CN112541129A - Method and device for processing interaction event - Google Patents
Method and device for processing interaction event Download PDFInfo
- Publication number
- CN112541129A CN112541129A CN202011409575.0A CN202011409575A CN112541129A CN 112541129 A CN112541129 A CN 112541129A CN 202011409575 A CN202011409575 A CN 202011409575A CN 112541129 A CN112541129 A CN 112541129A
- Authority
- CN
- China
- Prior art keywords
- interaction
- event
- node
- target
- nodes
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9536—Search customisation based on social or collaborative filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
- G06Q40/04—Trading; Exchange, e.g. stocks, commodities, derivatives or currency exchange
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Software Systems (AREA)
- Business, Economics & Management (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Finance (AREA)
- Accounting & Taxation (AREA)
- General Business, Economics & Management (AREA)
- Technology Law (AREA)
- Strategic Management (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Marketing (AREA)
- Economics (AREA)
- Development Economics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
An embodiment of the present specification provides a method and an apparatus for processing an interaction event, where in the processing method, a newly added interaction event is obtained. In the existing dynamic interaction graph, two target nodes corresponding to the newly added interaction events are determined, and newly added connection edges are established. And at least fusing the node information of the two target nodes to obtain event interaction information. And respectively updating the implicit vectors of the two target nodes according to the interaction time difference between the recent historical interaction event where the two target nodes are respectively located and the newly added interaction event and the event interaction information. And determining neighbor nodes of each order. Aiming at any first neighbor node in each order of neighbor nodes, based on the interaction time difference between the latest historical interaction event where the first neighbor node is located and the newly added interaction event, the first distance between the first neighbor node and the corresponding target node, and the updated implicit vectors of the two target nodes, the propagation information corresponding to the first neighbor node is determined. And updating the implicit vector of the first neighbor node according to the propagation information.
Description
Technical Field
One or more embodiments of the present specification relate to the field of machine learning, and more particularly, to a method and apparatus for processing an interaction event using machine learning.
Background
In many scenarios, user interaction events need to be analyzed and processed. The interaction event is one of basic constituent elements of an internet event, for example, a click action when a user browses a page can be regarded as an interaction event between the user and a content block of the page, a purchase action in an e-commerce can be regarded as an interaction event between the user and a commodity, and an inter-account transfer action is an interaction event between the user and the user. The characteristics of fine-grained habit preference and the like of the user and the characteristics of an interactive object are contained in a series of interactive events of the user, and the characteristics are important characteristic sources of a machine learning model. Therefore, in many scenarios, it is desirable to characterize and model interaction participants, as well as interaction events, based on the history of the interaction.
However, an interactive event involves both interacting parties, and the status of each party itself may be dynamically changing, and thus it is very difficult to accurately characterize the interacting parties comprehensively considering their multi-aspect characteristics. Thus, improved solutions for more efficiently analyzing and processing interactive events are desired.
Disclosure of Invention
One or more embodiments of the present specification describe a method and an apparatus for processing an interaction event, which can effectively analyze and process the interaction event.
In a first aspect, a method for processing an interaction event is provided, including:
acquiring an existing dynamic interaction graph constructed based on a plurality of historical interaction events; the method comprises the steps of representing a plurality of nodes of a plurality of objects, wherein each node corresponds to an implicit vector; two nodes representing two objects in the same historical interaction event are connected through a connecting edge, and the connecting edge indicates the interaction time of the same historical interaction event;
acquiring a newly added interaction event;
determining two target nodes corresponding to the newly added interaction event in the existing dynamic interaction graph, and establishing a newly added connecting edge between the two target nodes;
at least fusing the node information of the two target nodes to obtain event interaction information; wherein the node information of each target node is determined based on at least its implicit vector;
respectively updating respective implicit vectors of the two target nodes according to respective target time differences of interaction time of the latest historical interaction events where the two target nodes are respectively located and interaction time of the newly added interaction events and the event interaction information;
respectively determining respective adjacent nodes of each order of the two target nodes;
aiming at any first neighbor node in the neighbor nodes of each order, determining propagation information corresponding to the first neighbor node based on a first time difference between the interaction time of the latest historical interaction event where the first neighbor node is located and the interaction time of the newly added interaction event, a first distance between the first neighbor node and a corresponding target node, and respective updated implicit vectors of the two target nodes;
and updating the implicit vector of the first neighbor node according to the propagation information.
