CN112085279A

CN112085279A - Method and device for training interaction prediction model and predicting interaction event

Info

Publication number: CN112085279A
Application number: CN202010955099.6A
Authority: CN
Inventors: 文剑烽; 常晓夫; 宋乐; 刘旭钦
Original assignee: Alipay Hangzhou Information Technology Co Ltd
Current assignee: Alipay Hangzhou Information Technology Co Ltd
Priority date: 2020-09-11
Filing date: 2020-09-11
Publication date: 2020-12-15
Anticipated expiration: 2040-09-11
Also published as: CN112085279B

Abstract

The embodiment of the specification provides a method and a device for training and using an interactive prediction model. In the method, a dynamic interaction graph is constructed based on an interaction event sequence, and a sample node pair comprising a first node and a second node is determined from the dynamic interaction graph and corresponds to a sample interaction event. And obtaining a generation vector and a discrimination vector corresponding to each of the two nodes from the first generation network and the first discrimination network. With the second generation network, the time of occurrence of the sample interaction event is predicted based on the two generation vectors. Then, a first input is formed based on the two generated vectors and the predicted time, a second input is formed based on the two discrimination vectors and the real time, and the probabilities that the first input and the second input are each a real event are discriminated through a second discrimination network. A second generation network and a second discrimination network are trained based on the confrontation. The trained first generation network and the second generation network are used as interaction prediction models for predicting interaction events.

Description

Method and device for training interaction prediction model and predicting interaction event

Technical Field

One or more embodiments of the present specification relate to the field of machine learning, and more particularly, to training an interaction prediction model, and a method and apparatus for predicting an interaction event using the trained interaction prediction model.

Background

In many scenarios, user interaction events need to be analyzed and processed. The interaction event is one of basic constituent elements of an internet event, for example, a click action when a user browses a page can be regarded as an interaction event between the user and a content block of the page, a purchase action in an e-commerce can be regarded as an interaction event between the user and a commodity, and an inter-account transfer action is an interaction event between the user and the user. The characteristics of fine-grained habit preference and the like of the user and the characteristics of an interactive object are contained in a series of interactive events of the user, and the characteristics are important characteristic sources of a machine learning model. Therefore, in many scenarios, it is desirable to perform feature expression and modeling on interaction participants according to interaction events, and further perform analysis on interaction objects and events, especially on the security of the interaction events, so as to guarantee the security of the interaction platform.

However, an interactive event involves both interacting parties, and the status of each party itself may be dynamically changing, and thus it is very difficult to accurately characterize the interacting parties comprehensively considering their multi-aspect characteristics. Thus, improved solutions for more efficiently analyzing interactive objects and interactive events are desired.

Disclosure of Invention

One or more embodiments of the present disclosure describe a method and an apparatus for training an interaction prediction model, in which an interaction prediction model is obtained by training in an opposing training manner by using a two-stage cascade generated opposing network based on a dynamic interaction graph, so that a next interaction object and an interaction time of the target object can be predicted more accurately.

According to a first aspect, there is provided a method of training an interactive predictive model, the method comprising:

acquiring a dynamic interaction graph, wherein the dynamic interaction graph comprises a plurality of node pairs, each node pair corresponds to an interaction event, two nodes respectively represent two objects participating in the interaction event, and any node points to two nodes corresponding to the last interaction event participated in by the object represented by the node through a connecting edge;

selecting a sample node pair from the dynamic interaction graph, wherein the sample node pair comprises a first node and a second node and is marked with the real occurrence time of a sample interaction event corresponding to the sample node pair;

obtaining, from a first generated network, a first generated vector corresponding to a first node, and a second generated vector corresponding to a second node; obtaining, from a first discrimination network, a first discrimination vector corresponding to a first node and a second discrimination vector corresponding to a second node; the first generating network and the first discriminating network are trained by a first countermeasure training process;

inputting the first generated vector, the second generated vector and the noise vector into a second generated network to obtain the predicted occurrence time of the sample interaction event;

forming a first input based on the first generated vector, a second generated vector and the predicted occurrence time, and obtaining a first probability that the first input corresponds to a real event through a second discrimination network;

forming a second input based on the first discrimination vector, a second discrimination vector and the real occurrence time, and obtaining a second probability that the second input corresponds to a real event through the second discrimination network;

performing a second countermeasure training procedure, including training the second decision network with the goal of increasing the second probability and decreasing the first probability; training the second generation network with a goal of increasing the first probability; the trained first generation network and the second generation network form the interactive prediction model.

In one embodiment, the first generation network and the first discrimination network are two neural networks with the same structure and independent parameters.

According to one embodiment, obtaining, from a first generated network, a first generated vector corresponding to a first node, comprises: determining a first sub-graph formed by nodes in a preset range starting from the root node and reaching through a connecting edge in the dynamic interaction graph by taking the first node as the root node; and inputting the first sub-graph into the first generation network, wherein the first generation network outputs a hidden vector of a root node as the first generation vector according to the node attribute characteristics of each node in the first sub-graph and the connection relation between the nodes.

Further, in an embodiment, the first generation network includes an LSTM layer, where the LSTM layer respectively uses each node from a leaf node to a root node in the input subgraph as a current node, and sequentially iteratively processes each node, where the iterative processing includes determining a hidden vector of the current node according to at least a node attribute feature of the current node and a hidden vector of two nodes pointed by the current node through a connecting edge.

According to one embodiment, when the first generation vector, the second generation vector and the noise vector are input into a second generation network, the first generation vector and the second generation vector can be spliced as an a priori condition vector; sampling from Gaussian distribution to obtain the noise vector; and after splicing the prior condition vector and the noise vector, inputting the prior condition vector and the noise vector into the second generation network.

In one embodiment, the first input is formed by: performing fusion operation on the first generated vector and the second generated vector to obtain a first fusion vector; and splicing and combining the first fusion vector and the predicted occurrence time, and taking a combined result as the first input.

According to one embodiment, the first antagonistic training process comprises:

determining a plurality of candidate nodes corresponding to the first node from the dynamic interaction graph, wherein the candidate nodes comprise the second node;

generating the first generated vector and a plurality of candidate generated vectors corresponding to the plurality of candidate nodes respectively by using the first generated network; generating the first discrimination vector and a plurality of candidate discrimination vectors corresponding to the plurality of candidate nodes respectively by using the first discrimination network;

determining probability distribution of interaction between each candidate node and the first node according to the first generated vector and the candidate generated vectors, and determining a prediction node interacting with the first node according to the probability distribution;

fitting the discrimination vectors of the prediction nodes according to the plurality of candidate discrimination vectors and the probability distribution;

determining a third probability that the predicted node is a real interactive node based on the first discrimination vector and the discrimination vector of the predicted node; determining a fourth probability that the second node is a real interactive node based on the first discrimination vector and the second discrimination vector;

training the first discriminant network with the goal of increasing the fourth probability and decreasing the third probability; training the first generation network with a goal of increasing the third probability.

In one embodiment, two nodes in each node pair in the dynamic interaction graph respectively represent a first class object and a second class object which participate in the interaction event; and, in case the first node belongs to a first class of objects, the plurality of candidate nodes all belong to a second class of objects.

