WO2021139524A1

WO2021139524A1 - Method and apparatus for processing interaction data by using lstm neural network model

Info

Publication number: WO2021139524A1
Application number: PCT/CN2020/138398
Authority: WO
Inventors: 常晓夫; 文剑烽; 刘旭钦; 宋乐
Original assignee: 支付宝(杭州)信息技术有限公司
Priority date: 2020-01-09
Filing date: 2020-12-22
Publication date: 2021-07-15
Also published as: CN115081589A; CN111210008A; CN111210008B

Abstract

A method and an apparatus for processing interaction data. The method comprises: first acquiring a dynamic interaction diagram (200) constructed according to an interaction event set, any node i in the diagram pointing to, by means of connection edges, M associated nodes corresponding to N associated events in which an object represented by the node i last participated (31), the object being allowed to participate in a plurality of associated events at the same time, and the node being allowed to be connected to more than two associated nodes; then, in the dynamic interaction diagram (200), determining a current sub-diagram corresponding to the current node to be analyzed (32), and inputting the current sub-diagram into a neural network model for processing, the neural network model comprising an LSTM layer; and the LSTM layer successively and iteratively processing the nodes according to the pointing relationship of connecting edges between the nodes in the current sub-diagram, so as to obtain an implicit vector of the current node (33).

Description

Method and device for processing interactive data using LSTM neural network model

Technical field

One or more embodiments of this specification relate to the field of machine learning, and more particularly to methods and devices for processing interactive data using machine learning.

Background technique

In many scenarios, user interaction events need to be analyzed and processed. Interaction events are one of the basic elements of Internet events. For example, the click behavior of a user when browsing a page can be regarded as an interaction event between the user and the page content block, and the purchase behavior in e-commerce can be regarded as the relationship between the user and the product. Interaction events between users, and transfer behavior between accounts are interaction events between users. A series of user interaction events contain the user's fine-grained habits and preferences, as well as the characteristics of interactive objects, which are important source of features for machine learning models. Therefore, in many scenarios, it is desirable to characterize and model interactive participants based on interactive events.

However, an interactive event involves both parties to the interaction, and the status of each participant itself can be dynamically changed. Therefore, it is very difficult to accurately express the characteristics of the interactive participants comprehensively considering the various characteristics of the interactive parties. Therefore, it is hoped that there will be an improved solution to more effectively analyze and process the interactive objects in the interactive events, so as to obtain feature vectors suitable for subsequent business analysis.

Summary of the invention

One or more embodiments of this specification describe methods and devices for processing interactive data, in which an LSTM neural network model is used to consider interactive events in which interactive objects participate and the influence of other objects in the interactive events, and process interactive objects as implicit features. So as to carry out subsequent business processing analysis.

According to a first aspect, there is provided a method for processing interaction data, the method comprising: obtaining a dynamic interaction graph constructed according to an interaction event set, wherein the interaction event set includes a plurality of interaction events, and each interaction event includes at least , The two objects at which the interaction behavior occurs and the interaction time; the dynamic interaction graph includes any first node, and the first node corresponds to the first object in the interaction event that occurred at the first time, and the first node Pointing to the M associated nodes corresponding to the N associated events through the connecting edge, the N associated events all occur at the second time, and all include the first object as one of the interactive objects, and the second time is, Looking back from the first time forward, the time before the interaction behavior of the first object occurred; the dynamic interaction graph includes at least one multi-element node whose number of associated nodes is greater than 2; in the dynamic interaction graph, it is determined A current subgraph corresponding to the current node to be analyzed, where the current subgraph includes nodes within a predetermined range that start from the current node and reach via connecting edges; the current subgraph is input into a neural network model, and the neural network model Including the LSTM layer, the LSTM layer sequentially iteratively process each node according to the direction relationship of the connecting edge between each node in the current subgraph, so as to obtain the implicit vector of the current node; wherein each node includes the first Two nodes, the sequential iterative processing of each node includes, at least according to the node characteristics of the second node, the respective intermediate vectors and implicit vectors of the k associated nodes pointed to by the second node, determining the value of the second node Hidden vector and intermediate vector; according to the hidden vector of the current node, perform business processing related to the current node.

According to an embodiment, the aforementioned object includes a user, and the interaction event includes at least one of the following: a click event, a social event, and a transaction event.

In one embodiment, the above-mentioned M associated nodes are 2N nodes, respectively corresponding to the two objects included in each associated event in the N associated events; or, in another embodiment, the M associated nodes are The N+1 nodes respectively correspond to the N other objects interacting with the first object in the N associated events, and the first object itself.

In different implementation manners, the nodes within the foregoing predetermined range may include nodes within a connecting edge of a preset order K; and/or nodes whose interaction time is within a preset time range.

In an embodiment, each interaction event further includes the behavior characteristics of the interaction behavior; in this case, the node characteristics of the second node may include the attribute characteristics of the object corresponding to the second node, and the The behavior characteristics of the interaction event that the second node participates in the corresponding interaction time.

In an embodiment, the implicit vector and the intermediate vector of the second node may be determined by the following method: combining the node characteristics of the second node with the k implicit vectors corresponding to the k associated nodes, respectively, Input the first transformation function and the second transformation function with the same algorithm and different parameters to obtain k first transformation vectors and k second transformation vectors respectively; and the intermediate vector of the i-th associated node among the k associated nodes, and The corresponding i-th first transform vector and the i-th second transform vector are combined to obtain k operation results, and the k operation results are summed to obtain the combined vector; the node characteristics of the second node are combined with The k implicit vectors are input into the third transformation function and the fourth transformation function respectively to obtain the third transformation vector and the fourth transformation vector; based on the combination vector and the third transformation vector, determine the value of the second node Intermediate vector; based on the intermediate vector of the second node and the fourth transform vector, the implicit vector of the second node is determined.

According to an embodiment, sequentially iteratively processing each node may include, according to the node characteristics of the second node, the respective intermediate vectors and implicit vectors of the k associated nodes pointed to by the second node, and the corresponding second node The time difference between the interaction time and the interaction time corresponding to the k associated nodes determines the implicit vector and the intermediate vector of the second node.

In an example of the foregoing implementation manner, the implicit vector and the intermediate vector of the second node can be determined in the following manner: the node characteristics of the second node and the time difference are determined by k corresponding to the k associated nodes. The implicit vectors are combined separately, and the first transformation function is input to obtain k first transformation vectors; the node characteristics of the second node and the k implicit vectors corresponding to the k associated nodes are respectively combined, and the second transformation is input Function to obtain k second transform vectors; combine the intermediate vector of the i-th associated node among the k associated nodes with the corresponding i-th first transform vector and the i-th second transform vector to obtain k operations As a result, the k operation results are summed to obtain a combined vector; the node feature of the second node and the k implicit vectors are input to the third transformation function and the fourth transformation function, respectively, to obtain the third transformation Vector and a fourth transformation vector; based on the combined vector and the third transformation vector, determine the intermediate vector of the second node; based on the intermediate vector and the fourth transformation vector of the second node, determine the intermediate vector of the second node Implied vector.

In another embodiment of the foregoing implementation manner, the implicit vector and the intermediate vector of the second node can be determined in the following manner: the node characteristics of the second node and the time difference are determined by k corresponding to the k associated nodes. After the two implicit vectors are combined separately, the first transformation function and the second transformation function with the same algorithm and different parameters are input, and k first transformation vectors and k second transformation vectors are obtained respectively; The intermediate vector of the i-associated node is combined with the corresponding i-th first transform vector and the i-th second transform vector to obtain k operation results, and the k operation results are summed to obtain the combined vector; The node features of the two nodes and the k hidden vectors are input into the third transformation function and the fourth transformation function, respectively, to obtain the third transformation vector and the fourth transformation vector; based on the combination vector and the third transformation vector, determine The intermediate vector of the second node; based on the intermediate vector of the second node and the fourth transform vector, the implicit vector of the second node is determined.

According to an embodiment, the neural network model may include multiple LSTM layers, wherein the implicit vector of the second node determined by the previous LSTM layer is input to the next LSTM layer as the node feature of the second node.

In an example of the foregoing implementation manner, the neural network model integrates the hidden vectors of the current node output by each of the multiple LSTM layers to obtain the final hidden vector of the current node.

In another example of the foregoing implementation manner, the neural network model uses the implicit vector of the current node output by the last LSTM layer among the plurality of LSTM layers as the final implicit vector of the current node.

According to one embodiment, the neural network model is trained in the following manner: acquiring historical interaction events, including a first sample object and a second sample object; in the dynamic interaction graph, determining the relationship with the first sample The first subgraph corresponding to the object, and the second subgraph corresponding to the second sample object; the first subgraph and the second subgraph are respectively input to the neural network model to obtain the first subgraph respectively The hidden vector of the sample object and the hidden vector of the second sample object; according to the hidden vector of the first sample object and the hidden vector of the second sample object, predict the first sample object and the second sample object. Whether the two sample objects will interact to obtain a prediction result; determine the prediction loss according to the prediction result; and update the neural network model according to the prediction loss.

According to another embodiment, the neural network model is trained in the following manner: a sample object is selected from a plurality of sample objects involved in the interaction event set, and the classification label of the sample object is obtained; in the dynamic interaction diagram, Determine the sample subgraph corresponding to the sample object; input the sample subgraph into the neural network model to obtain the hidden vector of the sample object; predict the hidden vector of the sample object according to the hidden vector of the sample object Classify to obtain a prediction result; determine a prediction loss according to the prediction result and the classification label; and update the neural network model according to the prediction loss.

