Abstract: System and method for training a recommender system (RS). The RS is configured to make recommendations in respect of a bipartite graph that comprises a plurality of user nodes, a plurality of item nodes, and an observed graph topology that defines edges connecting at least some of the user nodes to some of the item nodes, the RS including an existing graph neural network (GNN) model configured by an existing set of parameters. The method includes: applying a loss function to compute an updated set of parameters for an updated GNN model that is trained with a new graph using the first set of parameters as initialization parameters, the loss function being configured to distil knowledge based on node embeddings generated by the existing GNN model in respect of an existing graph, wherein the new graph includes a plurality of user nodes and a plurality of item nodes that are also included in the existing graph; and replacing the existing GNN model of the RS with the updated GNN model.

Abstract: A system control method includes receiving a control task and randomly selecting a chromosome from an evolution pool according to the control task. The selected chromosome is decoded to obtain (N+1) ensemble policies, where the chromosome includes (N+1) gene fragments, and where N is a positive integer greater than or equal to 1. Each gene fragment can uniquely correspond to an ensemble policy, and each ensemble policy can uniquely correspond to a preset function. One ensemble policy is used for assigning a weight to a preset function that uniquely corresponds to the ensemble policy. The evolution pool can maintain two or more chromosomes. An ensemble calculation is performed according to weights assigned by the (N+1) ensemble policies to obtain an ensemble control output. A control signal is generated according to the ensemble control output, where the control signal is used for performing system control.

Abstract: A network data transmission method is provided. A switch device receives one or more data flows, classifies each of the received data flows into one of two classes according to data features of the data flow by using a decision tree model established by a flow table pipeline of the switch device. If a data flow belongs to a first class, the switch device reports the data flow to a controller, so that the controller computes a transmission path for the data flow. If a data flow belongs to a second class, the switch device obtains a transmission path for the data flow according to local flow table information, and transmits the data flow according to the obtained transmission path. Data flows are classified and filtered by using a switch, so as to improve network transmission efficiency while ensuring bearing capability of a network control system.

Abstract: Embodiments of this application provide a service survivability analysis method and apparatus, and relate to the field of communications technologies, so as to shorten duration of service survivability analysis and improve efficiency of the service survivability analysis. The method includes: obtaining a link fault record and network topology information that are in a preset time period; determining a similarity between any two links in all faulty links based on fault occurrence time and fault removal time of the any two links in the link fault record and connection information of network devices on the any two links, to obtain a link similarity matrix; performing clustering on all the faulty links based on the link similarity matrix, to obtain at least one link cluster; and performing survivability analysis on services on at least two preset links based on each of the at least one link cluster.

Abstract: A network data flow classification method related to artificial intelligence includes collecting an information set, including a plurality of pieces of dimension information, of a to-be-processed data flow, establishing a static behavior model and a dynamic behavior model of each piece of dimension information in the information set, where the static behavior model represents a value selection rule of the dimension information, and the dynamic behavior model represents a correlation relationship of the dimension information between two adjacent time moments, obtaining, using the static behavior model and the dynamic behavior model respectively, a static model distance and a dynamic model distance between the to-be-processed data flow and a data flow of each target application type, determining an application type of the to-be-processed data flow based on the static model distance and the dynamic model distance.

Abstract: A data visualization method and apparatus, where the method includes displaying a first density distribution diagram on a first map, where the first density distribution diagram represents density distribution, in a region, of source locations of flow events whose destinations are located in a target reference region, and displaying a second density distribution diagram on a second map, where the second density distribution diagram represents density distribution, in a region, of destinations of flow events whose source locations are located in the target reference region. Hence, bidirectional density distribution associated with each other using the target reference region are displayed on two maps in a linked manner, thereby implementing visualization of bidirectional density distribution data.