In a second aspect, an apparatus for processing an interaction event is provided, including:
the acquisition unit is used for acquiring an existing dynamic interaction diagram constructed based on a plurality of historical interaction events; the method comprises the steps of representing a plurality of nodes of a plurality of objects, wherein each node corresponds to an implicit vector; two nodes representing two objects in the same historical interaction event are connected through a connecting edge, and the connecting edge indicates the interaction time of the same historical interaction event;
the acquiring unit is further used for acquiring a newly added interaction event;
a determining unit, configured to determine two target nodes corresponding to the newly added interaction event in the existing dynamic interaction graph, and establish a newly added connection edge between the two target nodes;
the fusion unit is used for fusing at least the node information of the two target nodes to obtain event interaction information; wherein the node information of each target node is determined based on at least its implicit vector;
the updating unit is used for respectively updating the implicit vectors of the two target nodes according to the target time difference between the interaction time of the recent historical interaction event where the two target nodes are respectively located and the interaction time of the newly added interaction event and the event interaction information;
the determining unit is further configured to determine respective neighboring nodes of each order of the two target nodes respectively;
the determining unit is further configured to determine, for any first neighbor node in the neighbor nodes of each order, propagation information corresponding to the first neighbor node based on a first time difference between an interaction time of a recent historical interaction event where the first neighbor node is located and an interaction time of the newly added interaction event, a first distance between the first neighbor node and a corresponding target node, and respective updated implicit vectors of the two target nodes;
the updating unit is further configured to update the implicit vector of the first neighbor node according to the propagation information.
In a third aspect, there is provided a computer storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method of the first aspect.
In a fourth aspect, there is provided a computing device comprising a memory having stored therein executable code and a processor that, when executing the executable code, implements the method of the first aspect.
According to the method and the device for processing the interaction events, which are provided by one or more embodiments of the specification, the dynamic interaction graph is constructed on the basis of a plurality of historical interaction events. For the newly added interaction event, two target nodes corresponding to the newly added interaction event are determined in the existing dynamic interaction graph, and a newly added connection edge is established between the two target nodes. And then determining event interaction information based on the node information of the two target nodes, and updating the implicit vector of each target node based on the time difference between the latest historical interaction event where the target node is located and the newly added interaction event and the event interaction information for each target node. In addition, for the respective neighbor nodes of the two target nodes, the corresponding propagation information is determined based on the time difference and the distance corresponding to each neighbor node and the respective updated implicit vectors of the two target nodes, and the implicit vectors of the neighbor nodes are updated based on the propagation information. Therefore, in the scheme, when the interaction event is captured, the implicit vector of the target node of the interaction event is updated, and the implicit vectors of the neighbor nodes of the event node are also updated, so that the feature expression of each node in the dynamic interaction graph can be updated in real time, and the interaction event can be analyzed more comprehensively and accurately.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present disclosure, and it is obvious for those skilled in the art to obtain other drawings based on the drawings without creative efforts.
FIG. 1 is a diagrammatic illustration of temporal interaction provided by the present specification;
FIG. 2 is a schematic diagram of a flow dynamics neural network provided herein;
FIG. 3 is a schematic diagram of an implementation scenario of a method for processing an interaction event provided in the present specification;
FIG. 4 is a flow diagram of a method for processing an interaction event provided by one embodiment of the present description;
FIG. 5a is one of the dynamic interaction pictorial diagrams provided by the present specification;
FIG. 5b is a second schematic view of the dynamic interaction provided in the present specification;
FIG. 6 is a flow chart of a method of training a flow dynamics neural network provided herein;
fig. 7 is a schematic diagram of an apparatus for processing an interaction event according to an embodiment of the present disclosure.
Detailed Description
The scheme provided by the specification is described below with reference to the accompanying drawings.
Before describing the solutions provided by the embodiments of the present disclosure, the following description is made of the main innovation points of the solutions.
In a traditional dynamic interaction graph construction method, when an interaction event is captured, two target nodes related to the interaction event are determined in a dynamic interaction graph, and then vector representations of the two target nodes are obtained by aggregating the latest states of respective neighbor nodes of the two target nodes. And finally, based on the vector representation of the two target nodes, performing downstream traffic prediction. Such as identifying an event category of the interaction event, etc. However, this method has problems of information loss and low prediction efficiency.
First, a problem of information loss will be described.