According to one embodiment, a Gumbel-softmax function can be utilized to fit to obtain a probability distribution of interaction between each candidate node and the first node.

According to one embodiment, the first antagonistic training process is completed before the second antagonistic training process.

According to another embodiment, the first antagonistic training process alternates with the second antagonistic training process.

According to a second aspect, there is provided a method of predicting an interaction event, the method comprising:

obtaining an interactive prediction model obtained by training according to the method of the first aspect, wherein the interactive prediction model comprises a first generation network and a second generation network;

acquiring a first target object to be analyzed, and constructing a first target node corresponding to the first target object in the dynamic interaction graph;

obtaining, by a first generating network, a first target vector corresponding to the first target node and a second target vector corresponding to a second target node; wherein the second target node is a node corresponding to an object predicted by the first generation network to have a predicted interaction event with the first target object;

and inputting the first target vector, the second target vector and the noise vector into a second generation network to obtain the predicted occurrence time aiming at the predicted interaction event.

According to a third aspect, there is provided a method of predicting an interaction event, the method comprising:

acquiring a first target node and a second target node to be analyzed from the dynamic interaction graph;

obtaining, by a first generating network, a first target vector corresponding to the first target node and a second target vector corresponding to a second target node;

and inputting the first target vector, the second target vector and the noise vector into a second generation network to obtain predicted occurrence time, wherein the predicted occurrence time represents the expected interaction time of the objects represented by the first target node and the second target node respectively.

According to a fourth aspect, there is provided an apparatus for training an interactive prediction model, the apparatus comprising:

the interactive graph obtaining unit is configured to obtain a dynamic interactive graph, wherein the dynamic interactive graph comprises a plurality of node pairs, each node pair corresponds to one interactive event, two nodes respectively represent two objects participating in the interactive event, and any node points to two nodes corresponding to the last interactive event participated by the object represented by the node through a connecting edge;

the node pair selection unit is configured to select a sample node pair from the dynamic interaction graph, wherein the sample node pair comprises a first node and a second node, and the actual occurrence time of a sample interaction event corresponding to the sample node pair is marked;

a vector acquisition unit configured to acquire, from a first generated network, a first generated vector corresponding to a first node, and a second generated vector corresponding to a second node; obtaining, from a first discrimination network, a first discrimination vector corresponding to a first node and a second discrimination vector corresponding to a second node; the first generating network and the first discriminating network are trained by a first countermeasure training process;

a time prediction unit configured to input the first generation vector, the second generation vector and the noise vector into a second generation network to obtain a predicted occurrence time for the sample interaction event;

a first input discrimination unit configured to form a first input based on the first generated vector, a second generated vector and the predicted occurrence time, and obtain a first probability that the first input corresponds to a real event through a second discrimination network;

a second input discrimination unit configured to form a second input based on the first discrimination vector, a second discrimination vector, and the real occurrence time, and obtain a second probability that the second input corresponds to a real event through the second discrimination network;

a second countermeasure training unit configured to perform a second countermeasure training process including training the second decision network with a target of increasing the second probability and decreasing the first probability; training the second generation network with a goal of increasing the first probability; the trained first generation network and the second generation network form the interactive prediction model.

According to a fifth aspect, there is provided an apparatus for predicting an interaction event, the apparatus comprising:

a model obtaining unit configured to obtain an interactive prediction model obtained by training the apparatus according to the fourth aspect, wherein the interactive prediction model includes a first generation network and a second generation network; (ii) a

the target node acquisition unit is configured to acquire a first target object to be analyzed and construct a first target node corresponding to the first target object in the dynamic interaction graph;

a target vector acquisition unit configured to acquire, through a first generation network, a first target vector corresponding to the first target node, and a second target vector corresponding to a second target node; wherein the second target node is a node corresponding to an object predicted by the first generation network to have a predicted interaction event with the first target object;

and the time prediction unit is configured to input the first target vector, the second target vector and the noise vector into a second generation network to obtain the predicted occurrence time of the predicted interaction event.

According to a sixth aspect, there is provided an apparatus for predicting an interactivity event, the apparatus comprising:

the target node acquisition unit is configured to acquire a first target node and a second target node to be analyzed from the dynamic interaction graph;

a target vector acquisition unit configured to acquire, through a first generation network, a first target vector corresponding to the first target node, and a second target vector corresponding to a second target node;

and the time prediction unit is configured to input the first target vector, the second target vector and the noise vector into a second generation network to obtain predicted occurrence time, and the predicted occurrence time represents the expected interaction time of the objects represented by the first target node and the second target node respectively.

According to a seventh aspect, there is provided a computer readable storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method of any one of the first to third aspects.

According to an eighth aspect, there is provided a computing device comprising a memory and a processor, wherein the memory has stored therein executable code, and the processor, when executing the executable code, implements the method of any one of the first to third aspects.

According to the method and the device provided by the embodiment of the specification, a dynamic interaction graph is constructed based on an interaction event sequence, and a two-stage cascade confrontation generation network GAN is trained based on the dynamic interaction graph. The primary GAN performs countermeasure learning on probability distribution among objects which are interacted to obtain vector representation of nodes in the interaction graph; the second-level GAN is based on the vector representation of the first-level GAN, and the interaction time between the interaction objects is learned and predicted through the countermeasure training, so that the interaction objects and the interaction time can be comprehensively predicted through the two-level cascade GAN.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.

FIG. 1 illustrates an implementation scenario diagram according to one embodiment;

FIG. 2 illustrates a flow diagram of a method of training an interactive predictive model, according to one embodiment;

FIG. 3 illustrates a dynamic interaction sequence and a dynamic interaction diagram constructed therefrom, in accordance with one embodiment;

FIG. 4 illustrates an example of a first sub-graph;

FIG. 5 illustrates a flow of steps of a first countermeasure training process in one embodiment;

FIG. 6 shows a schematic diagram of the structure and training process of a cascaded GAN in one embodiment;

FIG. 7 illustrates a flow diagram of a method of predicting an interaction event, according to one embodiment;

FIG. 8 illustrates a flow diagram of a method of predicting an interaction event, in accordance with another embodiment;

FIG. 9 shows a schematic block diagram of a training apparatus for an interactive predictive model according to an embodiment;

FIG. 10 shows a schematic block diagram of an apparatus for predicting an interaction event, according to an embodiment;

fig. 11 shows a schematic block diagram of an apparatus for predicting an interaction event according to an embodiment.

Detailed Description

The scheme provided by the specification is described below with reference to the accompanying drawings.

As previously mentioned, it is desirable to be able to characterize and model interactive objects and interactive events based on a series of interactive events that occur with the interactive objects.

In one approach, a static interaction relationship network graph is constructed based on historical interaction events, such that individual interaction objects and individual interaction events are analyzed based on the interaction relationship network graph. Specifically, the participants of the historical events can be used as nodes, and connection edges can be established between the nodes with the interaction relationship, so that the interaction network graph is formed. However, the above static network diagram, although it may show the interaction relationships between objects, does not contain timing information of these interaction events. The graph embedding is simply carried out on the basis of the interaction relation network graph, and the obtained feature vectors do not express the influence of the time information of the interaction events on the nodes. Moreover, such static graphs are not scalable enough, and are difficult to flexibly process for the situations of newly added interaction events and newly added nodes.

