Understanding the training dynamics of deep neural networks (DNNs) is important as it can lead to improved training efficiency and task performance. Recent works have demonstrated that representing the wirings of neurons in feedforward DNNs as graphs is an effective strategy for understanding how architectural choices can affect performance. However, these approaches fail to model training dynamics since a single, static graph cannot capture how DNNs change over the course of training. Thus, in this work, we propose a compact, expressive temporal graph framework that effectively captures the dynamics of many workhorse architectures in computer vision. Specifically, our framework extracts an informative summary of graph properties (e.g., degree, eigenvector centrality) over a sequence of DNN graphs obtained during training. We demonstrate that the proposed framework captures useful dynamics by accurately predicting trained, task performance when using a summary over early training epochs (<5) across four different architectures and two image datasets. Moreover, by using a novel, highly-scalable DNN graph representation, we further demonstrate that the proposed framework captures generalizable dynamics as summaries extracted from smaller-width networks are effective when evaluated on larger widths.

Temporal graph embedding has been widely studied thanks to its superiority in tasks such as prediction and recommendation. Despite the advances in algorithms and novel frameworks such as deep learning, there has been relatively little work on systematically studying the properties of temporal network models and their cornerstones, the graph time-series representations that are used in these approaches. This paper aims to fill this gap by introducing a general framework that extends an arbitrary existing static embedding approach to handle dynamic tasks, and conducting a systematic study of seven base static embedding methods and six temporal network models. Our framework generalizes static node embeddings derived from the time-series representation of stream data to the dynamic setting by modeling the temporal dependencies with classic models such as the reachability graph. While previous works on dynamic modeling and embedding have focused on representing a stream of timestamped edges using a time-series of graphs based on a specific time-scale (\eg, 1 month), we introduce the notion of an ε-graph time-series that uses a fixed number of edges for each graph, and show its superiority in practical settings over the standard solution. From the 42 methods that our framework subsumes, we find that leveraging the new ε-graph time-series representation and capturing temporal dependencies with the proposed reachability or summary graph tend to perform well. Furthermore, the new dynamic embedding methods based on our framework perform comparably and on average better than the state-of-the-art embedding methods designed specifically for temporal graphs in link prediction tasks.

While most network embedding techniques model the proximity between nodes in a network, recently there has been significant interest in structural embeddings that are based on node equivalences , a notion rooted in sociology: equivalences or positions are collections of nodes that have similar roles—i.e., similar functions, ties or interactions with nodes in other positions—irrespective of their distance or reachability in the network. Unlike the proximity-based methods that are rigorously evaluated in the literature, the evaluation of structural embeddings is less mature. It relies on small synthetic or real networks with labels that are not perfectly defined, and its connection to sociological equivalences has hitherto been vague and tenuous. With new node embedding methods being developed at a breakneck pace, proper evaluation, and systematic characterization of existing approaches will be essential to progress. To fill in this gap, we set out to understand what types of equivalences structural embeddings capture. We are the first to contribute rigorous intrinsic and extrinsic evaluation methodology for structural embeddings, along with carefully-designed, diverse datasets of varying sizes. We observe a number of different evaluation variables that can lead to different results (e.g., choice of similarity measure, classifier, and label definitions). We find that degree distributions within nodes’ local neighborhoods can lead to simple yet effective baselines in their own right and guide the future development of structural embedding. We hope that our findings can influence the design of further node embedding methods and also pave the way for more comprehensive and fair evaluation of structural embedding methods.

Tag inference is an important task in the business of video platforms with wide applications such as recommendation, interpretation, and more. Existing works are mainly based on extracting video information from multiple modalities such as frames or music, and then infer tags through classification or object detection. This, however, does not apply to inferring generic tags or taxonomy that are less relevant to video contents, such as video originality or its broader category, which are important in practice. In this paper, we claim that these generic tags can be modeled through the semantic relations between videos and tags, and can be utilized simultaneously with the multi-modal features to achieve better video tagging. We propose TransFusion, an end-to-end supervised learning framework that fuses multi-modal embeddings (e.g., vision, audio, texts, etc.) with the knowledge embedding to derive the video representation. To infer the diverse tags following heterogeneous relations, TransFusion adopts a dual attentive approach to learn both the modality importance in fusion and relation importance in inference. Besides, it is general enough and can be used with the existing translation-based knowledge embedding approaches. Extensive experiments show that TransFusion outperforms the baseline methods with lowered mean rank and at least 9.59% improvement in [email protected] on the real-world video knowledge graph.