Fig. 1 is a schematic diagram of time interaction provided by the present specification. In FIG. 1, v1-v6 represent 6 nodes, respectively, and e1-e6 represent 6 interaction events arranged in chronological order, respectively. As can be seen from fig. 1, the state of the neighbor node v2 of the node v1 has changed several times before the interaction event e6 occurs between the node v1 and the node v 6. However, in the conventional method, only when the interaction event e6 occurs, the latest state of the node v2 is converged and the state of the node v1 is updated, so that much state change information of the neighbor node v2 is lost.
To solve the above problems, the inventors of the present application propose the following innovations: when the interaction event is captured, the state of a target node (also called an event node) related to the interaction event is updated, and the state of a neighbor node (also called an influence node) of the event node is also updated, so that the state change information of the neighbor node can be comprehensively captured in real time.
Secondly, the problem of low prediction efficiency is solved.
In the conventional method, aggregating the latest states of the neighbor nodes of any target node may include: firstly, a subgraph containing all the neighbor nodes is constructed according to the target node, and then the latest state of all the neighbor nodes in the subgraph is aggregated through a convolution network model, so that the vector representation of the target node is obtained. The computational consumption of subgraph construction and message aggregation is usually relatively large. For example, to aggregate the latest states of k-order neighbor nodes, and the number of neighbor nodes to which each target node is directly connected is n, the total consumption of the subgraph construction and the message aggregation is as follows: o (n)k) Thereby resulting in less efficient prediction of downstream predicted traffic.
To solve the above problems, the inventors of the present application propose the following innovations: the message aggregation operation and the traffic prediction operation may be performed in parallel. Therefore, the calculation complexity of the service prediction can be reduced to: o (1).
The above are two main innovation points provided in the present specification, and for the two innovation points, the present solution can be specifically implemented by a Streaming Dynamic Graph Neural network (SDHNN). The flow dynamics neural network may be as shown in fig. 2, which may include the following three modules: a delta update module 202, a message propagation module 204, and a classifier module 206. The incremental update module 202, also referred to as a neural network model, is used to update the state of the event node in a stream manner. The classifier module 206, also called classifier, is used to predict traffic according to the latest status of the event node. The message propagation module 204 is also called a multi-Layer Perceptron (MLP), and is configured to send respective corresponding propagation information to each order of neighboring nodes of the event node, and update respective states based on the respective corresponding propagation information. The propagation information corresponding to any neighbor node integrates the latest state of the event node, the relative time difference of the neighbor node, the relative position difference between the event node and the neighbor node, and other information. The specific implementation process of each module can be referred to the subsequent method steps.
In general, unlike conventional approaches, the solution provided by the present specification performs aggregation of messages for each node at multiple times. For example, once the state of a neighbor node of a node changes, the new state of the neighbor node is aggregated to the node through the message propagation module 204, which helps to capture more useful information so that an accurate vector representation can be performed for the node. In addition, in the present solution, the message propagation module 204 and the classifier module 206 may run in parallel, so that the service prediction only depends on the increment updating module 202 and the classifier module 206, thereby significantly improving the service prediction efficiency.
Finally, regarding the flow dynamics neural network, the present application proposes the following new training method:
specifically, for the current sample interaction event, after two target nodes related to the sample interaction event are determined in the dynamic interaction graph, historical propagation information of the two target nodes is obtained first, and the message propagation module 204 is used to update the states of the two target nodes for the first time based on the historical propagation information. Then, the incremental update module 202 updates the states of the two target nodes again. Finally, the classifier module 206 performs service prediction based on the updated states of the two target nodes, and obtains a prediction result. And after the prediction result is obtained, determining the prediction loss according to the prediction result and the sample label of the sample interaction event. And jointly train the delta update module 202, the message propagation module 204, and the classifier module 206 based on the predicted loss.
The above is a description of several innovative points provided by the present specification, and the present solution is described in detail below.
Fig. 3 is a schematic view of an implementation scenario of the method for processing an interaction event provided in this specification. In fig. 3, it is assumed that a new interaction event from the node vs to the node vg occurs at time t 0. Nodes vs and vg may be referred to herein as event nodes. And the current state of the node vs can be represented as hiddenContaining vector quantity
The current state of the node vg can be represented as an implicit vector
Then the incremental update module 202 may update the implicit vector of the node vs according to the event interaction information of the newly added interaction event and the relative time difference between the node vs, so as to obtain the implicit vector of the node vs
And updating the implicit vector of the node vg according to the event interaction information of the newly added interaction event and the relative time difference of the node vg to obtain
The classifier module 206 may then base the updated implicit vector on
And
and predicting the event type of the newly added interaction event.