According to a second aspect, there is provided an apparatus for processing interaction data, the apparatus comprising: an interaction graph obtaining unit configured to obtain a dynamic interaction graph constructed according to an interaction event set, wherein the interaction event set includes a plurality of interaction events, Each interaction event includes at least two objects where the interaction behavior occurred and the interaction time; the dynamic interaction graph includes any first node, and the first node corresponds to the first object in the interaction event that occurred at the first time , The first node points to the M associated nodes corresponding to the N associated events through the connecting edge, and the N associated events all occur at the second time, and all include the first object as one of the interactive objects, so The second time is, going back from the first time, the time before the interaction of the first object occurs; the dynamic interaction graph includes at least one multi-node with more than 2 associated nodes; the subgraph is determined A unit configured to determine, in the dynamic interaction graph, a current subgraph corresponding to the current node to be analyzed, the current subgraph including nodes within a predetermined range starting from the current node and arriving via a connecting edge; subgraph processing Unit, configured to input the current subgraph into a neural network model, the neural network model including an LSTM layer, and the LSTM layer sequentially iteratively processes each node according to the direction relationship of the connecting edges between the nodes in the current subgraph Node, thereby obtaining the implicit vector of the current node; wherein each node includes a second node, and the sequential iterative processing of each node includes, at least according to the node characteristics of the second node, the second node points to The respective intermediate vectors and implicit vectors of the k associated nodes determine the implicit vector and the intermediate vector of the second node; the service processing unit is configured to perform correlation with the current node according to the implicit vector of the current node Business processing.

According to a third aspect, there is provided a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method of the first aspect.

According to a fourth aspect, there is provided a computing device, including a memory and a processor, characterized in that executable code is stored in the memory, and when the processor executes the executable code, the method of the first aspect is implemented .

According to the method and device provided in the embodiments of this specification, a dynamic interaction diagram is constructed based on a set of interaction events, and the dynamic interaction diagram reflects the timing relationship of each interaction event and the mutual influences between interactive objects transmitted through each interaction event. Considering the possibility of simultaneous occurrence of interaction events in practical operations, the dynamic interaction graph allows nodes to be connected to an unlimited number of associated nodes, thereby forming a mixed and diverse interaction graph. Using the trained LSTM neural network model, based on the subgraphs related to the interactive object to be analyzed in the dynamic interaction graph, the hidden vector of the interactive object can be extracted. The implicit vector thus obtained introduces the influence of other interactive objects in each interactive event on it, so that the in-depth characteristics of the interactive object can be expressed comprehensively and used for business processing.

Description of the drawings

In order to explain the technical solutions of the embodiments of the present invention more clearly, the following will briefly introduce the drawings used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. A person of ordinary skill in the art can obtain other drawings based on these drawings without creative work.

Figure 1A shows a two-part diagram of the interaction relationship in an example;

FIG. 1B shows an interactive relationship network diagram in another example;

Fig. 2 shows a schematic diagram of an implementation scenario according to an embodiment;

Fig. 3 shows a flowchart of a method for processing interactive data according to an embodiment;

Figure 4 shows a dynamic interaction diagram constructed according to an embodiment;

Figure 5 shows a dynamic interaction diagram constructed according to another embodiment;

Figure 6 shows an example of the current subgraph in one embodiment;

FIG. 7 shows an example of the current subgraph in another embodiment;

Figure 8 shows a schematic diagram of the work of the LSTM layer;

Figure 9 shows the structure of an LSTM layer according to an embodiment;

FIG. 10 shows the structure of an LSTM layer according to another embodiment;

FIG. 11 shows the structure of an LSTM layer according to another embodiment;

Figure 12 shows a flowchart of training a neural network model in one embodiment;

Figure 13 shows a flowchart of training a neural network model in another embodiment;

Fig. 14 shows a schematic block diagram of an apparatus for processing interactive data according to an embodiment.

Detailed ways

The following describes the solutions provided in this specification with reference to the accompanying drawings.

As mentioned above, it is hoped that based on the interaction event, the participants in the interaction event, that is, the interaction object, can be characterized and modeled.

In one solution, a static interaction relationship network diagram is constructed based on historical interaction events, so that each interaction object is analyzed based on the interaction relationship network diagram. Specifically, participants in each historical event can be used as nodes, and connecting edges can be established between nodes that have an interactive relationship, so as to form the foregoing interactive network graph.

Fig. 1A and Fig. 1B respectively show the interactive relationship network diagram in a specific example. More specifically, Figure 1A shows a bipartite graph, which contains user nodes (U1-U4) and product nodes (V1-V3). If a user has purchased a product, it will be between the user and the product. Construct a connecting edge between. Figure 1B shows a user transfer relationship diagram, in which each node represents a user, and there is a connecting edge between two users who have made transfer records.

However, it can be seen that although FIG. 1A and FIG. 1B show the interaction relationship between objects, they do not contain the timing information of these interaction events. Simply embedding the graph based on such an interactive relationship network graph, the obtained feature vector does not express the influence of the time information of the interactive event on the node. Moreover, the scalability of such a static graph is not strong enough, and it is difficult to flexibly handle the situation of newly-added interactive events and newly-added nodes.

In another solution, for each interactive object to be analyzed, a behavior sequence of the object is constructed, and the characteristic expression of the object is extracted based on the behavior sequence. However, such a behavior sequence only characterizes the behavior of the object to be analyzed, and interactive events are events involving multiple parties, and the participants will indirectly transmit influence through the interaction events. Therefore, this method does not express the influence between the participating objects in the interaction event.

Taking the above factors into consideration, according to one or more embodiments of this specification, a dynamically changing set of interaction events is constructed into a dynamic interaction graph, wherein each interaction object involved in each interaction event corresponds to each node in the dynamic interaction graph. An arbitrary node is connected to a number of associated nodes, where the associated node is the node corresponding to the interaction event that the object corresponding to the arbitrary node participated in last time. For the interactive object to be analyzed, the subgraph part related to the corresponding node is obtained from the dynamic interaction graph, and the subgraph part is input into the LSTM-based neural network model to obtain the feature vector expression of the interactive object.

Fig. 2 shows a schematic diagram of an implementation scenario according to an embodiment. As shown in Figure 2, the set of interaction events composed of multiple interaction events can be obtained. More particularly, the interaction may be a set of events more interaction events chronologically organized into a sequence of interactive events _{_{<E 1, E 2, ...}} , E N>, where each element represents an interactive event E _i, may be expressed is a form of interactive features set _{_{E i = (a i, b}} i, t i), where a _i and b _i are the two objects interact in the event E _i, t _i is the time of interaction. Due to factors such as the accuracy of time measurement, multiple interaction events are allowed to occur at the same time.

According to an embodiment of the present specification, a dynamic interaction diagram 200 is constructed based on the set of interaction events. In FIG. 200, the individual interactive objects each interaction event a _{_i,} b _i represented by nodes, and establishes a connection between the edges containing the same event object. Since multiple interaction events are allowed to occur at the same time, the dynamic interaction graph 200 contains at least one multi-element node, and the multi-element node can be connected to 3 or more associated nodes. The structure of the dynamic interaction graph 200 will be described in more detail later.

For a certain interactive object to be analyzed, the corresponding current node in the dynamic interaction graph can be determined, and the current subgraph related to the current node in the dynamic interaction graph can be obtained. Generally, the current subgraph includes a certain range of nodes that can be reached from the current node through the connecting edge. The current subgraph reflects the influence of other objects in the interaction event directly or indirectly associated with the current interaction object on the current node.

Then, the current sub-image is input to the neural network model based on long and short-term memory LSTM, and the feature vector of the current interactive object is output through the model. The feature vector obtained in this way can extract the time information of the associated interaction event and the influence between the interaction objects in each interaction event, so as to more accurately express the in-depth characteristics of the current interaction object. Such feature vectors can be subsequently applied to various machine learning models and various business processing scenarios. For example, reinforcement learning can be performed based on the feature vector thus obtained, or cluster analysis can be performed based on the feature vector, for example, clustering users into groups of people. You can also perform classification prediction based on such feature vectors, for example, predict whether there will be interaction between two objects (such as whether a user will buy a certain product), and predict the business type of a certain object (such as the risk of a certain user). Level), etc.

The specific implementation of the above concept is described below.

Fig. 3 shows a flowchart of a method for processing interactive data according to an embodiment. It is understandable and understandable that the method can be executed by any device, device, platform, or device cluster with computing and processing capabilities. The following describes each step in the method for processing interactive data as shown in FIG. 3 in conjunction with specific embodiments.

First, in step 31, a dynamic interaction graph constructed according to the set of interaction events is obtained.

As mentioned above, it is possible to obtain an interaction event set composed of multiple interaction events, where each interaction event has two interaction objects and interaction time. Thus, any interaction events E _i may be expressed as a set of interactive features _{_{E i = (a i, b}} i, t i), where a _i and b _i are two events E _i interactive object, such as a first target With the second object, t _i is the interaction time.

For example, in an e-commerce platform, the interaction event may be a user's purchase behavior, where the first object may be a certain user, and the second object may be a certain commodity. In another example, the interaction event may be a user's click behavior on a page block, where the first object may be a certain user, and the second object may be a certain page block. In another example, the interaction event may be a transaction event, for example, a user transfers money to another user, or a user makes a payment to a store or platform. In another example, the interaction event may be a social event that a user occurs through a social platform, such as chatting, calling, sending red envelopes, and so on. In other business scenarios, interaction events can also be other interaction behaviors that occur between two objects.

In one embodiment, according to the characteristics of the interaction event, the two objects interacting may be objects of different types, such as objects of the first type and objects of the second type. For example, when the interaction event is a purchase behavior in an e-commerce platform, the first type of object may be a certain user, and the second type of object may be a certain commodity. In other embodiments, the two objects involved in the interaction event may be objects of the same type. For example, in an instant messaging scenario, the interaction event may be an instant communication between two users. At this time, both the first object and the second object are users and belong to the same type of objects. In still other embodiments, whether to distinguish the types of two interactive objects can be set according to the needs of the business. For example, for the transfer interaction event, in the foregoing example, it can be considered that the two users belong to the same type of object. In other examples, according to business needs, the user who transfers the amount may be regarded as an object of the first type, and the user who is the recipient is regarded as an object of the second type.