Multi-source entity linkage focuses on integrating knowledge from multiple sources by linking the records that represent the same real world entity. This is critical in high-impact applications such as data cleaning and user stitching. The state-of-the-art entity linkage pipelines mainly depend on supervised learning that requires abundant amounts of training data. However, collecting well-labeled training data becomes expensive when the data from many sources arrives incrementally over time. Moreover, the trained models can easily overfit to specific data sources, and thus fail to generalize to new sources due to significant differences in data and label distributions. To address these challenges, we present AdaMEL, a deep transfer learning framework that learns generic high-level knowledge to perform multi-source entity linkage. AdaMEL models the attribute importance that is used to match entities through an attribute-level self-attention mechanism, and leverages the massive unlabeled data from new data sources through domain adaptation to make it generic and data-source agnostic. In addition, AdaMEL is capable of incorporating an additional set of labeled data to more accurately integrate data sources with different attribute importance. Extensive experiments show that our framework achieves state-of-the-art results with 8.21% improvement on average over methods based on supervised learning. Besides, it is more stable in handling different sets of data sources in less runtime.

Structural roles define sets of structurally similar nodes that are more similar to nodes inside the set than outside, whereas communities define sets of nodes with more connections inside the set than outside. Roles based on structural similarity and communities based on proximity are fundamentally different but important complementary notions. Recently, the notion of structural roles has become increasingly important and has gained a lot of attention due to the proliferation of work on learning representations (node/edge embeddings) from graphs that preserve the notion of roles. Unfortunately, recent work has sometimes confused the notion of structural roles and communities (based on proximity) leading to misleading or incorrect claims about the capabilities of network embedding methods. As such, this article seeks to clarify the misconceptions and key differences between structural roles and communities, and formalize the general mechanisms (e.g., random walks and feature diffusion) that give rise to community- or role-based structural embeddings. We theoretically prove that embedding methods based on these mechanisms result in either community- or role-based structural embeddings. These mechanisms are typically easy to identify and can help researchers quickly determine whether a method preserves community- or role-based embeddings. Furthermore, they also serve as a basis for developing new and improved methods for community- or role-based structural embeddings. Finally, we analyze and discuss applications and data characteristics where community- or role-based embeddings are most appropriate.

Email is ubiquitous in the workplace. Naturally, machine learning models that make third-party email clients "smarter" can dramatically impact employees' productivity and efficiency. Motivated by this potential, we study the task of professional role inference from email data, which is crucial for email prioritization and contact recommendation systems. The central question we address is: Given limited data about employees, as is common in third-party email applications, can we infer where in the organizational hierarchy these employees belong based on their email behavior? Toward our goal, in this paper we study professional role inference on a unique new email dataset comprising billions of email exchanges across thousands of organizations. Taking a network approach in which nodes are employees and edges represent email communication, we propose EMBER, or EMBedding Email-based Roles, which finds email-centric embeddings of network nodes to be used in professional role inference tasks. EMBER automatically captures behavioral similarity between employees in the email network, leading to embeddings that naturally distinguish employees of different hierarchical roles. EMBER often outperforms the state-of-the-art by 2-20% in role inference accuracy and 2.5-344x in speed. We also use EMBER with our unique dataset to study how inferred professional roles compare between organizations of different sizes and sectors, gaining new insights into organizational hierarchy.