In FIG. 3, the message propagation module 204 may separately forward the nodes while the classifier module 206 predicts the event categories
And transmitting the corresponding propagation information. Node here
And the nodes are the neighbor nodes of each order of the node vs and the node vg. Neighbor node with node vs
For the purpose of example, it is shown that,the corresponding propagation information integrates the latest state of the node vs and the node vg, and the node
Relative time difference of (2), and node vs and node
Relative positional difference of the two or more. Then, based on the respective propagation information of the neighboring nodes of each order, respectively updating the respective implicit vectors of the neighboring nodes of each order
To obtain
FIG. 4 is a flowchart of a method for processing an interaction event according to an embodiment of the present disclosure. The execution subject of the method may be a device with processing capabilities: a server or a system or device. As shown in fig. 4, the method may include the steps of:
step 402, obtaining an existing dynamic interaction graph constructed based on a plurality of historical interaction events.
Each of the historical interaction events may include: two objects where an interactive action takes place and an interaction time. In addition, event features may also be included.
The existing dynamic interaction graph may include a plurality of nodes representing objects in historical interaction events, where two nodes representing two objects in the same historical interaction event are connected by a connection edge, and the connection edge may indicate an interaction time of the same historical interaction event. Each node in the existing dynamic interaction graph may have a corresponding state, which may be represented as an implicit vector. The implicit vector at least integrates node information of the corresponding node, event interaction information of each historical interaction event where the corresponding node is located, influence information (namely propagation information) transmitted by each-order neighbor node and the like.
In this specification, the above-mentioned existing dynamic interaction graph may be represented as follows: g (V, E, T), where V and E are the set of nodes and edges, respectively, and T is the set of interaction times, which may also be referred to as a sequence of timestamps. Each edge represents an interaction event, which can be specifically represented as (vs, vg, t), and specifically means: an interaction event occurs from node vs to node vg at time t.
In one example, the existing dynamic interaction graph described above may be as shown in FIG. 5 a. In FIG. 5a, the existing dynamic interaction graph contains 18 nodes and 19 edges from interaction time t1-t 19. 19 of which are shown by solid arrows, representing that it is built based on 19 historical interaction events. And each edge may indicate a corresponding interaction time, i.e., t1-t 19.
Step 404, acquiring a new interaction event.
The new interactivity event here is similar to the historical interactivity event, and may also include: two objects where an interactive action takes place and an interaction time. In addition, event features may also be included. Taking the newly added interaction event as the transaction event as an example, the event characteristics may include at least one of the following: transaction type, transaction amount, transaction channel, and the like.
In one example, the obtained new interaction event may be: (v1, v2, t20) indicating that an interaction event from node v1 to node v2 occurred at interaction time t 20.
Step 406, determining two target nodes corresponding to the newly added interaction event in the existing dynamic interaction graph, and establishing a newly added connection edge between the two target nodes.
Here, the determining two target nodes may specifically include: and searching the two target nodes in the existing dynamic interaction graph, and if the two target nodes are searched, taking the two searched target nodes as the two finally determined target nodes. And if any one target node cannot be found, adding the target node in the existing dynamic interaction graph, and then taking the newly added target node as the finally determined target node.
In addition, for the two searched target nodes, if a connection edge already exists between the two target nodes, the establishing of the new connection edge may be understood as covering the currently existing connection edge, and the interaction time indicated by the connection edge is updated to the interaction time of the new interaction event. If there is no connecting edge between the two target nodes, the establishing of the new adding edge may be understood as establishing a new connecting edge, and establishing a corresponding relationship between the new connecting edge and the interaction time of the new interaction event, so that the new connecting edge may indicate that there is interaction time.
In summary, for the existing dynamic interaction graph, multiple edges may be established between any two nodes vi and vj at different interaction times, but only the edge established by the latest interaction time needs to be finally reserved, so that only one edge exists between any two nodes, the one edge represents the latest historical interaction event, and the interaction time indicated by the one edge is the interaction time of the latest interaction event.
For the new interaction event in the above example, two target nodes can be identified in fig. 5 a: node v1 and node v 2. Then, a connecting edge may be newly added between the node v1 and the node v2, which may be specifically shown by a dashed arrow. After adding the connecting edge in fig. 5a, it can be seen in fig. 5 b.
And step 408, at least fusing the node information of the two target nodes to obtain event interaction information.
The node information of each target node therein is determined based on at least its implicit vector.