Further, in an embodiment, the interaction feature group corresponding to each interaction event may also include event feature or behavior feature f. In this way, each interaction feature group can be expressed as X _i = (a _i , b _i , t _i ,f). Specifically, the event feature or behavior feature f may include background and context information of the occurrence of the interaction event, some attribute features of the interaction behavior, and so on.

For example, in the case where the interaction event is a user click event, the event feature f may include the type of terminal used by the user to click, browser type, app version, etc.; in the case where the interaction event is a transaction event, the event The feature f may include, for example, transaction type (commodity purchase transaction, transfer transaction, etc.), transaction amount, transaction channel, and so on.

Interactivity events described above as well as in the event interactive objects a _i, b _i of example.

For the interaction time t _i , it needs to be understood that in actual operation, the time is always measured and recorded in an appropriate duration unit. For example, in some service platforms, the unit duration for recording the interaction time can be hours h or minutes m. Therefore, multiple interaction events are likely to occur within the unit duration. Even if it takes a short duration as a unit, such as seconds or even milliseconds, for some service platforms that interact very frequently, such as Alipay, there will inevitably be multiple interaction events within a unit of time.

In addition, there are some cases of batch interaction. For example, the user edits a message in advance, selects a group of friends, and then performs a batch group sending operation. This is equivalent to initiating an interaction event with multiple friends at the same time. For another example, a user adds multiple items to the shopping cart and then selects batch settlement, which is equivalent to simultaneously initiating interaction events with multiple items.

For at least the above two reasons, it is often the case that the interaction time recorded by multiple interaction events is the same. For such situations, this article is sometimes referred to as multiple interactive events occurring at the same time, without distinguishing their precise time and sequence.

In a specific example, it is assumed that an interaction event set S is obtained. The interaction events in the interaction event set S are arranged in chronological order and expressed in the form of interaction feature groups, which can be recorded as follows:

Among them, a, b, c, d, e, f, u, v are interaction objects, interaction events E ₂ and E ₃ all occur at _{time t 2} and interaction events E ₄ , E ₅ and E ₆ all occur at t ₃ At time, the interaction events E ₇ and E ₈ both occurred at time _{t 4.}

For the set of interaction events described above, a dynamic interaction diagram can be constructed to depict the association relationships between interaction events and interaction objects. Specifically, the objects contained in the interaction events occurring at each time can be used as the nodes of the dynamic interaction graph. In this way, a node can correspond to an object that interacts at a time, but the same entity object may correspond to multiple nodes. For example, if the entity object v interacts with the object u at time _{t 6} _{, the node v(t 6} ) can be constructed correspondingly, and the node v(t ₅ ) can be constructed correspondingly by interacting with the object c at t _{5 time.} Therefore, it can be considered that the nodes in the dynamic interaction graph correspond to the interactive object at a certain interaction time, or in other words, correspond to the state of the interactive object at a certain interaction time.

For each node in the dynamic interaction graph, construct the connecting edge in the following way: For any node i, it is called the first node for simplicity; assuming that it corresponds to the first object at the first interaction time t, then in the interaction In the event sequence, backtracking from the first interaction time t, that is, backtracking in the direction earlier than the first interaction time t, it is determined that the last time when the first object interacted is the second time (t-), and The N interaction events that occur in the second time and the first object participates in are regarded as the N associated events of the first node, and the M nodes corresponding to the N associated events are regarded as the associated nodes, and it is established to point from the first node i to M The connecting edge of the associated node. Due to the possibility of multiple interaction events occurring at the same time, N may be greater than 1. In this way, the dynamic interaction graph may include multiple nodes, that is, nodes with more than two connected associated nodes.

In one embodiment, when constructing the dynamic interaction graph, corresponding nodes are respectively established for the two objects of each interaction event. In this way, the aforementioned N associated events correspond to 2N nodes, and these 2N nodes are regarded as the aforementioned M associated nodes.

Fig. 4 shows a dynamic interaction diagram constructed according to an embodiment. Specifically, the left side of FIG. 4 shows a schematic diagram of an interaction sequence that organizes the foregoing interaction event set S in chronological order, and the right side shows a dynamic interaction diagram. In this dynamic interaction graph, two interaction objects in each interaction event are respectively regarded as nodes. The following takes nodes u(t ₆ ) and v(t ₆ ) as examples to describe the construction of connecting edges.

It can be understood that the node u(t ₆ ) represents the object u at _{time t 6.} Therefore, starting from time t ₆ and backtracking, it can be determined that the time when the object u had an interactive behavior last time is t ₄ , during which time t ₄ participated in two related events E ₇ and E ₈ , namely, the interaction event E ₇ Both E and E ₈ contain the object u as one of the interactive objects. Therefore, the four nodes corresponding to the associated events E ₇ and E ₈ are the associated nodes of _{the node u(t 6 ).} In Fig. 4, in order to distinguish the object node u in the _{events E 7} and E ₈ _{, they are denoted as u 1} (t ₄ ) and u ₂ (t ₄ ). Thus, a connecting edge from the node u(t ₆ ) to its 4 associated nodes is established.

The node v(t ₆ ) represents the object v at _{time t 6.} Therefore, starting from time t ₆ and looking back forward, it can be determined that the time when the object v had an interactive behavior last time is t ₅ , during which time t ₅ participated in an associated event E ₉ . Therefore, the two nodes v(t ₅ ) and c(t ₅ ) corresponding to the associated event E ₉ are the associated nodes of node v(t _{6 ).} Then, a connecting edge from the node v(t ₆ ) to the two associated nodes is established. For each other node, the above-mentioned method can be used to determine its associated event and associated node, so as to establish a connection edge to the associated node. In the dynamic interaction graph shown in Figure 4, the nodes u(t ₆ ) and c(t ₅ ) are all multi-element nodes.

In another embodiment, when constructing a dynamic interaction graph, for multiple interaction events that occur at the same time, different interaction objects involved in the multiple interaction events are determined, and corresponding nodes are established for each different interaction object. In other words, if multiple interaction events that occur at the same time contain the same object, only one node is established for the same object. In this way, when establishing a connection edge, if there are N associated events corresponding to the first node of the first object, then these N associated events correspond to N+1 associated nodes, corresponding to the first object itself, and N N other objects interacting with the first object in the associated event.

Fig. 5 shows a dynamic interaction diagram constructed according to another embodiment. Specifically, the left side of FIG. 5 shows the aforementioned interaction event set S, and the right side shows a dynamic interaction diagram. In the dynamic interaction graph, corresponding nodes are respectively established for different interaction objects in the interaction events that occur at the same time. The difference between the dynamic interaction diagram of FIG. 5 and that of FIG. 4 is that the nodes of the same object in the multiple interaction events that occur at the same time in FIG. 4 are merged into one node. For example, for two interaction events E ₇ , E ₈ that both occur at time t ₄ , which involve three different interaction objects a, b, u, then three nodes a(t ₄ ) are established for the interaction event at that time , B(t ₄ ), u(t ₄ ). This is equivalent to _{merging u 1} (t ₄ ) and u ₂ (t ₄ ) in Fig. 4 into one node u(t ₄ ). In this case, in one example, can be shown by the dashed double arrows between the nodes of the interactions occur, for example, shown by the dashed double arrow in FIG. 5, at time t _4, there is an interaction with a target u Behavior, there is interaction between objects b and u, but there is no interaction between objects a and b.

The following still takes the nodes u(t ₆ ) and v(t ₆ ) as examples to describe the construction of connecting edges.

As mentioned earlier, the node u(t ₆ ) represents the object u at _{time t 6.} Starting from time t ₆ and looking back, it can be determined that the time when the object u had an interactive behavior last time is t ₄ , during which time t ₄ participated in two related events E ₇ and E ₈ , namely, interactive events E ₇ and E ₈ all contain the object u as one of the interactive objects. The three nodes a(t ₄ ), b(t ₄ ), and u(t ₄ ) corresponding to the associated events E ₇ and E ₈ are associated nodes of the node u(t _{6 ).} Then, a connecting edge from the node u(t ₆ ) to the three associated nodes is established.

From the node v(t ₆ ), a connecting edge pointing to the two nodes v(t ₅ ) and c(t ₅ ) corresponding to the associated event E ₉ can be established. This process is the same as the description in conjunction with FIG. 4 and will not be repeated. For each of the other nodes in Figure 5, the above-mentioned methods can be used to determine the associated events and associated nodes, so as to establish a connection edge to the associated node. In the dynamic interaction graph shown in Figure 5, the nodes u(t ₆ ) and c(t ₅ ) are all multi-element nodes.

The above describes the method and process of constructing a dynamic interaction graph based on the set of interaction events. For the method of processing interactive data shown in FIG. 3, the process of constructing a dynamic interactive graph can be performed in advance or on site. Correspondingly, in one embodiment, in step 31, a dynamic interaction diagram is constructed on-site according to the interaction event set. The construction method is as described above. In another embodiment, a dynamic interaction graph may be constructed based on a set of interaction events in advance. In step 31, read or receive the formed dynamic interaction graph.

It can be understood that the dynamic interaction graph constructed in the above manner has strong scalability and can be dynamically updated according to newly added interaction events very easily. Correspondingly, step 31 may also include a process of updating the dynamic interaction graph.

Specifically, an existing dynamic interaction diagram constructed based on an existing interaction event set can be obtained, and then as time is updated, new interaction events that occur during the update time are continuously detected, and the existing dynamic interaction diagram is updated according to the new interaction events .