Motivated by the computational and storage challenges that dense embeddings pose, we introduce the problem of latent network summarization that aims to learn a compact, latent representation of the graph structure with dimensionality that is independent of the input graph size (\i.e., #nodes and #edges), while retaining the ability to derive node representations on the fly. We propose Multi-LENS, an inductive multi-level latent network summarization approach that leverages a set of relational operators and relational functions (compositions of operators) to capture the structure of egonets and higher-order subgraphs, respectively. The structure is stored in low-rank, size-independent structural feature matrices, which along with the relational functions comprise our latent network summary. Multi-LENS is general and naturally supports both homogeneous and heterogeneous graphs with or without directionality, weights, attributes or labels. Extensive experiments on real graphs show 3.5-34.3% improvement in AUC for link prediction, while requiring 80-2152x less output storage space than baseline embedding methods on large datasets. As application areas, we show the effectiveness of Multi-LENS in detecting anomalies and events in the Enron email communication graph and Twitter co-mention graph.

As graph datasets grow, faster data mining methods become indispensable. Random Walk with Restart (RWR), belief propagation, semi-supervised learning, and more graph methods can be expressed as a set of linear equations. In this work, we focus on solving such equations fast and accurately when large number of queries need to be handled. We use RWR as a case study, since it is widely used not only to evaluate the importance of a node, but also as a basis for more complex tasks, e.g., representation learning and community detection. We introduce a new, intuitive two-step divide-and-conquer formulation and a corresponding parallelizable method, FlowR, for solving RWR with two goals: (i) fast and accurate computation under multiple queries; (ii) one-time message exchange between subproblems. We further speed up our proposed method by extending our formulation to carefully designed overlapping subproblems (FlowR-OV) and by leveraging the strengths of iterative methods (FlowR-Hyb). Extensive experiments on synthetic and real networks with up to ∼8 million edges show that our methods are accurate and outperform in runtime various state-of-the-art approaches, running up to 34× faster in preprocessing and up to 32× faster in query time.

Given the soaring amount of data being generated daily, graph mining tasks are becoming increasingly challenging, leading to tremendous demand for summarization techniques. Feature selection is a representative approach that simplifies a dataset by choosing features that are relevant to a specific task, such as classification, prediction, and anomaly detection. Although it can be viewed as a way to summarize a graph in terms of a few features, it is not well-defined for exploratory analysis, and it operates on a set of observations jointly rather than conditionally (i.e., feature selection from many graphs vs. selection for an input graph conditioned on other graphs). In this work, we introduce EAGLE (Exploratory Analysis of Graphs with domain knowLEdge), a novel method that creates interpretable, feature-based, and domain-specific graph summaries in a fully automatic way. That is, the same graph in different domains-e.g., social science and neuroscience-will be described via different EAGLE summaries, which automatically leverage the domain knowledge and expectations. We propose an optimization formulation that seeks to find an interpretable summary with the most representative features for the input graph so that it is: diverse, concise, domain-specific, and efficient. Extensive experiments on synthetic and real-world datasets with up to ~1M edges and ~400 features demonstrate the effectiveness and efficiency of EAGLE and its benefits over existing methods. We also show how our method can be applied to various graph mining tasks, such as classification and exploratory analysis.

Given a large graph with several millions or billions of nodes and edges, such as a social network, how can we explore it efficiently and find out what is in the data? In this demo we present P erseus , a large-scale system that enables the comprehensive analysis of large graphs by supporting the coupled summarization of graph properties and structures, guiding attention to outliers, and allowing the user to interactively explore normal and anomalous node behaviors. Specifically, P erseus provides for the following operations: 1) It automatically extracts graph invariants ( e.g. , degree, PageRank, real eigenvectors) by performing scalable, offline batch processing on H adoop ; 2) It interactively visualizes univariate and bivariate distributions for those invariants; 3) It summarizes the properties of the nodes that the user selects; 4) It efficiently visualizes the induced subgraph of a selected node and its neighbors, by incrementally revealing its neighbors. In our demonstration, we invite the audience to interact with P erseus to explore a variety of multi-million-edge social networks including a Wikipedia vote network, a friendship/foeship network in Slashdot, and a trust network based on the consumer review website Epinions.com.