In one example, the node information of the target node may be represented as an embedded vector u, which is used to predict the event category of the newly added interaction event. It should be appreciated that the node information of the target node may be different for different traffic predictions, and thus the embedded vector u used to represent the node information is also different.
It should be noted that, when the newly added interaction event further includes an event feature, the acquiring of the event interaction information specifically may include: and fusing the node information of the two target nodes and the event characteristics to obtain event interaction information.
In one example, the event interaction information may be obtained by the following formula 1:
ei＝σ(Ws·us+Wg·ug+wff) (equation 1)
Wherein e isiFor event interaction information, σ () is the activation function, usAnd ugRespectively, the embedded vectors of the two target nodes, and f is an event characteristic. Ws、WgAnd WfAre all learnable parameters, are respectively used for us、ugAnd f is mapped as a vector of predetermined dimensions (e.g., d dimensions).
Step 410, respectively updating the implicit vectors of the two target nodes according to the target time difference between the interaction time of the recent historical interaction event where the two target nodes are respectively located and the interaction time of the newly added interaction event, and the event interaction information.
It should be noted that, in addition to the event interaction information described above, the present application also considers the influence of the relative time difference when updating the state of the target node. The relative time difference here refers to a target time difference between the interaction time of the recent historical interaction event where the target node is located and the interaction time of the newly added interaction event. However, since the relative time difference and the implicit vector are located in different feature spaces, the relative time difference can be encoded based on a time encoding mechanism before the relative time difference is fused.
Taking the first target node of any two target nodes as an example, updating the implicit vector of the first target node while also considering the relative time difference may include: step a, adopting a first mapping algorithm to map a target time difference corresponding to a first target node (namely, a target time difference between interaction time of a recent historical interaction event where the first target node is located and interaction time of a newly added interaction event) into a target vector with a preset dimension. And b, fusing the target vector and the implicit vector of the first target node to obtain an intermediate vector. And c, inputting the intermediate vector and the event interaction information into the neural network model to obtain an updated implicit vector of the first target node.
For the step a, the target time difference corresponding to the first target node may be mapped through the following formula.
Where Δ t is a target time difference of the first target node, Base () is used to map the target time difference to a target vector of a predetermined dimension (e.g., d-dimension), Base (Δ t) is the target vector, i is a specific position of the target vector, and Base (Δ t) [ i ] is the ith element in the target vector.
For the step b, the target vector of the first target node and the implicit vector may be fused by the following formula.
ΦT(Δt)＝WrteBase (Δ t) (equation 3)
WrteIs a learnable parameter for mapping a target vector into a target feature space. The target feature space may refer to a feature space corresponding to an implicit vector of the target node. PhiT(Δ t) is a target time difference expressed in the target feature space.
Is an implicit vector of the first target node, WincAnd bincAre learnable parameters that are all learning time dependencies,
is the intermediate vector.
For the step c, the Neural Network model may be, for example, a Recurrent Neural Network (RNN), a long-short term memory (LSTM) model, or a GRU model. Taking the GRU model as an example, the step c can be expressed as the following formula:
wherein the content of the first and second substances,
an updated implicit vector for the first target node, eiIn order to exchange information for an event,
is the intermediate vector. The event interaction information can be regarded as input parameters of the GRU model,
which may be considered as the initial state of the first target node.
Similarly to the method of updating the implicit vector of the first target node, the implicit vector of the other of the two target nodes may be updated, so that an updated implicit vector of the other target node may be obtained.
In summary, the latest status of the two target nodes can be updated through the above step 410.
Step 412, determining respective neighbor nodes of each order of the two target nodes.
Taking the first target node of any of the two target nodes as an example, the respective-order neighbor nodes of the first target node may include: and starting from the first target node, each node is reached through the connecting edges within the preset number N. And/or starting from the first target node, and reaching each node through the connecting edge and with the interaction time within a preset time range. Wherein N is a positive integer.
In social, transactional, or other networks, an interaction event may affect not only the state of two target nodes, but also the state of its neighbor nodes. Therefore, the embodiments of the present specification also update the states of respective order neighbor nodes of the two target nodes.
Step 414, for any first neighboring node in the neighboring nodes of each order, determining propagation information corresponding to the first neighboring node based on a first time difference between the interaction time of the recent historical interaction event where the first neighboring node is located and the interaction time of the newly added interaction event, a first distance between the first neighboring node and a corresponding target node, and respective updated implicit vectors of the two target nodes.