In one embodiment, the existing dynamic interaction graph adopts the form of FIG. 4, and each interaction event corresponds to two nodes. In this case, assuming that P new interaction events that occurred at the first update time are obtained, then 2P new nodes are added to the existing dynamic interaction graph, and the 2P new nodes respectively correspond to P Two objects included in each of the new interactive events. Then, for each newly added node, find its associated events and associated nodes in the aforementioned manner. If there is an associated node, add a connecting edge from the newly added node to its associated node.

In another embodiment, the existing dynamic interaction graph adopts the form of FIG. 5, and different objects in simultaneous interaction events correspond to different nodes. In this case, after acquiring the P newly-added interaction events that occurred at the first update time, first determine the Q different objects involved in the P newly-added interaction events. If the same interaction object does not exist in the P newly added interaction events, then Q=2P; if the same interaction object exists in the P newly added interaction events, then Q<2P. Then, Q new nodes are added to the existing dynamic interaction graph, and the Q new nodes respectively correspond to Q different objects. Then, for each newly added node, find its associated events and associated nodes in the aforementioned manner. If there is an associated node, add a connecting edge from the newly added node to its associated node.

In summary, in step 31, a dynamic interaction graph constructed based on the set of interaction events is obtained.

Next, in step 32, in the acquired dynamic interaction graph, a current subgraph corresponding to the current node to be analyzed is determined, and the current subgraph includes nodes within a predetermined range that start from the current node and reach through the connecting edge.

The current node is the node corresponding to the interactive object to be analyzed. However, as mentioned earlier, an entity object can correspond to multiple nodes, expressing the state of the entity object at different times. In order to express the latest state of the interaction object to be analyzed, in one embodiment, such a node is selected as the current node, that is, in the dynamic interaction graph, there is no connecting edge pointing to the node. In other words, select the node corresponding to the time when the object to be analyzed recently interacted as the current node. For example, in the dynamic interaction diagrams shown in Figs. 4 and 5, when you want to analyze the interactive object u, you can select the node u(t ₆ ) as the current node. However, this is not required. In other embodiments, for example, for training purposes, another node may be selected as the current node. For example, in order to analyze the object u, the node u(t ₄ ) may also be selected as the current node.

Starting from the current node, the nodes within the predetermined range reached via the connecting edge constitute the current subgraph corresponding to the current node. In an embodiment, the nodes within the foregoing predetermined range may be nodes that are reachable by connecting edges of at most a preset order K. Here, the order K is a preset hyperparameter, which can be selected according to business conditions. It can be understood that the preset order K reflects the number of steps of historical interaction events that are traced forward when the information of the current node is expressed. The larger the number K is, the more historical interactive information of order is considered.

In another embodiment, the nodes within the foregoing predetermined range may also be nodes whose interaction time is within the predetermined time range. For example, backtracking from the interaction time of the current node for a duration of T (for example, one day), the nodes within the range of duration that can be reached by connecting edges.

In another embodiment, the aforementioned predetermined range considers both the number of connected edges and the time range. In other words, the nodes within the predetermined range refer to nodes whose connecting edges passing through the preset order K at most are reachable and whose interaction time is within the predetermined time range.

For simplicity, in the following example, the connection edge of the preset order K is taken as an example for description.

Figure 6 shows an example of the current subgraph in one embodiment. In the example in Figure 6, suppose that u(t ₆ ) in Figure 4 is the current node, and the preset order K=2, then starting from u(t ₆ ), traverse along the direction of the connecting edge, and connect via 2 levels The nodes that the edges can reach are shown in the gray nodes in the figure. These gray nodes and the connection relationship between them are the current subgraph corresponding to _{the current node u(t 6 ).}

Fig. 7 shows an example of the current subgraph in another embodiment. In the example in Figure 7, suppose u(t ₆ ) in Figure 5 is the current node, and the preset order K=2, then starting from u(t ₆ ), traverse along the direction of the connecting edge, and connect via 2 levels The nodes that the edges can reach are shown in the gray nodes in the figure. These gray nodes and the connection relationship between them are the current subgraph corresponding to _{the current node u(t 6 ).}

Next, in step 33, the current sub-image is input to the neural network model, which includes the LSTM layer. For any node in the current subgraph, it is called the second node for the convenience of expression, and the LSTM layer performs the following processing: at least according to the node characteristics of the second node, the middle of each of the k associated nodes pointed to by the second node The vector and the hidden vector are used to determine the hidden vector and the intermediate vector of the second node. In this way, the LSTM layer iteratively processes each node in turn according to the direction relationship of the connecting edge between each node in the current subgraph, so as to obtain the implicit vector of the current node.

Figure 8 shows the working schematic diagram of the LSTM layer. Assume that node Q points to k associated nodes: node J ₁ to node J _k . 8, at time T, were treated to give LSTM layer node to node J ₁ J _k represents a vector H ₁ to H _k, and a vector comprising an intermediate vector implicit; the next time T +, according to the node layer LSTM The node characteristics of Q, the representation vectors H ₁ to H _k _{of J 1} to J _k obtained by the previous processing, and the representation vector H _{Q of} node Q is obtained. It can be understood that the representation vector of the node Q can be used to obtain the representation vector of the node pointing to the node Q at a subsequent time, so as to implement iterative processing.

This process is described in conjunction with the current sub-figure of FIG. 7. For each lowest-level node in the graph, for example, node a(t ₂ ), its pointing node is not considered in the current subgraph, that is, it is considered that a(t ₂ ) does not have associated nodes. In this case, the intermediate vector c and the implicit vector h of the associated node pointed to by the node are generated by padding with a default value (for example, 0). Thus node features, LSTM layer based on the node a (t _2), and the default node associated intermediate vector c and vector H implied, determination node a (t ₂₎ of the intermediate vector c (a (t ₂₎₎ And the implicit vector h(a(t ₂ )). Do the same process for the other lowest-level nodes to get the corresponding intermediate vector and hidden vector.

For the intermediate node a(t ₄ ), it points to two associated nodes a(t ₂ ) and f(t ₂ ). Therefore, the LSTM layer is based on the node _{characteristics of the node a(t 4} ) itself, and the respective intermediate and hidden vectors of the two associated nodes a(t ₂ ) and f(t _{2) it points to, namely c(a(} t ₂ )), h(a(t ₂ )), c(f(t ₂ )) and h(f(t ₂ )), determine the intermediate vector c(a(t ₄ )) of _{node a(t 4)} And h(a(t ₄ )). Do the same process for other middle layer nodes to get their corresponding middle vector and hidden vector.

For node u(t ₆ ), it points to three associated nodes a(t ₄ ), u(t ₄ ) and b(t ₄ ). Therefore, the LSTM layer is based on the node _{characteristics of the node u(t 6} ) itself, and the respective intermediate vectors and implicit vectors of the three associated nodes a(t ₄ ), u(t ₄ ) and b(t _{4) it points to} , Namely c(a(t ₄ )), h(a(t ₄ )), c(u(t ₄ )), h(u(t ₄ )), c(b(t ₄ )) and h(b (t ₄ )), determine the intermediate vectors c(u(t ₆ )) and h(u(t ₆ )) of the _{node u(t 6 ).}

In this way, iterative processing can obtain the intermediate vector and implicit vector of the _{current node u(t 6 ).}

The following describes the internal structure and algorithm of the LSTM layer in order to achieve the above iterative processing.

Figure 9 shows the structure of the LSTM layer according to one embodiment. In the example in Figure 9, the currently processed node is denoted as u(t),

Indicates the node characteristics of the node u(t). In the case where the node represents the user, the node feature may include the user's attribute characteristics, such as age, occupation, education level, location, etc.; in the case of the node representing the commodity, the node feature may include the attribute feature of the commodity, such as a commodity Category, time on sale, sales volume, etc. In the case where the node represents other interactive objects, the original node characteristics can be obtained accordingly. In the case that the feature group of the interaction event also includes the event feature or the behavior feature f, the node feature may also include the behavior feature f of each interaction event participating in the interaction time corresponding to the node.

Assuming that the node u(t) points to k associated nodes, denoted as u ¹ (t), u ² (t),..., u ^k (t), each associated node has an intermediate vector c and a hidden vector h. Fig. 9 exemplarily shows an example of k=3, correspondingly,

Represents the intermediate vector of the associated node,

Represents the implicit vector of the associated node. In Figure 9, i is 1, 2, and 3 respectively. However, it can be understood that the calculation relationship shown in FIG. 9 can be applied to the case where k is other values. For example, if the node u(t) does not have a real associated node, then k=0. At this time, a default value, such as a zero vector, can be used as the intermediate vector and implicit vector of the associated node; if the node u(t) is Binary node, then k=2, at this time, for example, the default value of the zero vector can be used as ^{the intermediate node and implicit vector corresponding to the third associated node u 3} (t); if the number of associated nodes of node u(t) If it is greater than 3, then add more intermediate vectors and hidden vectors corresponding to the associated nodes as input on the basis of FIG. 9.

The LSTM layer performs the following operations on the input node features, intermediate vectors and hidden vectors.

Node feature

K implicit vectors corresponding to k associated nodes respectively

(Where i is from 1 to k) combination, respectively input the first transformation function g and the second transformation function f with the same algorithm and different parameters to obtain k first transformation vectors and k second transformation vectors respectively.

More specifically, in an example, the first transformation function g and the second transformation function f are calculated using the following formula (1) and formula (2), respectively:

In the above formula (1) and formula (2), σ is the activation function, such as the sigmoid function,

with

Is the parameter matrix of the linear transformation,

with

Is the offset parameter. It can be seen that formulas (1) and (2) have the same algorithm, only the parameters are different. Through the above transformation function, k first transformation vectors and k second transformation vectors can be obtained.

Of course, in other examples, similar but different transformation function forms can also be used, such as selecting different activation functions, modifying the form and number of parameters in the above formula, and so on.