In view of the above, in accordance with one or more embodiments of the present specification, a dynamically changing sequence of interactivity events is constructed as a dynamic interactivity graph, wherein each interactivity object involved in each interactivity event corresponds to each node in the dynamic interactivity graph. Such a dynamic interaction graph may reflect timing information of interaction events experienced by individual interaction objects. Further, in order to analyze and predict an interaction event based on the dynamic interaction graph, in an embodiment of the present specification, a two-stage cascade generation countermeasure network gan (generic adaptive networks) is used to analyze and predict nodes. The primary GAN performs countermeasure learning on probability distribution among objects which are interacted to obtain vector representation of nodes in the interaction graph; the second-level GAN is based on the vector representation of the first-level GAN, and the interaction time between the interaction objects is learned and predicted through the countermeasure training, so that the interaction objects and the interaction time can be comprehensively predicted through the two-level cascade GAN.

Fig. 1 shows a schematic illustration of an implementation scenario according to an embodiment. As shown in FIG. 1, multiple interaction events occurring in sequence may be organized chronologically into a dynamic interaction sequence<E₁,E₂,…,E_N>Wherein each element E_iRepresenting an interaction event, which may be represented in the form of an interaction feature set E_i＝(a_i,b_i,t_i) Wherein a is_iAnd b_iIs an event E_iTwo interacting objects of, t_iIs the interaction time.

According to an embodiment of the present specification, a dynamic interaction graph is constructed based on the dynamic interaction sequence. In the dynamic interaction graph, each interaction object a in each interaction event is divided into a plurality of interaction objects_i,b_iRepresented by nodes, and establishing parent-child relationship connection edges between nodes containing continuous events of the same object. The structure of the dynamic interaction graph will be described in more detail later.

To more efficiently perform node analysis and event analysis, the nodes in the dynamic interaction graph are first characterized as node vectors by a level one GAN, GAN 1. Specifically, the GAN1 includes a first generation network and a first discriminant network, which respectively use respective network parameters to characterize nodes in the dynamic interaction graph as vectors. For example, for node u and node v that interact at time t, the first generative network may generate vectors g (u) and g (v), respectively, and the first discriminant network may generate vectors d (u) and d (v), respectively. The first generating network and the first discriminating network are trained by counterlearning a probability distribution between objects with which interaction occurs.

The resulting node vector from GAN1 is further input to a secondary GAN, GAN 2. The GAN2 includes a second generation network and a second discrimination network. The second generation network generates an interaction time t' of the node u and the node v based on the node vectors g (u) and g (v) generated by the first generation network. The second discrimination network is used for discrimination and discrimination, based on the generated inputs of g (u), g (v) and t', and the true inputs of d (u), d (v) and t, against the second generation network.

Through the countermeasure training of two-stage cascade GAN, a first generation network in GAN1 can generate vector representation of each node and predict interactive nodes; the second generation network in GAN2 can predict the interaction time of the pairs of interaction nodes based on the vector representations of the pairs of interaction nodes obtained by GAN1, thereby enabling a comprehensive analysis and prediction of interaction events.

Specific implementations of the above concepts are described below.

FIG. 2 illustrates a flow diagram of a method of training an interactive predictive model, according to one embodiment. It is to be appreciated that the method can be performed by any apparatus, device, platform, cluster of devices having computing and processing capabilities. The following describes each step in the training method shown in fig. 2 with reference to a specific embodiment.

First, in step 21, a dynamic interaction graph reflecting an interaction event correlation is obtained.

Generally, a plurality of interaction events occurring in sequence can be organized into an interaction event sequence according to the time sequence as described above, and a dynamic interaction graph is constructed based on the interaction event sequence, so as to reflect the incidence relation of the interaction events. Sequences of interaction events, e.g. expressed as<E₁,E₂,…,E_N>May comprise a plurality of interactivity events arranged in chronological order, wherein each interactivity event E_iCan be represented as an interactive feature set E_i＝(a_i,b_i,t_i) Wherein a is_iAnd b_iIs an event E_iTwo interacting objects of, t_iIs the interaction time.

In one embodiment, the two interaction objects involved in each interaction event belong to two classes of objects, hereinafter referred to as first class objects and second class objects. For example, in an e-commerce platform, an interaction event may be a purchase behavior of a user, two of which may be a user object (first class object) and a merchandise object (second class object). In another example, the interaction event may be a click action of a user on a page tile, where the two objects may be a user object (first class object) and a page tile object (second class object). In yet another example, the interaction event may be a recommendation event, such as a user (object of the first type) accepting a recommendation (object of the second type) to push the recommendation, and the recommendation may be various types of pushable contents, such as a movie, a product, an article, and so on. In other business scenarios, an interaction event may also be other interaction behavior that occurs between two different classes of objects. It should be understood that although the user object is taken as the first type object in the above examples, this is not necessarily so, and the setting of the first type object and the second type object may be determined according to the interaction events, some interaction events do not include the user object, and in the interaction events including the user object, the user object may be taken as the second type object.

For the dynamic interaction sequence described above, a dynamic interaction graph may be constructed. Specifically, a pair of nodes (two nodes) is used for representing two objects related to one interactive event, and each object in each interactive event in the dynamic interactive sequence is represented by a node respectively. In case an interactivity event involves two types of objects, the two nodes of a node pair represent, respectively, one first type of object and one second type of object participating in an interactivity event. It is to be understood that one node may correspond to one object in one interaction event, but the same physical object may correspond to multiple nodes. For example, if user U1 purchased commodity a1 at time t1 and purchased commodity a2 at time t2, then there are two feature groups of interaction events (U1, a1, t1) and (U1, a2, t2), then two nodes U1(t1), U1(t2) are created for user U1 based on these two interaction events, respectively. It can therefore be considered that a node in the dynamic interaction graph corresponds to the state of an interaction object in one interaction event.

For each node in the dynamic interaction graph, a connecting edge is constructed in the following way: for any node i, assuming that it corresponds to an interactivity event i (with an interaction time t), in the dynamic interaction sequence, tracing back from the interactivity event i, i.e. tracing back to a direction earlier than the interaction time t, determines the first interactivity event j (with an interaction time t-, t-earlier than t) which also contains the object represented by the node i as the last interactivity event in which the object participates. Thus, a connecting edge is established pointing from node i to both nodes in the last interactivity event j. The two pointed-to nodes are then also referred to as the associated nodes of node i.