The determining propagation information corresponding to the first neighboring node may specifically include: and step x, mapping the first time difference into a first vector with a preset dimension by adopting a first mapping algorithm. And step y, mapping the first distance into a second vector with a preset dimension by adopting a second mapping algorithm. And step z, fusing the first vector, the second vector and the updated implicit vectors of the two target nodes to obtain the propagation information corresponding to the first neighbor node.
For step x, the specific mapping step can be referred to the above formula 2, and this description is not repeated herein. It should be noted that after the mapping step is performed to obtain the first vector, the first vector may be further transformed by formula 3 to obtain the first time difference Φ expressed in the target feature spaceT(Δt1)。
For step y, the first distance may be mapped by the following formula.
ΦP(k) Base (k) (equation 6)
Where k is a first distance, Base () is used to map the first distance to a second vector of a predetermined dimension (e.g., d dimension), Base (k), and Φp(k) Is the second vector, i is the particular position of the second vector, base (k) i]Is the ith element in the second vector.
For step z, the propagation information corresponding to the first neighboring node may be determined by the following formula.
Wherein e ispFor the propagation information corresponding to the first neighbor node, Winc、WsAnd WgAs described above in the above-mentioned publication,
and
implicit vectors, phi, respectively updated for the two target nodesT(Δ t1) is a first time difference, Φ, expressed in the target feature spacep(k) Is the second vector.
It should be noted that, in formula 7, the first time difference is fused with the implicit vectors of the two target nodes, and then the fusion result is fused with the second vector, because the mapping manner of the first time difference and the first distance is similar. It should be understood that, by this way of fusing the present description, the model may be enabled to unambiguously distinguish between the first time difference and the first distance.
It should be understood that, similarly to the above method for determining the propagation information of the first neighboring node, the propagation information corresponding to each of the other neighboring nodes may also be determined, and the description of this specification is not repeated herein.
In addition, as can be seen from the above equations 6 and 7, the embodiments of the present specification respectively determine corresponding propagation information by combining the respective related information (e.g., time difference and distance) of each neighboring node. That is to say, for different neighbor nodes, the determined propagation information may be different, which may make the determination manner of the propagation information more reasonable, and may further improve the accuracy of vector expression of the neighbor nodes.
Step 416, updating the implicit vector of the first neighboring node according to the propagation information.
Specifically, the propagation information and the implicit vector of the first neighbor node may be integrated by using an attention mechanism to obtain an integrated result. And updating the implicit vector of the first neighbor node by utilizing the multilayer perceptron based on the integration result.
Wherein, Wq、WkAnd WvIs a parameter that can be learned by the user,
as an implicit vector of the first neighbor node, epTo correspond to the propagated information of the first neighboring node,
in order to integrate the results of the process,
an updated implicit vector for the first neighboring node. It should be noted that K and K in equation 8 are both parameters used in the attention mechanism, which are different from the first distance K described above.
Similarly, the implicit vectors of other neighboring nodes can be updated according to the respective propagation information of other neighboring nodes, so as to obtain the respective updated implicit vectors of other neighboring nodes.
It should be noted that, in the embodiment of the present specification, the distances between the neighboring nodes and the target node are integrated into the propagation information, and the self-attention mechanism is used to update the states of the neighboring nodes, so that the computational complexity of the state update of the neighboring nodes only increases linearly with the number of nodes.
In addition, when the classifier module 206 is used to predict the event category of the newly added interaction event, the method provided by the embodiment of the present specification may further include the following steps: and predicting the event category of the newly-added interaction event by utilizing a classifier based on the updated implicit vectors of the two target nodes respectively.
Specifically, the activation function may be utilized to activate the updated implicit vectors of the two target nodes, respectively, so as to obtain two embedded vectors used for representing the updated node information of the two target nodes. And inputting the two embedded vectors into a classifier to obtain the event category of the newly added interaction event.
It should be noted that, the step of obtaining the embedded vector of each of the two target nodes here can be understood as the embedded vector (u) mentioned in the above step 408sAnd ug) And (6) updating.
In one example, the embedded vectors of two target nodes may be obtained by the following formula:
wherein, Cs、Cg、bs_cAnd bg_cFor parameters that can be learned when making different traffic predictions, e.g. parameters that are learned when making a prediction of an event class of an interactivity event, σ () is an activation function,
and
respectively updated implicit vectors, U, for the two target nodessAnd UgRespectively, the updated embedded vectors of each of the two target nodes.