Abstract: A method for use by a network controller is provided. The controller collects network status data of a current moment, and estimates network status information of the current moment according to the network status data of the current moment. The controller predicts network status information of a next moment according to the network status information of the current moment. The controller receives and preprocessing a dataflow entering the network at the current moment, to obtain a preprocessing result of the dataflow. The controller generates a control action by using a network control policy of the current moment and according to the network status information of the current moment, the predicted network status information of the next moment, and the preprocessing result of the dataflow. The controller obtains feedback information of the network resulted from the control action, and generates a network control policy of the next moment.

Abstract: Embodiments of this application provide a method for generating a routing control action in a software-defined network and a related device, to provide optimum control actions for the SDN. The method includes: obtaining a current network state parameter of the SDN; determining a Q function of the SDN based on the current network state parameter of the SDN and a deep neural network model, where the deep neural network model is determined based on a current topology structure of the SDN; and determining a routing control action for the SDN based on the Q function and a link state parameter of each link in the SDN. In the technical solution, the deep neural network model is combined with a Q-learning algorithm of reinforcement learning, and optimum control actions can be determined.

Abstract: A coflow identification method includes: obtaining a weighted matrix by means of learning according to historical data in the network, where the weighted matrix is used to minimize a feature distance between data streams belonging to a same coflow and maximize a feature distance between data streams belonging to different coflows; computing a feature distance between any two data streams in the network according to metrics in the data stream layer data feature, the application layer data stream feature distance, the terminal aspect data feature distance, and the weighted matrix; and dividing the data streams in the network into several cluster sets by using a clustering algorithm and according to the feature distance between the any two data streams in the network, where each of the several cluster sets is a coflow.

Abstract: Embodiments of this application provide a service survivability analysis method and apparatus, and relate to the field of communications technologies, so as to shorten duration of service survivability analysis and improve efficiency of the service survivability analysis. The method includes: obtaining a link fault record and network topology information that are in a preset time period; determining a similarity between any two links in all faulty links based on fault occurrence time and fault removal time of the any two links in the link fault record and connection information of network devices on the any two links, to obtain a link similarity matrix; performing clustering on all the faulty links based on the link similarity matrix, to obtain at least one link cluster; and performing survivability analysis on services on at least two preset links based on each of the at least one link cluster.

Abstract: A network data flow classification method related to artificial intelligence includes collecting an information set, including a plurality of pieces of dimension information, of a to-be-processed data flow, establishing a static behavior model and a dynamic behavior model of each piece of dimension information in the information set, where the static behavior model represents a value selection rule of the dimension information, and the dynamic behavior model represents a correlation relationship of the dimension information between two adjacent time moments, obtaining, using the static behavior model and the dynamic behavior model respectively, a static model distance and a dynamic model distance between the to-be-processed data flow and a data flow of each target application type, determining an application type of the to-be-processed data flow based on the static model distance and the dynamic model distance.

Abstract: A method for detecting a data flow type includes obtaining a header of a first data packet of a current data flow and a pattern vector of the current data flow from the header; comparing the at least one feature dimension in the pattern vector of the current data flow with a corresponding feature dimension in a pattern vector of at least one historical data flow, so as to obtain at least one pattern similarity of the current data flow; predicting a length of the current data flow according to the at least one pattern similarity of the current data flow and a length of the corresponding at least one historical data flow; and comparing the predicted length of the current data flow with a preset threshold, and determining whether the current data flow is a large data flow or a small data flow according to a comparison result.

Abstract: Embodiments of this application provide a method for generating a routing control action in a software-defined network and a related device, to provide optimum control actions for the SDN. The method includes: obtaining a current network state parameter of the SDN; determining a Q function of the SDN based on the current network state parameter of the SDN and a deep neural network model, where the deep neural network model is determined based on a current topology structure of the SDN; and determining a routing control action for the SDN based on the Q function and a link state parameter of each link in the SDN. In the technical solution, the deep neural network model is combined with a Q-learning algorithm of reinforcement learning, and optimum control actions can be determined.

