Abstract: Embodiments of the present invention provide systems, methods, and computer storage media for determining an increased matching for large graphs in which an increased matching is generated for the graph by leveraging an initial matching for a small fraction of edges of the large graph. An initial matching for a random subset of edges of an input graph is leveraged to generate alternating paths based on the initially matched edges and the remaining edges, not included in the random subset. An increased matching for the entire graph includes the alternating paths without the initial matched edges, thus increasing the number of matched edges in the increased matching by at least one for every initially matched edge. Graph-based tasks may then be triggered based on the increased matching.

Abstract: In implementations of multi-item influence maximization, a computing device can obtain updates to a user association graph that indicates social correspondence between users, and obtain updates to a user-item graph that indicates user correspondence with one or more items. The computing device includes an influence maximization module that can update an item association graph that indicates item correspondence of each item with one or more other items, where the item association graph can be updated based on the user-item graph that indicates the user correspondence with one or more of the items. The influence maximization module can then iteratively determine a resource allocation for each of the users to maximize user influence of multiple items that are associated in the item association graph and based on the social correspondence between the users, as well as assign a variable portion of the resource allocation to any number of the users.

Abstract: Techniques and systems are described for analytics system entity resolution. Typed higher-order node combinations are determined within a dataset, and an amount of similarity between two arbitrary nodes within the dataset is determined based on the typed higher-order node combinations. The amount of similarity enables the digital analytics to accurately perform source resolution of portions of the dataset to a respective source, and may be utilized to control output of digital content to a client device.

Abstract: In some embodiments, a network analysis system receives network data in the form of a temporal graph that includes nodes and edges. Each node represents an entity involved in a network. An edge connects two nodes to indicate an association between the two nodes. Each edge also has a temporal value indicating a time point when the association between the two nodes was created. The network analysis system generates a sequence of nodes by traversing the nodes in the temporal graph along edges with non-decreasing temporal values or with non-increasing temporal values. The network analysis system further replaces the identifiers of the nodes in the sequence to generate a sequence of feature values. Based on the sequence of feature values, the network analysis system determines network embeddings for the nodes in the temporal graph. Using the network embeddings, the network analysis system identifies two or more of the nodes in the temporal graph that belong to the same entity.

Abstract: A system is disclosed for identifying and counting typed graphlets in a heterogeneous network. A methodology implementing techniques for the disclosed system according to an embodiment includes identifying typed k-node graphlets occurring between any two selected nodes of a heterogeneous network, wherein the nodes are connected by one or more edges. The identification is based on combinatorial relationships between (k?1)-node typed graphlets occurring between the two selected nodes of the heterogeneous network. Identification of 3-node typed graphlets is based on computation of typed triangles, typed 3-node stars, and typed 3-paths associated with each edge connecting the selected nodes. The method further includes maintaining a count of the identified k-node typed graphlets and storing those graphlets with non-zero counts. The identified graphlets are employed for applications including visitor stitching, user profiling, outlier detection, and link prediction.

Abstract: In implementations of higher-order graph clustering and embedding, a computing device receives a heterogeneous graph representing a network. The heterogeneous graph includes nodes that each represent a network entity and edges that each represent an association between two of the nodes in the heterogeneous graph. To preserve node-type and edge-type information, a typed graphlet is implemented to capture a connectivity pattern and the types of the nodes and edges. The computing device determines a frequency of the typed graphlet in the graph and derives a weighted typed graphlet matrix to sort graph nodes. Sorted nodes are subsequently analyzed to identify node clusters having a minimum typed graphlet conductance score. The computing device is further implemented to determine a higher-order network embedding for each of the nodes in the graph using the typed graphlet matrix, which can then be concatenated into a matrix representation of the network.

Abstract: The present disclosure relates to training a recommendation model to generate trait recommendations using one permutation hashing and populated-value-slot-based densification. In particular, the disclosed systems can train the recommendation model by computing sketch vectors corresponding to traits using one permutation hashing. The disclosed systems can then fill in unpopulated value slots of the sketch vectors using populated-value-slot-based densification. The disclosed systems can combine the resulting densified sketches to generate the trained recommendation model. For example, in some embodiments, the disclosed systems can combine the sketches by generating a plurality of locality sensitive hashing tables based on the sketches. In some embodiments, the disclosed systems generate a count sketch matrix based on the sketches and generate trait embeddings based on the count sketch matrix using spectral embedding.

Abstract: Embodiments of the present invention are generally directed to generating figure captions for electronic figures, generating a training dataset to train a set of neural networks for generating figure captions, and training a set of neural networks employable to generate figure captions. A set of neural networks is trained with a training dataset having electronic figures and corresponding captions. Sequence-level training with reinforced learning techniques are employed to train the set of neural networks configured in an encoder-decoder with attention configuration. Provided with an electronic figure, the set of neural networks can encode the electronic figure based on various aspects detected from the electronic figure, resulting in the generation of associated label map(s), feature map(s), and relation map(s).

Abstract: Various embodiments describe techniques for making inferences from graph-structured data using graph convolutional networks (GCNs). The GCNs use various pre-defined motifs to filter and select adjacent nodes for graph convolution at individual nodes, rather than merely using edge-defined immediate-neighbor adjacency for information integration at each node. In certain embodiments, the graph convolutional networks use attention mechanisms to select a motif from multiple motifs and select a step size for each respective node in a graph, in order to capture information from the most relevant neighborhood of the respective node.

