representation learning KGs in such an inductive setting aims to capture and transfer structural patterns from existing entities. However, performance methods limited by sparsity implicit transfer. In this paper, we propose VMCL, a Contrastive Learning (CL) framework with graph guided Variational autoencoder on Meta-KGs setting. We first representation generation encoded generated representations...