Abstract: The present application provides a path planning method and a controller. The method includes: acquiring data flow information of a to-be-transmitted job in a software-defined network, where the job includes at least one target data flow, and the data flow information of the job includes: a source address, a destination address, and a volume of each target data flow; and performing path planning according to the data flow information, and obtaining a job transmission path used to ensure that the job is transmitted in the software-defined network in a shortest job transmission time, where the job transmission path includes a transmission path corresponding to each target data flow in the job. The present application improves a data transmission speed of a job in an SDN network.

Abstract: A coflow identification method includes: obtaining a weighted matrix by means of learning according to historical data in the network, where the weighted matrix is used to minimize a feature distance between data streams belonging to a same coflow and maximize a feature distance between data streams belonging to different coflows; computing a feature distance between any two data streams in the network according to metrics in the data stream layer data feature, the application layer data stream feature distance, the terminal aspect data feature distance, and the weighted matrix; and dividing the data streams in the network into several cluster sets by using a clustering algorithm and according to the feature distance between the any two data streams in the network, where each of the several cluster sets is a coflow.

Abstract: A data visualization method and apparatus, where the method includes displaying a first density distribution diagram on a first map, where the first density distribution diagram represents density distribution, in a region, of source locations of flow events whose destinations are located in a target reference region, and displaying a second density distribution diagram on a second map, where the second density distribution diagram represents density distribution, in a region, of destinations of flow events whose source locations are located in the target reference region. Hence, bidirectional density distribution associated with each other using the target reference region are displayed on two maps in a linked manner, thereby implementing visualization of bidirectional density distribution data.

Abstract: A method for use by a network controller is provided. The controller collects network status data of a current moment, and estimates network status information of the current moment according to the network status data of the current moment. The controller predicts network status information of a next moment according to the network status information of the current moment. The controller receives and preprocessing a dataflow entering the network at the current moment, to obtain a preprocessing result of the dataflow. The controller generates a control action by using a network control policy of the current moment and according to the network status information of the current moment, the predicted network status information of the next moment, and the preprocessing result of the dataflow. The controller obtains feedback information of the network resulted from the control action, and generates a network control policy of the next moment.

Abstract: A network data transmission method is provided. A switch device receives one or more data flows, classifies each of the received data flows into one of two classes according to data features of the data flow by using a decision tree model established by a flow table pipeline of the switch device. If a data flow belongs to a first class, the switch device reports the data flow to a controller, so that the controller computes a transmission path for the data flow. If a data flow belongs to a second class, the switch device obtains a transmission path for the data flow according to local flow table information, and transmits the data flow according to the obtained transmission path. Data flows are classified and filtered by using a switch, so as to improve network transmission efficiency while ensuring bearing capability of a network control system.

Abstract: A system control method includes: receiving a control task; randomly selecting a chromosome from an evolution pool according to the control task, and decoding the chromosome to obtain (N+1) ensemble policies, where the chromosome includes (N+1) gene fragments, each gene fragment uniquely corresponds to an ensemble policy, each ensemble policy uniquely corresponds to a preset function, one ensemble policy is used for assigning a weight to a preset function that uniquely corresponds to the ensemble policy, the evolution pool maintains two or more chromosomes, and N is a positive integer greater than or equal to 1; performing an ensemble calculation according to weights assigned by the (N+1) ensemble policies, to obtain an ensemble control output; and generating a control signal according to the ensemble control output, where the control signal is used for performing system control.

Abstract: A method for detecting a data flow type includes obtaining a header of a first data packet of a current data flow and a pattern vector of the current data flow from the header; comparing the at least one feature dimension in the pattern vector of the current data flow with a corresponding feature dimension in a pattern vector of at least one historical data flow, so as to obtain at least one pattern similarity of the current data flow; predicting a length of the current data flow according to the at least one pattern similarity of the current data flow and a length of the corresponding at least one historical data flow; and comparing the predicted length of the current data flow with a preset threshold, and determining whether the current data flow is a large data flow or a small data flow according to a comparison result.