Then, take the intermediate vector of the i-th associated node u ⁱ (t) among the k associated nodes

And the corresponding i-th first transformation vector

And the i-th second transform vector

The combination operation is performed, and then k operation results are obtained, and the k operation results are summed to obtain the combined vector V.

Specifically, in an example, the above-mentioned combination operation can be a bitwise multiplication between three vectors, namely

The ⊙ symbol means bitwise multiplication. In other examples, the above-mentioned combination operation may also be other vector operations such as addition. When the combination operation is bitwise multiplication, the resulting combination vector V can be expressed as:

In addition, the node characteristics of the node

With k implicit vectors

(Where i is from 1 to k), input the third transformation function p and the fourth transformation function o, respectively, to obtain the third transformation vector p _u (t) and the fourth transformation vector o _u (t).

Specifically, in the example shown in Figure 9, the third transformation function p can be obtained by first obtaining the vectors z _u (t) and s _u (t), and then _{performing z u} (t) and s _u (t) Combining operations, thereby obtaining the third transformation vector p _u (t). For example, in a specific example:

p _u (t) = z _u (t)⊙s _u (t) (4)

Among them, ⊙ means bitwise multiplication.

More specifically, z _u (t) and s _u (t) can be calculated according to the following formula:

Among them, W _z ,

W _s and

Is the parameter matrix of linear transformation, and b _z and b _s are offset parameters.

The fourth transformation function o can be obtained by obtaining the fourth transformation vector o _u (t) through the following formula:

Among them, W _o and

Is the parameter matrix of linear transformation, and b _o is the offset parameter.

Next, based on the above-mentioned combination vector V and the third transformation vector p _u (t), the intermediate vector c _{u(t) of the} node u(t) is determined. For example, the combined vector V and the third transformation vector p _u (t) can be summed to obtain the intermediate vector c _{u(t) of u(t)} . In a specific example, the intermediate vector c _u(t) can be expressed as:

In other examples, other combination methods, such as weighted summation and bitwise multiplication, can be used to combine the combination vector V and the third transform vector, and the intermediate vector c _{u(t) can} be obtained according to the combination result.

In addition, based on the intermediate vector c _u(t) and the fourth transformation vector o _u (t) of the node u(t) thus obtained, the implicit vector h _{u(t) of} the node u(t) is determined.

In the specific example shown in Figure 9, the intermediate vector c _u(t) can be combined with the fourth transformation vector o _u (t) after performing the tanh function operation, such as bitwise multiplication, and the combined result is used as the node The implicit vector h _{u(t) of u(t)} , namely:

h _u(t) = o _u(t) ⊙tanh(c _u(t) ) (9)

Therefore, according to the structure and algorithm shown in Figure 9, the LSTM layer determines the node u(t) based on the node characteristics of the current processing node u(t), the intermediate vectors and hidden vectors of the k associated nodes pointed to by the node ) Is the intermediate vector c _u(t) and the implicit vector h _u(t) .

In one embodiment, in the process of iteratively processing each node u(t) to determine its intermediate vector and implicit vector, the interaction time corresponding to the current processing node u(t) and the k associated nodes pointed to are further introduced The time difference Δ between the corresponding interaction times. Specifically, assuming that the current processing node u(t) corresponds to the first interaction time t1, then, according to the previous description of the dynamic interaction graph, the connected k associated nodes are: the object corresponding to node u(t) participated in the last time For nodes corresponding to several interaction events that occur at the same time, the time when these interaction events occur simultaneously is recorded as the second interaction time t2. Then, the above-mentioned time difference Δ is the time difference between the first interaction time t1 and the second interaction time t2. In this way, the LSTM layer can determine the node u(t) according to the node characteristics of the current processing node u(t), the respective intermediate vectors and hidden vectors of the k associated nodes pointed to by the node u(t), and the above-mentioned time difference Δ. ) Implied vector h _u(t) and intermediate vector c _u(t) .

More specifically, on the basis of the method shown in FIG. 9, the factor of the time difference Δ can be introduced to similarly obtain the hidden vector and the intermediate vector of the node u(t). Specifically, a processing process combining the time difference may include: combining the node characteristics of the second node u(t) and the time difference Δ with the k implicit vectors corresponding to the k associated nodes, respectively, and inputting the first transformation function g , Obtain k first transformation vectors; combine the node characteristics of the second node with the k implicit vectors corresponding to the k associated nodes, and input the second transformation function f to obtain k second transformation vectors; The intermediate vector of the i-th associated node among the k associated nodes is combined with the corresponding i-th first transform vector and the i-th second transform vector to obtain k operation results, and sum the k operation results , Obtain the combination vector; input the node feature of the second node together with the k implicit vectors, respectively, input the third transformation function and the fourth transformation function to obtain the third transformation vector and the fourth transformation vector; based on the combination vector and the third transformation vectors, determining a second point u (t) of the intermediate vector C _{u (t);} the second node based on C intermediate vector _{u (t)} and a fourth transformation vectors, determining a second point u The implicit vector h _{u(t) of (t)} .

FIG. 10 shows the structure of an LSTM layer according to another embodiment. Comparing Fig. 10 and Fig. 9, it can be seen that the structure of Fig. 10 and the implemented algorithm are similar to Fig. 9, except that the time difference Δ(u, t) is further introduced on the basis of Fig. 9. In the example in Figure 10, the time difference Δ(u,t) and the node characteristics of the node u(t)

Together, they are combined with the implicit vector of each associated node and input into the first transformation function g. Correspondingly, the first transformation function g can be modified to:

Among them, formula (10) further introduces the time term corresponding to the time difference Δ(u,t) on the basis of formula (1), correspondingly,

It is a parameter for the time term, which can be embodied as a vector.

The other transformation functions in FIG. 10 and the calculation process between the functions can be the same as the example described in conjunction with FIG. 9.

According to another embodiment, the process of combining the time difference may include the following steps: After the node characteristics of the second node u(t) and the time difference Δ are respectively combined with the k implicit vectors corresponding to the k associated nodes, Input the first transformation function g and the second transformation function f with the same algorithm and different parameters to obtain k first transformation vectors and k second transformation vectors respectively; compare the intermediate vector of the i-th associated node among the k associated nodes with The corresponding i-th first transform vector and the i-th second transform vector are combined to obtain k operation results, and the k operation results are summed to obtain the combined vector; the node feature of the second node is combined with the k implicit vectors, input the third transformation function and the fourth transformation function, respectively, to obtain the third transformation vector and the fourth transformation vector; based on the combination vector and the third transformation vector, determine the second node u(t ) C intermediate vector _{u (t);} the second node based on C intermediate vector _{u (t)} and a fourth transformation vectors, determining a second point u (t) implicit vector h _{u (t).}

FIG. 11 shows the structure of an LSTM layer according to still another embodiment. It can be seen that the LSTM layer of Fig. 11 also introduces a time difference Δ(u, t), and, compared to Fig. 10, the time difference Δ(u, t) in Fig. 11 is further input to the second transformation function f. In other words, the time difference Δ(u,t) and the node characteristics of the node u(t)

At the same time, they are combined with the implicit vector of each associated node and input into the first transformation function g and the second transformation function f.

More specifically, the first transformation function g in FIG. 11 can still take the form of formula (10). Further, the second transformation function f may take the following form:

Among them, formula (11) further introduces the time term corresponding to the time difference Δ(u,t) on the basis of formula (2), correspondingly,

It is a parameter for the time term, which can be embodied as a vector.

The other transformation functions in FIG. 11 and the operation process between the functions can be the same as the example described in conjunction with FIG. 9.

In more embodiments, the above-mentioned time difference may be further input to the third transformation function p and/or the fourth transformation function o. In this case, part or all of the aforementioned formulas (5), (6), (7) can be modified, and the time term for the time difference can be similarly introduced on the original basis, which will not be detailed here.

Through the LSTM layer described above in conjunction with Figures 9-11, each node in the current subgraph is processed iteratively in order to obtain the intermediate vector and the implicit vector of the current node. In one embodiment, the hidden vector thus obtained can be used as the output of the neural network model to characterize the current node.

It can be seen that the above LSTM-based neural network model is different from the conventional LSTM network, but has been modified and optimized for the processing of multiple dynamic interactive graphs. It can be called the dynamic graph LSTM neural network model.

According to an embodiment, in order to further improve the effect, the dynamic graph LSTM neural network model may include multiple LSTM layers, where the implicit vector of a certain node determined by the previous LSTM layer is input to the next LSTM layer as the node Node characteristics. In other words, each LSTM layer still processes each node iteratively, and determines the implicit vector of the node i according to the node characteristics of the current processing node i, the respective intermediate vectors and implicit vectors of the k associated nodes pointed to by the node i Vector and intermediate vector, but the bottom LSTM layer uses the original feature of node i as the node feature, and the subsequent LSTM layer uses the hidden vector h _i of the node i determined by the previous LSTM layer as the node feature. In one embodiment, the above-mentioned multiple LSTM layers are stacked in a residual network manner to form a neural network model.

In the case that the neural network model has multiple LSTM layers, it can be understood that each LSTM layer can determine the hidden vector of the current node. In one embodiment, the neural network model synthesizes the hidden vectors of the current node output by each of the multiple LSTM layers to obtain the final hidden vector of the current node. More specifically, each implicit vector output by each LSTM layer can be weighted and combined to obtain the final implicit vector. The weight of the weighted combination can be simply set as each layer corresponds to a weight factor, and its size is adjusted through training. Alternatively, a more complex attention mechanism can also be used to determine the weight factor.

In another embodiment, the neural network model may also use the hidden vector of the current node output by the last LSTM layer among the multiple LSTM layers as the final hidden vector of the current node.