Abstract: In implementations of time-dependent network embedding, a computing device maintains time-dependent interconnected data in the form of a time-based graph that includes nodes and node associations that each represent an edge between two of the nodes in the time-based graph based at least in part on a temporal value that indicates when the two nodes were associated. The computing device includes a network embedding module that is implemented to traverse one or more of the nodes in the time-based graph along the node associations, where the traversal is performed with respect to the temporal value of each of the edges that associate the nodes. The network embedding module is also implemented to determine a time-dependent embedding for each of the nodes traversed in the time-based graph, the time-dependent embedding for each of the respective nodes being representative of feature values that describe the respective node.

Abstract: In implementations of higher-order network embedding, a computing device maintains interconnected data in the form of a graph that represents a network, the graph including nodes that each represent entities in the network and node associations that each represent edges between the nodes in the graph. The computing device includes a network embedding module that is implemented to determine a frequency of k-vertex motifs for each of the edges in the graph, and derive motif-based matrices from the frequency of each of the k-vertex motifs in the graph. The network embedding module is also implemented to determine a higher-order network embedding for each of the nodes in the graph from each of the motif-based matrices. The network embedding module can then concatenate the higher-order network embeddings into a matrix representation.

Abstract: Embodiments of the present invention provide systems, methods, and computer storage media for latent summarization of a graph. Structural features can be captured from feature vectors associated with each node of the graph by applying base functions on the feature vectors and iteratively applying relational operators to successive feature matrices to derive deeper inductive relational functions that capture higher-order structural information in different subgraphs of increasing size (node separations). Heterogeneity can be summarized by performing capturing features in appropriate subgraphs (e.g., node-centric neighborhoods associated with each node type, edge direction, and/or edge type). Binning and/or dimensionality reduction can be applied to the resulting feature matrices.

Abstract: A system is provided for facilitating multi-label classification. During operation, the system maintains a set of training vectors. A respective vector represents an object and is associated with one or more labels that belong to a label set. After receiving an input vector, the system determines a similarity value between the input vector and one or more training vectors. The system further determines one or more labels associated with the input vector based on the similarity values between the input vector and the training vectors and their corresponding associated labels.

Abstract: In implementations of higher-order network embedding, a computing device maintains interconnected data in the form of a graph that represents a network, the graph including nodes that each represent entities in the network and node associations that each represent edges between the nodes in the graph. The computing device includes a network embedding module that is implemented to determine a frequency of k-vertex motifs for each of the edges in the graph, and derive motif-based matrices from the frequency of each of the k-vertex motifs in the graph. The network embedding module is also implemented to determine a higher-order network embedding for each of the nodes in the graph from each of the motif-based matrices. The network embedding module can then concatenate the higher-order network embeddings into a matrix representation.

Abstract: In implementations of time-dependent network embedding, a computing device maintains time-dependent interconnected data in the form of a time-based graph that includes nodes and node associations that each represent an edge between two of the nodes in the time-based graph based at least in part on a temporal value that indicates when the two nodes were associated. The computing device includes a network embedding module that is implemented to traverse one or more of the nodes in the time-based graph along the node associations, where the traversal is performed with respect to the temporal value of each of the edges that associate the nodes. The network embedding module is also implemented to determine a time-dependent embedding for each of the nodes traversed in the time-based graph, the time-dependent embedding for each of the respective nodes being representative of feature values that describe the respective node.

Abstract: One embodiment provides a system for facilitating anomaly detection and characterization. During operation, the system determines, by a computing device, a first set of testing data which includes a plurality of data points, wherein the first set includes a data series for a first variable and one or more second variables. The system identifies anomalies by dividing the first set into a number of groups and performing an inter-quartile range analysis on data in each respective group. The system obtains, from the first set, a second set of testing data which includes a data series from a recent time period occurring before a current time, and which further includes a first data point from the identified anomalies. The system classifies the first data point as a first type of anomaly based on whether a magnitude of a derivative of the second set is greater than a first predetermined threshold.

Abstract: One embodiment provides a system for facilitating anomaly detection. During operation, the system determines, by a computing device, a set of testing data which includes a plurality of data points, wherein the set includes a data series for a first variable and one or more second variables, and wherein the one or more second variables are dependent on the first variable. The system divides the set of testing data into a number of groups based on a type of the data series. The system determines an inter-quartile range for a respective group. The system classifies a first testing data point in the respective group as an anomaly based on the inter-quartile range for the respective group, thereby enhancing data mining and outlier detection for the data series for multiple variables.

Abstract: A method of deep graph representation learning includes: deriving a set of base features; and automatically developing, by a processing device, a multi-layered hierarchical graph representation based on the set of base features, wherein each successive layer of the multi-layered hierarchical graph representation leverages an output from a previous layer to learn features of a higher-order.

Abstract: System and methods for relational time-series learning are provided. Unlike traditional time series forecasting techniques, which assume either complete time series independence or complete dependence, the disclosed system and method allow time series forecasting that can be performed on multivariate time series represented as vertices in graphs with arbitrary structures and predicting a future classification for data items represented by one of nodes in the graph. The system and methods also utilize non-relational, relational, temporal data for classification, and allow using fast and parallel classification techniques with linear speedups. The system and methods are well-suited for processing data in a streaming or online setting and naturally handle training data with skewed or unbalanced class labels.

Abstract: One embodiment provides a system for facilitating anomaly detection. During operation, the system determines, by a computing device, a set of training instances, wherein a training instance represents a single class of data within a predefined range. The system computes a similarity score for each testing instance in a set of testing instances, wherein the similarity score is based on a similarity function which takes as input a respective testing instance and the set of training instances. The system determines a boundary threshold based on an ordering of the similarity score for each testing instance. The system classifies a first testing instance as an anomaly responsive to determining that the first testing instance falls outside the boundary threshold, thereby enhancing data mining and outlier detection in the single class of data using unlabeled training instances.