The following description is made in conjunction with specific examples. FIG. 3 illustrates a dynamic interaction sequence and a dynamic interaction diagram constructed therefrom, according to one embodiment. In particular, the left side of FIG. 3 shows a dynamic interaction sequence organized in time order, wherein an exemplary illustration is given at t respectively₁,t₂,…,t₆Interaction event E occurring at a moment₁,E₂,…,E₆Each interaction event contains two interaction objects involved in the interaction and an interaction time (for clarity of illustration, the event feature is omitted), wherein the first column of objects is user class objects such as David, Lucy, etc., and the second column of objects is item class objects such as movie names, which are denoted by M1 to M4 in the figure. The right side of fig. 3 shows a dynamic interaction diagram constructed according to the dynamic interaction sequence on the left side, wherein two interaction objects in each interaction event are respectively taken as nodes. And any node points to two nodes corresponding to the last interactive event participated by the object represented by the node through the connecting edge. E.g. node u (t)₆) Representing an interaction event E₆The last interactive event involved by the user is E₄Then, node u (t)₆) By connecting edge direction E₄Corresponding two nodes u (t)₄) And w (t)₄). Similarly, u (t)₄) Direction E₂The corresponding two nodes, and so on. In this manner, connecting edges are constructed between nodes, thereby forming the dynamic interaction graph of FIG. 3.

The above describes a way and a process for constructing a dynamic interaction graph based on a dynamic interaction sequence. For the training process shown in fig. 2, the process of constructing the dynamic interaction graph may be performed in advance or in the field. Accordingly, in one embodiment, at step 21, a dynamic interaction graph is constructed in the field according to the dynamic interaction sequence. Constructed as described above. In another embodiment, the dynamic interaction graph can be constructed in advance based on the dynamic interaction sequence. In step 21, the formed dynamic interaction graph is read or received.

On the basis of obtaining the dynamic interaction graph, in step 22, a sample node pair is selected from the dynamic interaction graph, wherein the sample node pair includes a first node and a second node, and the actual occurrence time of the sample interaction event corresponding to the sample node pair is marked.

It is to be understood that a sample node pair is a pair of nodes chosen for training the model that are known to interact and whose interaction time is known. The two nodes in the pair are referred to as the first node and the second node, respectively, and the interaction between the two nodes is referred to as a sample interaction event.

For example, in FIG. 3, assume that event E was selected₆The corresponding two nodes are sample node pairs, and the first node can be u (t)₆) The second node is v (t)₆)，E₆Is a sample interaction event with real occurrence time t₆. Of course, other interactive node pairs may be selected as the sample node pairs.

For simplicity of description, the first node is denoted as node u, the second node is denoted as node v, and the actual occurrence time of the sample interaction event is denoted as t.

Next, in step 23, the first node u and the second node v are characterized by using the first generation network and the first discriminant network included in the first stage GAN, respectively. That is, a first generated vector g (u) corresponding to a first node u, and a second generated vector g (v) corresponding to a second node v are acquired from a first generated network; a first discriminant vector d (u) corresponding to the first node u and a second discriminant vector d (v) corresponding to the second node v are obtained from the first discriminant network.

In one embodiment, the first generation network and the first discrimination network are two neural networks with the same structure and independent parameters, and the two neural networks can respectively use network parameters thereof to characterize any target node in the dynamic interaction graph to obtain a corresponding node vector. The following description will take an example of a process in which the first generation network performs vector representation with the first node u as a target node. It is to be understood, however, that the process may also be applied to the characterization of any node by the first discrimination network.

Specifically, according to an embodiment, when the first node u needs to be characterized, a subgraph including the first node u, that is, a first subgraph, may be first extracted from the dynamic interaction graph. Specifically, in one example, the first node u is taken as a root node, and a sub-graph formed by nodes in a predetermined range, starting from the root node and reaching via a connecting edge, is determined in the dynamic interaction graph as a first sub-graph.

Fig. 4 shows an example of a first sub-graph. In FIG. 4, assume u (t)₆) Is the first node. Then, with this node u (t)₆) Traversing along the connection edge direction of the parent-child relationship for the root node, and determining the nodes in the preset range. In the example of fig. 4, it is assumed that the predetermined range is a child node reached via at most a preset number K of connection edges of 2. Then, from the current root node u (t)₆) Starting from this, the nodes that can be reached via 2 connecting edges are shown as dashed areas in the figure. The node and the connection relation in the region are the node u (t)₆) The corresponding sub-graph, i.e. the first sub-graph. It will be appreciated that in other examples, the predetermined range may be set in other manners, including, for example, a predetermined length of time to backtrack forward, etc.

Upon obtaining a first subgraph for the first node u, such first subgraph may be input to a first generation network. The first generation network outputs the hidden vector of the first node as a corresponding first generation vector according to the node attribute characteristics of each node in the first subgraph and the connection relation between the nodes.

In one example, the first generation network may be a graph neural network that is graph-processed by way of graph embedding, such as a graph convolution neural network GNN. In the graph embedding process, according to the connection relation among the nodes, each node in the input subgraph is used as a K-order neighbor node of a root node, neighbor node aggregation is carried out, and a hidden vector of a target node is obtained and used as a generated vector of the hidden vector.

In another example, the first generation network may include a temporal recursion layer, such as a recurrent neural network RNN layer, or a long-short term memory LSTM layer, so that a subgraph with the target node as the root node is graph-processed in a recursive iterative manner. Specifically, in one embodiment, the first generation network includes an LSTM layer, and the LSTM layer respectively uses each node from a leaf node to a root node in the input subgraph as a current node, and sequentially iteratively processes each node, where the iterative processing includes determining a hidden vector of the current node according to at least a node attribute feature of the current node and hidden vectors of two nodes pointed by the current node through a connecting edge. Through iterative processing from leaf nodes to root nodes, a hidden vector of the root node is finally obtained and used as a generated vector.

The above-mentioned node attribute characteristics differ depending on the different classes of objects that the node represents. For example, where a node represents a user, the node attribute characteristics may include attribute characteristics of the user, such as age, occupation, education, location, registration duration, crowd labels, and so forth; where the nodes represent items, the node attribute characteristics may include attribute characteristics of the items, such as item category, time on shelf, sales volume, number of reviews, and so forth. And under the condition that the node represents other interactive objects, the node attribute characteristics can be obtained correspondingly based on the attributes of the interactive objects.

Thus, in the above various ways, the first generation network performs vector characterization for the first node u, which is characterized as a first generation vector g (u). The first generated network may be similarly characterized for other nodes in the dynamic interaction graph. Thus, for the second node v, the first generated network may characterize it as a second generated vector g (v).

As described above, the first discrimination network and the first generation network have the same network structure and algorithm. Therefore, the first discrimination network may also process any target node in the dynamic interaction graph to obtain a discrimination vector of the target node based on the subgraph including the target node in the above manner. Accordingly, correspondingly, for the first node u and the second node v which are taken as the samples above, a first discriminant vector d (u) corresponding to the first node u and a second discriminant vector d (v) corresponding to the second node v can be obtained through the first discriminant network.

The manner in which the first generation network and the first discrimination network process to obtain the characterization vector of each node is described above. As described above, the first generation network and the first discrimination network have the same network structure and algorithm, but have independent network parameters. As two sub-networks of the first-level GAN, the respective network parameters of the first generation network and the first discriminant network are trained by a first countermeasure training process that targets interaction probability distribution between objects for countermeasure learning.

The steps of the first antagonistic training process are described in detail below.