It should be noted that the above step of predicting the event category of the newly added interaction event may be executed in parallel with the above steps 412 to 416.
In summary, the method for processing an interaction event provided in the embodiments of the present specification updates, when a newly added interaction event is obtained, the states of two target nodes related to the newly added interaction event and the states of neighbor nodes of the newly added interaction event, so that information loss can be reduced. In addition, in the scheme, the message aggregation operation and the traffic prediction operation can be executed in parallel. For example, updating the state of the neighbor node and predicting the event category of the newly-added interaction event are executed in parallel, so that the service prediction efficiency can be greatly improved.
Finally, the embodiments of the present disclosure provide a new training method for the above-mentioned flow dynamics neural network, and the following description is directed to the training method.
Fig. 6 is a flowchart of a method for training a flow dynamics neural network provided in the present specification. As shown in fig. 6, the method may include:
step 602, obtaining an existing dynamic interaction graph constructed based on a plurality of historical interaction events.
Including a plurality of nodes representing a plurality of objects, each node corresponding to an implicit vector. Two nodes representing two objects in the same historical interaction event are connected through a connecting edge, and the connecting edge indicates the interaction time of the same historical interaction event.
Step 604, a sample interaction event is obtained.
The sample interaction event has a corresponding sample label.
Here, the updating step may refer to step 416, and the description is not repeated here.
And step 610, at least fusing the node information of the two target nodes to obtain event interaction information.
The node information of the target node may be represented as an embedded vector u, which is used to predict the event class of the sample interaction event described above.
Step 612, respectively updating the first updated implicit vectors of the two target nodes again according to the target time difference between the interaction time of the recent historical interaction event where the two target nodes are respectively located and the interaction time of the sample interaction event, and the event interaction information.
And 614, predicting the event category of the sample interaction event by using the classifier based on the updated implicit vectors of the two target nodes respectively.
Step 616, determining the predicted loss according to the sample label and the event category.
And step 618, jointly training the multi-layer perceptron and the classifier according to the prediction loss.
Of course, when step 612 is executed based on the neural network model, step 618 may be replaced with: and jointly training the multilayer perceptron, the neural network model and the classifier according to the prediction loss.
Therefore, in the training method of the flow dynamic graph neural network, the original flow node state updating process is modified, that is, the execution sequence of the message aggregation is restored. Therefore, each module in the neural network of the flow dynamics diagram can be effectively trained.
Corresponding to the method for processing an interactivity event described above, an embodiment of the present specification further provides an apparatus for processing an interactivity event, as shown in fig. 7, the apparatus may include:
an obtaining unit 702, configured to obtain an existing dynamic interaction graph constructed based on a plurality of historical interaction events. Including a plurality of nodes representing a plurality of objects, each node corresponding to an implicit vector. Two nodes representing two objects in the same historical interaction event are connected through a connecting edge, and the connecting edge indicates the interaction time of the same historical interaction event.
The obtaining unit 702 is further configured to obtain a new interaction event.
A determining unit 704, configured to determine two target nodes corresponding to the newly added interaction event in the existing dynamic interaction graph, and establish a newly added connection edge between the two target nodes.
And a fusion unit 706, configured to fuse node information of at least two target nodes to obtain event interaction information. The node information of each target node therein is determined based on at least its implicit vector.
Optionally, the newly added interaction event may have an event feature. When the newly added interaction event is a transaction event, the event characteristics may include at least one of the following: transaction type, transaction amount, and transaction channel.
The fusion unit 706 may specifically be configured to: and fusing the node information of the two target nodes and the event characteristics to obtain event interaction information.
The updating unit 708 is configured to update the implicit vectors of the two target nodes respectively according to target time differences between the interaction time of the recent historical interaction event in which the two target nodes are respectively located and the interaction time of the newly added interaction event, and the event interaction information.
Here, the two target nodes may include a first target node, and the updating unit 708 may specifically be configured to:
a target time difference corresponding to the first target node is mapped to a target vector having a predetermined dimension using a first mapping algorithm.
And fusing the target vector and the implicit vector of the first target node to obtain an intermediate vector.
And inputting the intermediate vector and the event interaction information into the neural network model to obtain an updated implicit vector of the first target node.
The determining unit 704 is further configured to determine respective neighbor nodes of each order of the two target nodes.
Wherein, each order neighbor node of the arbitrary first target node includes:
starting from the first target node, and reaching each node through a connecting edge within a preset number N; and/or starting from the first target node, and reaching each node through the connecting edge and with the interaction time within a preset time range.