In this way, in a variety of ways, the LSTM-based neural network model is based on the current subgraph corresponding to the current node to be analyzed, and obtains the hidden vector of the current node as its feature vector. Since the current subgraph reflects the time-series interaction history information related to the interactive object corresponding to the current node, the feature vector of the current node thus obtained not only expresses the characteristics of the interactive object itself, but also expresses the fact that the interactive object is in The impact of previous interaction events, which fully characterizes the characteristics of the interactive objects.

Therefore, in step 34, the business processing related to the current node is performed according to the implicit vector of the current node.

In an embodiment, the foregoing business processing may be to predict the classification category of the object corresponding to the current node based on the implicit vector obtained above.

For example, in the case where the object corresponding to the current node is a user, the user category of the user can be predicted based on the implicit vector, such as the category of the group to which it belongs, the category of risk level, and so on. In the case where the object corresponding to the current node is an item, the category of the item can be predicted based on the implicit vector, such as the category of the business to which it belongs, the category of suitable people, the category of the scene being purchased, and so on.

In an embodiment, the business processing may further include analyzing and predicting interaction events related to the current node. Since interaction events generally involve two objects, it is also necessary to analyze the feature vector of another node.

Specifically, another node different from the aforementioned current node may be selected in the dynamic interaction graph, for example, v(t ₆ ) in FIG. 4 and FIG. 5. In a manner similar to

steps

32 and 33 in FIG. 3, the hidden vector corresponding to the other node is determined. In one embodiment, it is possible to predict whether the objects represented by the two nodes will interact based on the implicit vectors respectively corresponding to the current node and the other node. In another embodiment, the aforementioned current node and another node are two nodes corresponding to the first interaction event that has occurred. Then, the event category of the first interaction event can be predicted according to the implicit vectors corresponding to the two nodes respectively.

For example, in an example, the user represented by the current node has confirmed to purchase the commodity represented by the other node, and the first interaction event is thus generated. When the user requests payment, the implicit vector of the two nodes can be used to predict whether the first interaction event is a fraudulent transaction involving account embezzlement, so as to determine whether the payment is allowed. In another example, the user represented by the current node has already performed a comment operation on an item (such as a movie) represented by another node, such as liking or posting a text comment, thereby generating the first interaction event. After that, it can be predicted whether the first interaction event is a real operation based on the implicit vectors of the two nodes.

It can be understood that the basis of the above-mentioned service processing is the implicit vector of the node determined by the LSTM neural network model based on the dynamic interaction graph. As mentioned above, the calculation process of the LSTM neural network model to determine the hidden vector of the node depends on a large number of parameters, such as the parameters in the aforementioned transformation functions. These parameters need to be determined by training the neural network model. In different embodiments, the neural network model can be trained through different tasks.

In one embodiment, the neural network model is trained by predicting the interaction behavior. Fig. 12 shows a flowchart of training a neural network model in this embodiment. As shown in FIG. 12, in step 121, a historical interaction event is acquired, and the historical interaction event is an interaction event that has indeed occurred. In a specific example, historical interaction events can be obtained from the aforementioned collection of interaction events. The two objects included in the historical interaction event are called the first sample object and the second sample object.

In step 122, in the dynamic interaction graph, a first subgraph corresponding to the first sample object and a second subgraph corresponding to the second sample object are respectively determined. Specifically, the first sample node corresponding to the first sample object and the second sample node corresponding to the second sample object are respectively determined in the dynamic interaction graph, and the first sample node and the second sample node are respectively taken as The current node determines the corresponding first sub-graph and second sub-graph in a similar manner to step 32 in FIG. 3.

Then, in step 123, the above-mentioned first subgraph and the second subgraph are respectively input into the neural network model, and the hidden vector of the first sample object and the hidden vector of the second sample object are obtained respectively. The specific process of the neural network model to determine the hidden vector corresponding to the sample object based on the pointing relationship of the nodes in the subgraph is as described above in conjunction with step 33, and will not be repeated.

Next, in step 124, according to the implicit vector of the first sample object and the implicit vector of the second sample object, predict whether the first sample object and the second sample object will interact, and obtain the prediction result. Generally, a two-class classifier can be used to predict whether two sample objects will interact, and the obtained prediction result is usually expressed as the probability of the two sample objects interacting.

Therefore, in step 125, the prediction loss is determined based on the above prediction result. It can be understood that the above-mentioned first sample object and second sample object come from historical interaction events, so the interaction has actually occurred, which is equivalent to knowing the relationship label between the two sample objects. According to the loss function form such as the cross entropy calculation method, the loss of this prediction can be determined based on the above prediction result.

Then, in step 126, the neural network model is updated based on the predicted loss. Specifically, methods such as gradient descent and back propagation can be used to adjust the parameters in the neural network to update the neural network model until the prediction accuracy of the neural network model reaches a certain requirement.

The foregoing uses two sample objects in historical interaction events to predict the relationship between objects, which is equivalent to using positive samples for training. In an embodiment, two sample objects that have not interacted with each other can also be found in the dynamic interaction graph as negative samples for further training, so as to achieve a better training effect.

According to another embodiment, the neural network model is trained by predicting the classification of interactive objects. FIG. 13 shows a flowchart of training the neural network model in this embodiment. As shown in FIG. 13, in step 131, a sample object is selected from each object involved in the interaction event set, and the classification label of the sample object is obtained. The sample object may be any interaction object in any event included in the interaction event set, and the classification label for the sample object may be a label related to a business scenario. For example, in the case where the sample object is a user, the classification label may be a pre-set group classification label or a user risk level classification label; in the case where the sample object is a commodity, the classification label may be a commodity classification label. Such labels can be generated by manual labeling, or generated through other business-related processing.

In step 132, in the dynamic interaction graph, a sample subgraph corresponding to the sample object is determined. Specifically, the node corresponding to the sample object can be determined in the dynamic interaction graph, and the node is used as the current node to determine the corresponding sample subgraph in a similar manner to step 32 in FIG. 3.

Then, in step 133, the above-mentioned sample subgraph is input into the neural network model to obtain the hidden vector of the sample object. This process is the same as that described in step 33, and will not be repeated here.

Next, in step 134, the classification of the sample object is predicted according to the hidden vector of the sample object, and the prediction result is obtained. A classifier can be used to predict each probability that the sample object belongs to each category as the prediction result.

Then, in step 135, the prediction loss is determined based on the prediction result and the classification label. Specifically, for example, a cross-entropy calculation method can be used to predict each probability and classification label in the result, and determine the loss of this prediction.

In step 136, the neural network model is updated based on the predicted loss. In this way, the neural network model is trained by predicting the task of classifying sample objects.

In summary, in the solution of the embodiment of the present specification, a dynamic interaction diagram is constructed based on a set of interaction events, and the dynamic interaction diagram can reflect the timing relationship of each interaction event and the mutual influence between interactive objects transmitted through each interaction event. Considering the possibility of interaction events occurring at the same time, the dynamic interaction graph allows nodes to be connected to an unlimited number of associated nodes, thereby forming a mixed and diverse interaction graph. Using the trained LSTM neural network model, based on the subgraphs related to the interactive object to be analyzed in the dynamic interaction graph, the hidden vector of the interactive object can be extracted. The implicit vector thus obtained introduces the influence of other interactive objects in each interactive event on it, so that the in-depth characteristics of the interactive object can be comprehensively expressed for business processing.

According to another embodiment, an apparatus for processing interactive data is provided. The apparatus can be deployed in any device, platform, or device cluster with computing and processing capabilities. Fig. 14 shows a schematic block diagram of an apparatus for processing interactive data according to an embodiment. As shown in FIG. 14, the processing device 140 includes the following units.

The interaction graph obtaining unit 141 is configured to obtain a dynamic interaction graph constructed according to an interaction event set, where the interaction event set includes a plurality of interaction events, and each interaction event includes at least two objects on which an interaction behavior occurs and an interaction time; The dynamic interaction graph includes any first node, the first node corresponds to the first object in the interaction event that occurs at the first time, and the first node points to the M corresponding to the N associated events through the connecting edge. Associated nodes, the N associated events all occur at a second time, and all include the first object as one of the interactive objects, and the second time is backtracking from the first time, the The time before the interaction behavior of the first object occurs; the dynamic interaction graph includes at least one multi-element node whose number of associated nodes is greater than 2.

The subgraph determining unit 142 is configured to determine, in the dynamic interaction graph, a current subgraph corresponding to the current node to be analyzed, and the current subgraph includes nodes within a predetermined range that start from the current node and reach via the connecting edge. ；

The subgraph processing unit 143 is configured to input the current subgraph into a neural network model, the neural network model including an LSTM layer, and the LSTM layer is based on the direction relationship of the connecting edges between the nodes in the current subgraph, Iteratively process each node in turn to obtain the implicit vector of the current node; wherein each node includes a second node, and the iterative processing of each node in turn includes, at least according to the node characteristics of the second node, the second node The respective intermediate vectors and implicit vectors of the k associated nodes pointed to by the node determine the implicit vector and the intermediate vector of the second node.

The service processing unit 144 is configured to perform service processing related to the current node according to the implicit vector of the current node.

In one embodiment, the object includes a user, and the interaction event includes at least one of the following: a click event, a social event, and a transaction event.

In different implementations, the above-mentioned M associated nodes may be 2N nodes, respectively corresponding to the two objects included in each associated event in the N associated events; or, there may be N+1 nodes, respectively corresponding to N other objects interacting with the first object in the N associated events, and the first object itself.

In different embodiments, the nodes within the predetermined range may include: nodes within a connecting edge of a preset order K; and/or nodes whose interaction time is within a preset time range.

According to an embodiment, the aforementioned current node is a node: in the dynamic interaction graph, there is no connecting edge pointing to the node.