FIG. 5 illustrates a flow of steps of a first countermeasure training process in one embodiment. As shown in fig. 5, to train the first level GAN, a plurality of candidate nodes corresponding to the first node, including the second node, are first determined from the dynamic interaction graph at step 51. It is to be understood that the candidate nodes are all possible nodes that are selected for the purpose of the training of the GAN that are likely to be another node of the pair of nodes where the first node is located. Thus, the candidate nodes include a second node that actually interacts with the first node to form the node pair. The set of candidate nodes constitutes a sampling space V.

In one embodiment, the objects represented by the nodes in the dynamic interaction graph may be divided into a first class of objects and a second class of objects, and each interaction event occurs between the two classes of objects, so that two nodes in each node pair represent a first class of objects and a second class of objects, respectively. It is assumed here that the first node u is an object of the first type. In one example, when selecting the candidate node corresponding to the first node u, all nodes belonging to the second class object in the dynamic interaction graph may be considered as candidate nodes. In another embodiment, when the dynamic interaction graph is relatively large and the number of nodes exceeds a certain threshold, nodes which are within a certain range from the first node u and belong to the second class of objects in the dynamic interaction graph are taken as candidate nodes. The certain range may be, for example, a node within a certain time range from the occurrence time of the interaction event where the first node u is located.

The description is continued with the example of fig. 3. Suppose node u (t)₆) Is a first node which representsObject David, a user object. Therefore, the node corresponding to the item class object in the dynamic interaction graph can be selected as the alternative node. In one example, the nodes corresponding to all the item class objects (M1 to M4) in FIG. 3 are all used as the alternative nodes, wherein the second node v (t) is necessarily included₆)。

Next, in step 52, a first generated vector g (u) corresponding to the first node and a plurality of candidate generated vectors corresponding to the plurality of candidate nodes are generated using the first generated network; then, a first discriminant vector d (u) corresponding to the first node u and a plurality of candidate discriminant vectors corresponding to the plurality of candidate nodes are generated by using the first discriminant network. The process of generating the corresponding vector for any target node by the first generating network and the first judging network is as described above, and is not described again.

Fig. 6 shows a schematic diagram of the structure and training process of a cascaded GAN in one embodiment. In fig. 6, the first generation network is denoted as G1 and the first discriminant network is denoted as D1. Each alternative node in the sampling space V corresponding to the first node u is respectively represented as V₁,v₂,…,v_n. Correspondingly, the first generation network G1 generates a first generation vector G (u) for the first node u, and generates candidate generation vectors G (v) for the candidate nodes, respectively₁),g(v₂),…,g(v_n). On the other hand, the first discrimination network D1 generates a first discrimination vector D (u) for the first node u, and generates candidate discrimination vectors D (v) for the candidate nodes, respectively₁),d(v₂),…,d(v_n)。

Then, in step 53, a probability distribution of interaction between each candidate node and the first node is determined according to the first generated vector and the candidate generated vectors, and a predicted node interacting with the first node is determined according to the probability distribution.

In one example, for alternate node v_iThe probability p (v) that it interacts with the first node u can be determined by the following equation (1)_i| u), thereby obtaining the probability distribution P of interaction between each candidate node and the node u.

Wherein v is_jTo belong to any candidate node of the sampling space V, f (u, V)_i) For computing node u and node v_iA function of interaction probability dependent on a first generation vector g (u) corresponding to node u and node v_iCorresponding generated vector g (v)_i)。

For example, in one specific example:

f(u,v_i)＝g(u)·g(v_i) (2)

i.e., f (u, v)_i) Is vector g (u) and vector g (v)_i) Dot product result of (c).

In another example, f (u, v)_i) Can be defined as based on vectors g (u) and g (v)_i) And calculating the cosine similarity. In more examples, more ways based on g (u) and g (v) may be used_i) Obtaining a candidate node v_iAnd the probability of interaction with node u.

According to one embodiment, on the basis of the formula (1) above, a softmax function is adopted to fit the candidate nodes v_jProbability p (v) of interaction with first sample node u_i|u)：

In another example, the Gumbel-softmax function is utilized to fit to obtain the probability distribution of interaction between each candidate node and the first node u. The Gumbel-softmax function is based on softmax fitting, random sampling under Gumbel distribution is further introduced, so that the problem of possible gradient disconnection caused by sampling from discrete distribution is avoided, and the Gumbel-softmax function is more beneficial to network training and parameter adjustment.

On the basis of obtaining the probability distribution of interaction between each candidate node and the first node, a prediction node v' interacting with the first node can be determined according to the probability distribution.

In one embodiment, the candidate node with the highest probability may be selected as the predicted node v' from the probability distribution. In another embodiment, a virtual node is used as the predicted node v', the virtual node can be represented by a one-hot vector, and the virtual node corresponds to the probability distribution.

The prediction node v' and the probability distribution may then be passed to the first discrimination network. Accordingly, at step 54, a plurality of candidate decision vectors d (v) corresponding to the plurality of candidate nodes, respectively, are determined₁),d(v₂),…,d(v_n) And fitting the discrimination vector d (v ') of the prediction node v' according to the probability distribution.

As described above, in one embodiment, the predicted node v 'is one node selected from the plurality of candidate nodes, and in this case, the discrimination vector corresponding to the selected node is directly used as the discrimination vector d (v') of the predicted node. In another embodiment, the predicted node v' is a virtual node. In this case, based on the probability distribution, the discrimination vectors of the candidate nodes are weighted and summed, and the discrimination vector d (v ') of the predicted node v' is obtained by fitting:

d(v′)＝∑w_i*d(v_i) (4)

wherein the alternative node v_iCorresponding weight w_iInteraction probability p (v) obtained according to formula (1) or formula (3)_i| u). Thus, a discrimination vector d (v ') of the predicted node v' is obtained.

It is understood that the prediction node is a node interacting with the first node u, which is determined according to the generation vector of the first generation network for each node, and thus the first discriminant vector d (u) of the first node u and the discriminant vector d (v') corresponding to the prediction node can be used as the representation of the interaction event generated by the first generation network. On the other hand, since it is known that the first node u and the second node v constitute a node pair, the first discrimination vector d (u) corresponding to the first node u and the second discrimination vector d (v) corresponding to the second node may be used as representations of the real interaction event. Then, the above two representations are respectively input into the first discrimination network for discrimination.

That is, in step 55, based on the first discriminant vector d (u) and the discriminant vector d (v') of the predicted node, the probability that the input is represented as a true event, or the probability that the predicted node is a true interaction node, is determined. This probability is referred to as a third probability P3. On the other hand, based on the first discrimination vector d (u) and the second discrimination vector d (v), the probability that the second node is the true interactive node, i.e. the fourth probability P4, is determined.

Next, a countermeasure training is performed based on the probabilities, i.e., the first discriminant network is trained at step 56 with the goal of increasing the fourth probability and decreasing the third probability. That is, the training goal of the first discriminant network is to be able to distinguish whether the input representation is a real event. On the other hand, the first generation network is trained with a goal of increasing the third probability. That is, the training goal of the first generated network is that the generated interactivity events are sufficient to be determined to be real interactivity events.

