Abstract: In an example embodiment, a deep learning model is used to learn embedding representations of a heterogeneous information network, where the embedding represents entity-specific properties and network environment properties. Position-aware embeddings specific to the heterogeneous information network may be used as input features of the deep learning model. Furthermore, meta-path embedding specific to the heterogeneous information network may also be used as input features of the deep learning model. Modified embedding propagation methods are further designed to explore better ways to capture network meta-path properties.

Abstract: In an example embodiment, a deep learning model is used to learn embedding representations of a heterogeneous information network, where the embedding represents entity-specific properties and network environment properties. Position-aware embeddings specific to the heterogeneous information network may be used as input features of the deep learning model. Furthermore, meta-path embedding specific to the heterogeneous information network may also be used as input features of the deep learning model. Modified embedding propagation methods are further designed to explore better ways to capture network meta-path properties.

Abstract: An online connection server is configured to more accurately predict connections for a viewing member of an online connection network. The online connection server may implement a machine-learning model that uses prior interactions by the viewing member to determine those connections that are likely to lead to more substantial interactions with the viewing member. The machine-learning model may be implemented using a reinforcement learning technique, such as a Deep Q network. The online connection server may further implement a state representation module that generates a state from a graph-based embedding of the viewing member profile, where the state is used to train the machine-learning model and determine an optimal candidate to recommend as a connection for the viewing member.

Abstract: Techniques for generating latent representations for entities based on a network graph are provided. In one technique, an artificial neural network is trained based on feature vectors of entities and feature vectors of neighbors of those entities. The neighbors are determined based on a graph of nodes representing the entities. Two nodes are connected if, for example, the corresponding entities are connected in an online network, one entity transmitted an online communication to the other entity, or one entity interacted with content associated with the other entity. Once trained, the artificial neural network is used to generate latent representations for entities represented by the graph. Latent representations may be used in multiple ways. For example, a similarity between two latent representations may be used to determine an order of candidate content items to present to an entity corresponding to one of the latent representations.

Abstract: An online connection server is configured to more accurately predict connections for a viewing member of an online connection network. The online connection server may implement a machine-learning model that uses prior interactions by the viewing member to determine those connections that are likely to lead to more substantial interactions with the viewing member. The machine-learning model may be implemented using a reinforcement learning technique, such as a Deep Q network. The online connection server may further implement a state representation module that generates a state from a graph-based embedding of the viewing member profile, where the state is used to train the machine-learning model and determine an optimal candidate to recommend as a connection for the viewing member.

Abstract: Techniques for generating latent representations for entities based on a network graph are provided. In one technique, an artificial neural network is trained based on feature vectors of entities and feature vectors of neighbors of those entities. The neighbors are determined based on a graph of nodes representing the entities. Two nodes are connected if, for example, the corresponding entities are connected in an online network, one entity transmitted an online communication to the other entity, or one entity interacted with content associated with the other entity. Once trained, the artificial neural network is used to generate latent representations for entities represented by the graph. Latent representations may be used in multiple ways. For example, a similarity between two latent representations may be used to determine an order of candidate content items to present to an entity corresponding to one of the latent representations.

Abstract: In an embodiment, the disclosed technologies include computing a score for a node pair including first and second nodes of a digital connection graph; where nodes of the digital connection graph represent members of an online system; where the online system uses the digital connection graph to determine a runtime decision related to a member represented by the first node; where the score indicates a predicted likelihood of interaction, during a time interval, after a digital connection between the first and second nodes of the node pair; where the predicted likelihood of interaction is determined by comparing a set of statistics computed for the node pair to a digital model; where the digital model has been created using data extracted from post-connection interactions in the online system between members whose nodes are connected in the digital connection graph; causing the score to modify the runtime decision.