In one embodiment, the LSTM layer in the neural network model used by the subgraph processing unit 143 is specifically used to: use the node characteristics of the second node to k implicit vectors corresponding to the k associated nodes. Combine, input the first transformation function and the second transformation function with the same algorithm and different parameters to obtain k first transformation vectors and k second transformation vectors respectively; the intermediate vector of the i-th associated node among the k associated nodes , Perform a combination operation with the corresponding i-th first transform vector and the i-th second transform vector to obtain k operation results, and sum the k operation results to obtain the combined vector; and the node of the second node The features together with the k hidden vectors are input into the third transformation function and the fourth transformation function respectively to obtain the third transformation vector and the fourth transformation vector; based on the combination vector and the third transformation vector, the second transformation vector is determined The intermediate vector of the node; based on the intermediate vector of the second node and the fourth transform vector, the implicit vector of the second node is determined.

According to an embodiment, the LSTM layer in the neural network model used by the subgraph processing unit 143 is used to: according to the node characteristics of the second node, the respective intermediate vectors of the k associated nodes pointed to by the second node and The implicit vector and the time difference between the interaction time corresponding to the second node and the interaction time corresponding to the k associated nodes determine the implicit vector and the intermediate vector of the second node.

More specifically, in an embodiment, the above-mentioned LSTM layer is specifically configured to: combine the node characteristics of the second node and the time difference with the k implicit vectors corresponding to the k associated nodes, respectively, and input the first A transformation function to obtain k first transformation vectors; respectively combine the node characteristics of the second node with the k implicit vectors corresponding to the k associated nodes, and input the second transformation function to obtain k second transformations Vector; Combine the intermediate vector of the i-th associated node among the k associated nodes with the corresponding i-th first transform vector and the i-th second transform vector to obtain k operation results, and the k operation results Sum to obtain a combined vector; input the node feature of the second node together with the k hidden vectors into the third transform function and the fourth transform function, respectively, to obtain the third transform vector and the fourth transform vector; based on The combined vector and the third transformation vector determine the intermediate vector of the second node; and the implicit vector of the second node is determined based on the intermediate vector and the fourth transformation vector of the second node.

In another embodiment, the above-mentioned LSTM layer is specifically used to: after the node characteristics of the second node and the time difference are combined with the k implicit vectors corresponding to the k associated nodes, the input algorithm is the same, The first transformation function and the second transformation function with different parameters respectively obtain k first transformation vectors and k second transformation vectors; compare the intermediate vector of the i-th associated node among the k associated nodes with the corresponding i-th The first transform vector and the i-th second transform vector are combined to obtain k operation results, and the k operation results are summed to obtain a combined vector; the node feature of the second node is combined with the k implicit Vector, input the third transformation function and the fourth transformation function respectively to obtain the third transformation vector and the fourth transformation vector; based on the combination vector and the third transformation vector, determine the intermediate vector of the second node; based on the The intermediate vector and the fourth transform vector of the second node determine the implicit vector of the second node.

According to an embodiment, the neural network model includes multiple LSTM layers, wherein the implicit vector of the second node determined by the previous LSTM layer is input to the next LSTM layer as the node feature of the second node.

In this case, in one embodiment, the neural network model synthesizes the hidden vectors of the current node output by each of the multiple LSTM layers to obtain the final hidden vector of the current node.

In another embodiment, the neural network model uses the hidden vector of the current node output by the last LSTM layer in the plurality of LSTM layers as the final hidden vector of the current node.

According to an embodiment, the neural network model is trained by the model training unit 145. The model training unit 145 may be included in the device 140 or located outside it. The model training unit 145 may include (not shown): a sample acquisition module configured to acquire historical interaction events, which includes a first sample object and a second sample object; and a sub-picture determining module, configured to be in the dynamic interaction diagram , Respectively determine a first sub-picture corresponding to the first sample object and a second sub-picture corresponding to the second sample object; the vector acquisition module is configured to combine the first sub-picture and the first sub-picture The two sub-images are input into the neural network model respectively to obtain the implicit vector of the first sample object and the implicit vector of the second sample object; the prediction module is configured to be based on the implicit vector of the first sample object The vector and the implicit vector of the second sample object predict whether the first sample object and the second sample object will interact to obtain the prediction result; the loss determination module is configured to determine the prediction loss according to the prediction result; update Module, configured to update the neural network model according to the predicted loss.

In another embodiment, the model training unit 145 may include (not shown): a sample acquisition module configured to select a sample object from a plurality of sample objects involved in the set of interaction events, and obtain a classification label of the sample object A subgraph determining module, configured to determine a sample subgraph corresponding to the sample object in the dynamic interaction graph; a vector acquisition module, configured to input the sample subgraph into the neural network model to obtain the sample A hidden vector of the object; a prediction module configured to predict the classification of the sample object according to the hidden vector of the sample object to obtain a prediction result; a loss determination module configured to according to the prediction result and the classification label, Determine the prediction loss; an update module configured to update the neural network model according to the prediction loss.

Through the above device, based on the dynamic interactive graph, the neural network model is used to process the interactive objects, and feature vectors suitable for subsequent analysis are obtained.

According to another embodiment, there is also provided a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method described in conjunction with FIG. 3.

According to an embodiment of still another aspect, there is also provided a computing device, including a memory and a processor, the memory is stored with executable code, and when the processor executes the executable code, it implements the method described in conjunction with FIG. 3 method.

Those skilled in the art should be aware that, in one or more of the above examples, the functions described in the present invention can be implemented by hardware, software, firmware, or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on the computer-readable medium.

The specific embodiments described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention, and are not intended to limit the scope of the present invention. The protection scope, any modification, equivalent replacement, improvement, etc. made on the basis of the technical solution of the present invention shall be included in the protection scope of the present invention.

Claims

A method for processing interactive data, the method comprising:

Acquire a dynamic interaction diagram constructed according to an interaction event set, wherein the interaction event set includes a plurality of interaction events, each interaction event includes at least two objects where the interaction behavior occurs and the interaction time; the dynamic interaction diagram includes any The first node, the first node corresponds to the first object in the interaction event occurring at the first time, the first node points to the M associated nodes corresponding to the N associated events through the connecting edge, and the N The associated events all occur at the second time, and they all include the first object as one of the interactive objects. The second time is, going back from the first time, before the first object interacts. At a time; the dynamic interaction graph includes at least one multi-node with more than 2 associated nodes;

In the dynamic interaction graph, determine a current subgraph corresponding to the current node to be analyzed, and the current subgraph includes nodes within a predetermined range that start from the current node and reach via a connecting edge;

The current subgraph is input into a neural network model, and the neural network model includes an LSTM layer, and the LSTM layer processes each node in turn iteratively according to the direction relationship of the connecting edge between each node in the current subgraph, so as to obtain The implicit vector of the current node; wherein each node includes a second node, and the sequential iterative processing of each node includes, at least according to the node characteristics of the second node, the k associated nodes pointed to by the second node Respective intermediate vectors and hidden vectors, determining the hidden vector and intermediate vector of the second node;

Perform business processing related to the current node according to the implicit vector of the current node.
The method according to claim 1, wherein the object includes a user, and the interaction event includes at least one of the following: a click event, a social event, and a transaction event.
The method of claim 1, wherein:

The M associated nodes are 2N nodes, respectively corresponding to two objects included in each associated event in the N associated events; or,

The M associated nodes are N+1 nodes, respectively corresponding to N other objects interacting with the first object in the N associated events, and the first object itself.
The method according to claim 1, wherein the nodes within the predetermined range include:

Nodes within the connecting edges of the preset order K; and/or

Nodes whose interaction time is within the preset time range.
The method according to claim 1, wherein each interaction event further includes behavior characteristics of the interaction behavior;

The node characteristics of the second node include the attribute characteristics of the object corresponding to the second node, and the behavior characteristics of the interaction events that the second node participates in during the corresponding interaction time.
The method according to claim 1, wherein the determining the implicit vector and the intermediate vector of the second node comprises:

Combine the node characteristics of the second node with the k implicit vectors corresponding to the k associated nodes, and input the first transformation function and the second transformation function with the same algorithm and different parameters to obtain k first transformation functions. Transformation vector and k second transformation vectors;

Combine the intermediate vector of the i-th associated node among the k associated nodes with the corresponding i-th first transform vector and the i-th second transform vector to obtain k operation results, and the k operation results Sum to get the combined vector;

Inputting the node feature of the second node together with the k implicit vectors into a third transformation function and a fourth transformation function, respectively, to obtain a third transformation vector and a fourth transformation vector;

Determining the intermediate vector of the second node based on the combination vector and the third transformation vector;

Based on the intermediate vector and the fourth transform vector of the second node, the implicit vector of the second node is determined.
The method according to claim 1, wherein the sequentially iteratively processing each node comprises, according to the node characteristics of the second node, the respective intermediate vectors and implicit vectors of the k associated nodes pointed to by the second node, And the time difference between the interaction time corresponding to the second node and the interaction time corresponding to the k associated nodes to determine the implicit vector and the intermediate vector of the second node.
The method according to claim 7, wherein said determining the implicit vector and the intermediate vector of the second node comprises:

Combining the node characteristics of the second node and the time difference with the k implicit vectors corresponding to the k associated nodes, respectively, and inputting a first transformation function to obtain k first transformation vectors;

Combine the node features of the second node and the k implicit vectors corresponding to the k associated nodes respectively, and input a second transformation function to obtain k second transformation vectors;

Combine the intermediate vector of the i-th associated node among the k associated nodes with the corresponding i-th first transform vector and the i-th second transform vector to obtain k operation results, and sum the k operation results , Get the combined vector;

Inputting the node feature of the second node together with the k implicit vectors into a third transformation function and a fourth transformation function, respectively, to obtain a third transformation vector and a fourth transformation vector;

Determining the intermediate vector of the second node based on the combination vector and the third transformation vector;

Based on the intermediate vector and the fourth transform vector of the second node, the implicit vector of the second node is determined.
The method according to claim 7, wherein said determining the implicit vector and the intermediate vector of the second node comprises:

After combining the node characteristics of the second node and the time difference with the k implicit vectors corresponding to the k associated nodes, respectively, the first transformation function and the second transformation function with the same algorithm and different parameters are input, respectively Obtain k first transform vectors and k second transform vectors;

Combine the intermediate vector of the i-th associated node among the k associated nodes with the corresponding i-th first transform vector and the i-th second transform vector to obtain k operation results, and sum the k operation results , Get the combined vector;

Inputting the node feature of the second node together with the k implicit vectors into a third transformation function and a fourth transformation function, respectively, to obtain a third transformation vector and a fourth transformation vector;

Determining the intermediate vector of the second node based on the combination vector and the third transformation vector;

Based on the intermediate vector and the fourth transform vector of the second node, the implicit vector of the second node is determined.
The method according to claim 1, wherein the neural network model includes a plurality of LSTM layers, wherein the implicit vector of the second node determined by the previous LSTM layer is input to the next LSTM layer as the first The node characteristics of the two nodes.
The method according to claim 10, wherein the neural network model integrates the hidden vectors of the current node output by each of the multiple LSTM layers to obtain the final hidden vector of the current node.
The method according to claim 10, wherein the neural network model uses the implicit vector of the current node output by the last LSTM layer among the plurality of LSTM layers as the final implicit vector of the current node.
The method according to claim 1, wherein the neural network model is trained in the following manner:

Obtain historical interaction events, including the first sample object and the second sample object;

In the dynamic interaction graph, a first subgraph corresponding to the first sample object and a second subgraph corresponding to the second sample object are respectively determined;

Input the first subgraph and the second subgraph into the neural network model respectively to obtain the hidden vector of the first sample object and the hidden vector of the second sample object;

According to the implicit vector of the first sample object and the implicit vector of the second sample object, predict whether the first sample object and the second sample object will interact, and obtain a prediction result;

Determine the predicted loss according to the predicted result;

According to the predicted loss, the neural network model is updated.
The method according to claim 1, wherein the neural network model is trained in the following manner:

Selecting a sample object from a plurality of sample objects involved in the interaction event set, and obtaining a classification label of the sample object;

In the dynamic interaction diagram, determine the sample sub-image corresponding to the sample object;

Input the sample sub-image into the neural network model to obtain the implicit vector of the sample object;

Predict the classification of the sample object according to the implicit vector of the sample object to obtain the prediction result;

Determine the prediction loss according to the prediction result and the classification label;

According to the predicted loss, the neural network model is updated.
A device for processing interactive data, the device comprising:

The interaction diagram obtaining unit is configured to obtain a dynamic interaction diagram constructed according to a set of interaction events, wherein the set of interaction events includes a plurality of interaction events, and each interaction event includes at least two objects on which the interaction behavior occurs and the interaction time; The dynamic interaction graph includes any first node, the first node corresponds to the first object in the interaction event occurring at the first time, and the first node points to M corresponding to the N associated events through the connecting edge. Associated node, the N associated events all occur at the second time, and all include the first object as one of the interactive objects, the second time is, going back from the first time, the first time The time before the interaction of an object; the dynamic interaction graph includes at least one multi-node with more than 2 associated nodes;

A subgraph determining unit configured to determine, in the dynamic interaction graph, a current subgraph corresponding to the current node to be analyzed, the current subgraph including nodes within a predetermined range that start from the current node and reach via a connecting edge;

The subgraph processing unit is configured to input the current subgraph into a neural network model, the neural network model including an LSTM layer, and the LSTM layer sequentially Iteratively process each node to obtain the implicit vector of the current node; wherein each node includes a second node, and the sequential iterative processing of each node includes, at least according to the node characteristics of the second node, the second node Determine the hidden vector and the intermediate vector of the respective intermediate vectors and hidden vectors of the k associated nodes pointed to;

The service processing unit is configured to perform service processing related to the current node according to the implicit vector of the current node.
The apparatus according to claim 15, wherein the object includes a user, and the interaction event includes at least one of the following: a click event, a social event, and a transaction event.
The apparatus according to claim 15, wherein the M associated nodes are 2N nodes, respectively corresponding to two objects included in each associated event in the N associated events; or,

The M associated nodes are N+1 nodes, respectively corresponding to N other objects interacting with the first object in the N associated events, and the first object itself.
The apparatus according to claim 15, wherein the nodes in the predetermined range comprise:

Nodes within the connecting edges of the preset order K; and/or

Nodes whose interaction time is within the preset time range.
The apparatus according to claim 15, wherein each of the interaction events further comprises behavior characteristics of the interaction behavior;

The node characteristics of the second node include the attribute characteristics of the object corresponding to the second node, and the behavior characteristics of the interaction events that the second node participates in during the corresponding interaction time.
The apparatus according to claim 15, wherein the LSTM layer is used to:

Combine the node characteristics of the second node with the k implicit vectors corresponding to the k associated nodes, and input the first transformation function and the second transformation function with the same algorithm and different parameters to obtain k first transformation functions. Transformation vector and k second transformation vectors;

Combine the intermediate vector of the i-th associated node among the k associated nodes with the corresponding i-th first transform vector and the i-th second transform vector to obtain k operation results, and the k operation results Sum to get the combined vector;

Inputting the node feature of the second node together with the k implicit vectors into a third transformation function and a fourth transformation function, respectively, to obtain a third transformation vector and a fourth transformation vector;

Determining the intermediate vector of the second node based on the combination vector and the third transformation vector;

Based on the intermediate vector and the fourth transform vector of the second node, the implicit vector of the second node is determined.
The apparatus according to claim 15, wherein the LSTM layer is used to: according to the node characteristics of the second node, the respective intermediate vectors and implicit vectors of the k associated nodes pointed to by the second node, and the The time difference between the interaction time corresponding to the second node and the interaction time corresponding to the k associated nodes determines the implicit vector and the intermediate vector of the second node.
The device according to claim 21, wherein the LSTM layer is specifically used for:

Combining the node characteristics of the second node and the time difference with the k implicit vectors corresponding to the k associated nodes, respectively, and inputting a first transformation function to obtain k first transformation vectors;

Combine the node features of the second node and the k implicit vectors corresponding to the k associated nodes respectively, and input a second transformation function to obtain k second transformation vectors;

Combine the intermediate vector of the i-th associated node among the k associated nodes with the corresponding i-th first transform vector and the i-th second transform vector to obtain k operation results, and sum the k operation results , Get the combined vector;

Inputting the node feature of the second node together with the k implicit vectors into a third transformation function and a fourth transformation function, respectively, to obtain a third transformation vector and a fourth transformation vector;

Determining the intermediate vector of the second node based on the combination vector and the third transformation vector;

Based on the intermediate vector and the fourth transform vector of the second node, the implicit vector of the second node is determined.
The device according to claim 21, wherein the LSTM layer is specifically used for:

After combining the node characteristics of the second node and the time difference with the k implicit vectors corresponding to the k associated nodes, respectively, the first transformation function and the second transformation function with the same algorithm and different parameters are input, respectively Obtain k first transform vectors and k second transform vectors;

Combine the intermediate vector of the i-th associated node among the k associated nodes with the corresponding i-th first transform vector and the i-th second transform vector to obtain k operation results, and sum the k operation results , Get the combined vector;

Inputting the node feature of the second node together with the k implicit vectors into a third transformation function and a fourth transformation function, respectively, to obtain a third transformation vector and a fourth transformation vector;

Determining the intermediate vector of the second node based on the combination vector and the third transformation vector;

Based on the intermediate vector and the fourth transform vector of the second node, the implicit vector of the second node is determined.
The device according to claim 15, wherein the neural network model includes a plurality of LSTM layers, wherein the implicit vector of the second node determined by the previous LSTM layer is input to the next LSTM layer as the first The node characteristics of the two nodes.
The device according to claim 24, wherein the neural network model integrates the implicit vectors of the current node output by each of the multiple LSTM layers to obtain the final implicit vector of the current node.
The device according to claim 24, wherein the neural network model uses the implicit vector of the current node output by the last LSTM layer among the plurality of LSTM layers as the final implicit vector of the current node.
The device according to claim 15, wherein the neural network model is trained by a model training unit, and the model training unit comprises:

The sample acquisition module is configured to acquire historical interaction events, including the first sample object and the second sample object;

A subgraph determining module, configured to respectively determine a first subgraph corresponding to the first sample object and a second subgraph corresponding to the second sample object in the dynamic interaction graph;

A vector acquisition module configured to input the first sub-image and the second sub-image into the neural network model respectively to obtain the implicit vector of the first sample object and the implicit vector of the second sample object ；

A prediction module configured to predict whether the first sample object and the second sample object will interact according to the implicit vector of the first sample object and the implicit vector of the second sample object, to obtain a prediction result;

A loss determining module, configured to determine a predicted loss according to the prediction result;

The update module is configured to update the neural network model according to the predicted loss.
The device according to claim 15, wherein the neural network model is trained by a model training unit, and the model training unit comprises:

A sample acquisition module, configured to select a sample object from a plurality of sample objects involved in the set of interaction events, and obtain a classification label of the sample object;

A subgraph determining module, configured to determine a sample subgraph corresponding to the sample object in the dynamic interaction graph;

A vector acquisition module configured to input the sample sub-image into the neural network model to obtain the implicit vector of the sample object;

A prediction module configured to predict the classification of the sample object according to the implicit vector of the sample object to obtain a prediction result;

A loss determination module, configured to determine a prediction loss according to the prediction result and the classification label;

The update module is configured to update the neural network model according to the predicted loss.
A computer-readable storage medium with a computer program stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method of any one of claims 1-14.
A computing device, comprising a memory and a processor, characterized in that executable code is stored in the memory, and when the processor executes the executable code, the device described in any one of claims 1-14 is implemented method.