In this way, the first generation network and the first discrimination network are trained by a first countermeasure training process of performing countermeasure learning on a probability distribution between objects with which interaction occurs. The first generation network and the first discriminant network trained in the way can better represent each node in the dynamic interaction graph, and the represented vector can well fit the real interaction situation among the nodes.

Returning to fig. 2. Based on the node characterization capabilities of the first generation network and the first discriminant network in the first stage GAN, as described above, in step 23, a first generation vector g (u) and a second generation vector g (v) corresponding to the first node u and the second node v, respectively, are obtained from the first generation network, and a first discriminant vector d (u) corresponding to the first node u and a second discriminant vector d (v) corresponding to the second node are obtained from the first discriminant network. The second stage GAN can then be trained based on the above vectors.

Specifically, in step 24, the first generation vector g (u), the second generation vector g (v) and the noise vector are input into the second generation network to obtain the predicted occurrence time t' for the sample interaction event.

In one embodiment, the first and second generated vectors g (u), (g), (v) may be input as a priori conditions for the second generating network, such that the second generating network generates the predicted time of occurrence based on the a priori conditions. More specifically, in one example, the first generated vector g (u) and the second generated vector g (v) may be concatenated as a priori conditional vectors. In addition, sampling is performed from the gaussian distribution to obtain a noise vector. And after splicing the prior condition vector and the noise vector, inputting the spliced prior condition vector and the noise vector into a second generation network. The second generation network then outputs a predicted occurrence time t' based on the input stitching vector.

The second discrimination network is used for discriminating the generated event representation and the real event representation containing time. Thus, the generated event representation and the real event representation may be constructed separately and input to the second discrimination network for discrimination.

Specifically, in step 25, a first input is formed based on the first generated vector g (u), the second generated vector g (v) and the predicted occurrence time t', and a first probability P1 that the first input corresponds to the real event is obtained through the second discriminant network. It will be appreciated that the first and second generated vectors g (u), (g), (v) are both vectors generated by the first generated network, and the predicted time of occurrence t' is the time generated by the second generated network, so that the first input corresponds to the generation of the event representation.

The first input may be configured and formed in a variety of ways. In one example, a first generated vector g (u) and a second generated vector g (v) may be fused to obtain a first fused vector. The fusion operation may include various operations such as splicing, summing, weighted summing, bit-wise multiplying, etc. Then, the first fusion vector and the predicted occurrence time t' may be subjected to stitching combination, and the combination result may be used as a first input.

On the other hand, in step 26, a second input is formed based on the first discriminant vector d (u), the second discriminant vector d (v) and the real occurrence time t, and a second probability P2 that the second input corresponds to the real event is obtained through a second discriminant network. It will be appreciated that the first discrimination vector d (u) and the second discrimination vector d (v) are vectors processed by the first discrimination network, and t is the true time of occurrence of the sample event, and thus the second input corresponds to a true event representation. The second input is configured and formed in correspondence with the first input.

Next, a second countermeasure training process for the second stage GAN is performed based on the first probability P1 and the second probability P2 described above. Specifically, in step 27, the second decision network is trained with the goal of increasing the second probability P2 and decreasing the first probability P1; the second generation network is trained with the goal of increasing the first probability P1.

It can be seen that the training goal of the second decision network is to give as small a prediction probability as possible for the generated event representation (first input) (which indicates the probability that the input is a true event), and as large a prediction probability as possible for the true event representation (second input), so as to distinguish the generated event representation from the true event representation as much as possible. And the training target of the second generation network is that the generation event representation constructed based on the generated predicted occurrence time t 'is enough to be judged as a real event representation, that is, the predicted occurrence time t' is close to the real occurrence time t as much as possible. The second generating network and the second discriminating network form a countermeasure against the training target of the first probability P1.

An example of a process for second countermeasure training is described with continued reference to fig. 6. In fig. 6, the second stage GAN includes a second generation network G2 and a second discrimination network D2.

As described above, in the first countermeasure training process, the first generation network G1 generates corresponding generation vectors for the first node u and each candidate node. A first generated vector G (u) corresponding to the first node and a second generated vector G (v) generated by the second node may be obtained therefrom and input as a priori conditions to the second generating network G2, so that G2 generates the predicted occurrence time t' based on the a priori conditions and noise. Further, the generation vector g (u), g (v) and the predicted occurrence time t' may form a first input to the second decision network D2.

On the other hand, the first discrimination network D1 generates corresponding discrimination vectors for the first node u and each candidate node. A first discrimination vector D (u) corresponding to the first node and a second discrimination vector D (v) corresponding to the second node may be obtained therefrom, and together with the real occurrence time t of the sample event, a second input may be formed and input to the second discrimination network D2.

Thus, the second discrimination network D2 may discriminate and distinguish the first input and the second input, respectively, with the discrimination result for the first input competing against the training targets of the second generation network G2.

In one embodiment, a first countermeasure training process may be performed first, after completing training of the first generation network and the first discriminant network in the first stage GAN, parameters in the first stage GAN are fixed, and then a second countermeasure training process is performed to train the second generation network and the second discriminant network.

In another embodiment, the first and second antagonistic training processes may be performed alternately, so as to train the first stage GAN and the second stage GAN alternately, and the prediction performance of the GANs is improved alternately and gradually.

Through the first antagonistic training and the second antagonistic training, the two-stage cascade GAN network can be finally obtained. After training is completed, the interactive prediction model may be formed using only the generating networks in the two levels of GAN, i.e., the first generating network and the second generating network.

The following describes a process of performing interactive event prediction using the interactive prediction model formed above.

FIG. 7 illustrates a flow diagram of a method of predicting an interaction event, according to one embodiment. The method may be performed by any apparatus, device, platform, cluster of devices having computing, processing capabilities. As shown in fig. 7, first, in step 71, an interactive prediction model trained according to the above method is obtained, where the interactive prediction model includes a first generation network and a second generation network. In other words, the first discriminant network and the second discriminant network are used for training the first generating network and the second generating network in an antagonistic manner, and in a prediction stage after the training is completed, the first discriminating network and the second discriminating network can be not used any more, but the first generating network and the second generating network form an interactive prediction model.

Then, in step 72, a dynamic interaction graph reflecting the association relationship of the interaction events is obtained, which is similar to step 21 of fig. 2 and is not repeated.

Next, in step 73, a first target object to be analyzed is obtained, and a first target node corresponding to the first target object is constructed in the dynamic interaction graph. Specifically, a single node may be newly added in the dynamic interaction graph, the single node representing the first target object. Similar to the other nodes, a connecting edge is established between the single node and two nodes corresponding to the last interactive event in which the first target object participates, so that the single node points to the two nodes of the last interactive event through the connecting edge. However, unlike other nodes, since the interaction object is unknown, the single node exists independently and does not have a corresponding node forming the node pair.

Then, at step 74, a first target vector corresponding to the first target node and a second target vector corresponding to a second target node are obtained via a first generation network; and the second target node is a node which is predicted by the first generation network and corresponds to an object which is predicted to generate a predicted interaction event with the first target object.

In the above step 74, node vector characterization and interactive object prediction are performed using the first generative network. As mentioned above, the trained first generation network may characterize any node in the dynamic interaction graph as a node vector, and thus, for the first target node, the first generation network may output a corresponding first target vector.

For interactive object prediction, in one embodiment, a plurality of candidate nodes corresponding to the first target node may be determined. For example, when the first target object is an object of a first class, a plurality of candidate nodes belonging to an object of a second class may be determined from the dynamic interaction graph. Then, an alternative generating vector corresponding to each alternative node is generated by using the first generating network. And (3) according to the first target vector and each candidate generation vector, calculating the interaction probability of each candidate node and the first target node by using a formula (1) or (3), and selecting the node with the maximum probability as the second target node.

Then, in addition to outputting the first target vector corresponding to the first target node, the first generation network may predict a second target node at which an interaction event (hereinafter, referred to as a predicted interaction event) will occur with the first target object and output a corresponding second target vector.

Then, in step 75, the first target vector, the second target vector and the noise vector are input into a second generation network, and the predicted occurrence time of the predicted interaction event is obtained. This step is similar to the process of step 24 and will not be described again.

In the process, the first generation network is used for representing the node vector, and an interaction object of an interaction event with the first target node is predicted; the second generation network is then used to predict the time of occurrence of the interaction event. Thus, for a given target object, the interaction prediction model may give a comprehensive predictive analysis from interaction object to interaction time.

According to another embodiment, the interaction prediction model can also be used for predicting the future interaction time of the assumed two object nodes.

FIG. 8 shows a flow diagram of a method of predicting an interaction event, according to another embodiment. The method may be performed by any apparatus, device, platform, cluster of devices having computing, processing capabilities. As shown in fig. 8, first, in step 81, an interactive prediction model trained according to the above method is obtained, where the interactive prediction model includes a first generation network and a second generation network.

In step 82, a dynamic interaction graph is obtained, where the dynamic interaction graph includes a plurality of node pairs, each node pair corresponds to an interaction event, two nodes of the dynamic interaction graph respectively represent two objects participating in the interaction event, and any node points to two nodes corresponding to the last interaction event in which the object represented by the node participates through a connecting edge. This step is similar to step 21 of fig. 2 and will not be described again.

In step 83, a first target node and a second target node to be analyzed are obtained from the dynamic interaction graph. Here, the first target node and the second target node are two designated nodes, and it is assumed that the objects represented by the two nodes will interact in the future.

Next, at step 84, a first target vector corresponding to the first target node and a second target vector corresponding to the second target node are obtained via the first generation network. In this step, only the first generative network is needed for node vector characterization. The characterization method and process are as described above and will not be repeated.

Then, in step 85, inputting the first target vector, the second target vector and the noise vector into a second generation network to obtain predicted occurrence time; the predicted occurrence time represents a time at which the objects represented by the first target node and the second target node, respectively, are expected to interact.

In the above process, for two designated target nodes, the node vector is characterized by using the first generation network, and then the future interaction occurrence time is predicted by using the second generation network. Thus, for a given two target objects, the interaction prediction model can give accurate prediction analysis of their interaction times.

Referring back to the above process, in the solution of the embodiment of the present specification, a dynamic interaction graph is constructed based on a sequence of interaction events, and based on such dynamic interaction graph, a two-stage cascade of confrontation generating networks GAN is trained. The primary GAN performs countermeasure learning on probability distribution among objects which are interacted to obtain vector representation of nodes in the interaction graph; the second-level GAN is based on the vector representation of the first-level GAN, and the interaction time between the interaction objects is learned and predicted through the countermeasure training, so that the interaction objects and the interaction time can be comprehensively predicted through the two-level cascade GAN. The prediction of the interactive objects can be used for various scenes such as article recommendation, content recommendation and the like.

According to another aspect, an apparatus for training an interactive prediction model is provided, which may be deployed in any device, platform or device cluster having computing and processing capabilities. FIG. 9 shows a schematic block diagram of a training apparatus for an interactive predictive model according to one embodiment. As shown in fig. 9, the training device 90 includes:

the interaction graph obtaining unit 91 is configured to obtain a dynamic interaction graph, where the dynamic interaction graph includes a plurality of node pairs, each node pair corresponds to an interaction event, two nodes of the interaction graph respectively represent two objects participating in the interaction event, and an arbitrary node points to two nodes corresponding to a previous interaction event where an object represented by the node participates through a connecting edge;

a node pair selecting unit 92 configured to select a sample node pair from the dynamic interaction graph, where the sample node pair includes a first node and a second node and is marked with a real occurrence time of a sample interaction event corresponding to the sample node pair;

a vector acquisition unit 93 configured to acquire, from the first generated network, a first generated vector corresponding to the first node, and a second generated vector corresponding to the second node; obtaining, from a first discrimination network, a first discrimination vector corresponding to a first node and a second discrimination vector corresponding to a second node; the first generating network and the first discriminating network are trained by a first countermeasure training process;

a time prediction unit 94 configured to input the first generated vector, the second generated vector and the noise vector into a second generation network, resulting in a predicted occurrence time for the sample interaction event;

a first input discrimination unit 95 configured to form a first input based on the first generated vector, a second generated vector and the predicted occurrence time, and obtain a first probability that the first input corresponds to a real event through a second discrimination network;

a second input discrimination unit 96 configured to form a second input based on the first discrimination vector, a second discrimination vector and the real occurrence time, and obtain a second probability that the second input corresponds to a real event through the second discrimination network;

a second antagonistic training unit 97 configured to perform a second antagonistic training process including training the second decision network with a goal of increasing the second probability and decreasing the first probability; training the second generation network with a goal of increasing the first probability; the trained first generation network and the second generation network form the interactive prediction model.

According to an embodiment, the vector obtaining unit 93 is specifically configured to: determining a first sub-graph formed by nodes in a preset range starting from the root node and reaching through a connecting edge in the dynamic interaction graph by taking the first node as the root node; and inputting the first sub-graph into the first generation network, wherein the first generation network outputs a hidden vector of a root node as the first generation vector according to the node attribute characteristics of each node in the first sub-graph and the connection relation between the nodes.

According to one embodiment, the temporal prediction unit 94 is specifically configured to concatenate the first generated vector and the second generated vector as an a priori condition vector; sampling from Gaussian distribution to obtain the noise vector; and after splicing the prior condition vector and the noise vector, inputting the prior condition vector and the noise vector into the second generation network.

In one embodiment, the first input discrimination unit 95 is specifically configured to: performing fusion operation on the first generated vector and the second generated vector to obtain a first fusion vector; and splicing and combining the first fusion vector and the predicted occurrence time, and taking a combined result as the first input.

According to one embodiment, the first antagonistic training process comprises:

According to an embodiment of yet another aspect, an apparatus for predicting an interaction event is provided, which may be deployed in any device, platform or cluster of devices having computing and processing capabilities. Fig. 10 shows a schematic block diagram of an apparatus for predicting an interaction event according to an embodiment. As shown in fig. 10, the prediction apparatus 100 includes:

a model obtaining unit 101 configured to obtain a trained interactive prediction model, which includes a first generation network and a second generation network;

an interaction graph obtaining unit 102, configured to obtain a dynamic interaction graph, where the dynamic interaction graph includes a plurality of node pairs, each node pair corresponds to an interaction event, two nodes of the dynamic interaction graph respectively represent two objects participating in the interaction event, and an arbitrary node points to two nodes corresponding to a previous interaction event where an object represented by the node participates through a connecting edge;

a target node obtaining unit 103, configured to obtain a first target object to be analyzed, and construct a first target node corresponding to the first target object in the dynamic interaction graph;

a target vector acquisition unit 104 configured to acquire, through a first generation network, a first target vector corresponding to the first target node, and a second target vector corresponding to a second target node; wherein the second target node is a node corresponding to an object predicted by the first generation network to have a predicted interaction event with the first target object;

a time prediction unit 105 configured to input the first target vector, the second target vector and the noise vector into a second generation network, resulting in a predicted occurrence time for the predicted interaction event.

According to an embodiment of a further aspect, an apparatus for predicting an interaction event is provided, which may be deployed in any device, platform or cluster of devices having computing and processing capabilities. Fig. 11 shows a schematic block diagram of an apparatus for predicting an interaction event according to an embodiment. As shown in fig. 11, the prediction apparatus 110 includes:

a model obtaining unit 111 configured to obtain an interactive prediction model trained by the apparatus according to the fourth aspect, wherein the interactive prediction model includes a first generation network and a second generation network; (ii) a

An interaction graph obtaining unit 112 configured to obtain a dynamic interaction graph, where the dynamic interaction graph includes a plurality of node pairs, each node pair corresponds to an interaction event, two nodes of the dynamic interaction graph respectively represent two objects participating in the interaction event, and an arbitrary node points to two nodes corresponding to a previous interaction event where the object represented by the node participates in through a connecting edge;

a target node obtaining unit 113 configured to obtain a first target node and a second target node to be analyzed from the dynamic interaction graph;

a target vector acquisition unit 114 configured to acquire, through a first generation network, a first target vector corresponding to the first target node, and a second target vector corresponding to a second target node;

a time prediction unit 115, configured to input the first target vector, the second target vector and the noise vector into a second generation network, to obtain a predicted occurrence time, where the predicted occurrence time represents a time when an interaction is expected to occur for objects represented by the first target node and the second target node, respectively.

Through the training device 90, based on the dynamic interaction diagram, an interaction prediction model based on the two-stage cascade GAN is obtained through training; through the

above prediction apparatuses

100 and 110, the interaction object and the interaction time of the target object can be predicted and evaluated by using the trained interaction prediction model.

According to an embodiment of another aspect, there is also provided a computer-readable storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method described in connection with fig. 2.

According to an embodiment of yet another aspect, there is also provided a computing device comprising a memory and a processor, the memory having stored therein executable code, the processor, when executing the executable code, implementing the method described in connection with fig. 2.

Those skilled in the art will recognize that, in one or more of the examples described above, the functions described in this invention may be implemented in hardware, software, firmware, or any combination thereof. When implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium.

The above-mentioned embodiments, objects, technical solutions and advantages of the present invention are further described in detail, it should be understood that the above-mentioned embodiments are only exemplary embodiments of the present invention, and are not intended to limit the scope of the present invention, and any modifications, equivalent substitutions, improvements and the like made on the basis of the technical solutions of the present invention should be included in the scope of the present invention.

Claims

1. A method of training an interactive predictive model, the method comprising:

executing a second countermeasure training process, wherein the second judgment network is trained with the goal of increasing the second probability and decreasing the first probability; training the second generation network with a goal of increasing the first probability; the trained first generation network and the second generation network form the interactive prediction model.

2. The method of claim 1, wherein the first generating network and the first discriminating network are two neural networks with the same structure and independent parameters.

3. The method of claim 1, wherein obtaining, from a first generated network, a first generated vector corresponding to a first node comprises:

determining a first sub-graph formed by nodes in a preset range starting from the root node and reaching through a connecting edge in the dynamic interaction graph by taking the first node as the root node;

and inputting the first sub-graph into the first generation network, wherein the first generation network outputs a hidden vector of a root node as the first generation vector according to the node attribute characteristics of each node in the first sub-graph and the connection relation between the nodes.

4. The method of claim 3, wherein the first generation network comprises an LSTM layer, the LSTM layer respectively takes each node from a leaf node to a root node in the input subgraph as a current node, and sequentially and iteratively processes each node, and the iterative process comprises determining a hidden vector of the current node at least according to the node attribute characteristics of the current node and the hidden vectors of two nodes pointed by the current node through a connecting edge.

5. The method of claim 1, wherein inputting the first generated vector, the second generated vector, and the noise vector into a second generating network comprises:

splicing the first generated vector and the second generated vector to be used as prior condition vectors;

sampling from Gaussian distribution to obtain the noise vector;

and after splicing the prior condition vector and the noise vector, inputting the prior condition vector and the noise vector into the second generation network.

6. The method of claim 1, wherein forming a first input based on the first generated vector, a second generated vector, and the predicted time of occurrence comprises:

performing fusion operation on the first generated vector and the second generated vector to obtain a first fusion vector;

and splicing and combining the first fusion vector and the predicted occurrence time, and taking a combined result as the first input.

7. The method of any of claims 1-6, wherein the first antagonistic training process comprises:

8. The method of claim 7, wherein two nodes in each node pair in the dynamic interaction graph represent a first class object and a second class object, respectively, that participate in the interaction event;

the first node belongs to a first class of objects, and the plurality of candidate nodes all belong to a second class of objects.

9. The method of claim 7, wherein determining a probability distribution of each candidate node interacting with the first node comprises: and fitting to obtain the probability distribution by using a Gumbel-softmax function.

10. The method of claim 7, wherein the first antagonistic training process is completed before the second antagonistic training process.

11. The method of claim 7, wherein the first antagonistic training process alternates with the second antagonistic training process.

12. A method of predicting an interaction event, the method comprising:

obtaining an interactive prediction model trained according to the method of claim 1, including a first generative network, and a second generative network;

13. The method of claim 12, wherein obtaining, via a first generating network, a first target vector corresponding to the first target node and a second target vector corresponding to a second target node comprises:

determining a plurality of alternative nodes corresponding to the first target node;

generating alternative generating vectors corresponding to the alternative nodes by using the first generating network;

and calculating the interaction probability of each candidate node and the first target node according to the first target vector and each candidate generation vector, and selecting the node with the maximum probability as the second target node.

14. A method of predicting an interaction event, the method comprising:

15. An apparatus to train an interactive predictive model, the apparatus comprising:

16. An apparatus to predict an interaction event, the apparatus comprising:

a model obtaining unit configured to obtain an interactive prediction model trained by the apparatus according to claim 15, wherein the interactive prediction model comprises a first generation network and a second generation network; (ii) a

17. An apparatus to predict an interaction event, the apparatus comprising:

18. A computer-readable storage medium, on which a computer program is stored which, when executed in a computer, causes the computer to carry out the method of any one of claims 1-14.

19. A computing device comprising a memory and a processor, wherein the memory has stored therein executable code that, when executed by the processor, performs the method of any of claims 1-14.