The determining unit 704 is further configured to determine, for any first neighboring node in the neighboring nodes of each order, propagation information corresponding to the first neighboring node based on a first time difference between an interaction time of a recent historical interaction event in which the first neighboring node is located and an interaction time of a newly added interaction event, a first distance between the first neighboring node and a corresponding target node, and respective updated implicit vectors of the two target nodes.
The determining unit 704 may specifically be configured to:
the first time difference is mapped into a first vector having a predetermined dimension using a first mapping algorithm.
The first distances are mapped to a second vector having a predetermined dimension using a second mapping algorithm.
And fusing the first vector, the second vector and the updated implicit vectors of the two target nodes to obtain the propagation information corresponding to the first neighbor node.
The updating unit 708 is further configured to update the implicit vector of the first neighboring node according to the propagation information.
The updating unit 708 may specifically be configured to:
and integrating the transmission information and the implicit vector of the first neighbor node by adopting an attention mechanism to obtain an integration result.
And updating the implicit vector of the first neighbor node by utilizing the multilayer perceptron based on the integration result.
Optionally, the newly added interaction event may be a sample interaction event, and the sample interaction event has a corresponding sample tag.
The obtaining unit 702 is further configured to obtain historical propagation information received by each of the two target nodes, and update the implicit vectors of each of the two target nodes for the first time by using the multi-layer sensor based on the historical propagation information.
The updating unit 708 may further specifically be configured to:
and respectively updating the implicit vectors of the two target nodes which are updated for the first time.
Optionally, the apparatus may further include:
and a predicting unit 710, configured to predict event classes of the newly added interaction events by using the classifier based on the updated implicit vectors of the two target nodes, respectively.
The determining unit 704 is further configured to determine the predicted loss according to the sample label and the event category.
And a training unit 712, configured to jointly train the multi-layer perceptron and the classifier according to the predicted loss.
The predicting unit 710 is further configured to predict event classes of the newly added interaction events by using a classifier based on the updated implicit vectors of the two target nodes, respectively.
The prediction unit 710 may specifically be configured to:
and respectively activating the updated implicit vectors of the two target nodes by using an activation function to obtain two embedded vectors for representing the updated node information of the two target nodes.
And inputting the two embedded vectors into a classifier to obtain the event category of the newly added interaction event.
The functions of each functional module of the device in the above embodiments of the present description may be implemented through each step of the above method embodiments, and therefore, a specific working process of the device provided in one embodiment of the present description is not repeated herein.
The device for processing the interaction event, provided in one embodiment of the present specification, can update the feature expression of each node in the dynamic interaction graph in real time, and can further analyze the interaction event more comprehensively and accurately.
According to an embodiment of another aspect, there is also provided a computer-readable storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method described in connection with fig. 4 or fig. 6.
According to an embodiment of yet another aspect, there is also provided a computing device comprising a memory and a processor, the memory having stored therein executable code, the processor, when executing the executable code, implementing the method described in connection with fig. 4 or fig. 6.
The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for the apparatus embodiment, since it is substantially similar to the method embodiment, the description is relatively simple, and for the relevant points, reference may be made to the partial description of the method embodiment.
The steps of a method or algorithm described in connection with the disclosure herein may be embodied in hardware or may be embodied in software instructions executed by a processor. The software instructions may consist of corresponding software modules that may be stored in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such the processor can read information from, and write information to, the storage medium. Of course, the storage medium may also be integral to the processor. The processor and the storage medium may reside in an ASIC. Additionally, the ASIC may reside in a server. Of course, the processor and the storage medium may reside as discrete components in a server.
Those skilled in the art will recognize that, in one or more of the examples described above, the functions described in this invention may be implemented in hardware, software, firmware, or any combination thereof. When implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media includes both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage media may be any available media that can be accessed by a general purpose or special purpose computer.
The foregoing description has been directed to specific embodiments of this disclosure. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing may also be possible or may be advantageous.
The above-mentioned embodiments, objects, technical solutions and advantages of the present specification are further described in detail, it should be understood that the above-mentioned embodiments are only specific embodiments of the present specification, and are not intended to limit the scope of the present specification, and any modifications, equivalent substitutions, improvements and the like made on the basis of the technical solutions of the present specification should be included in the scope of the present specification.
Claims (24)
1. A computer-implemented method of processing an interaction event, comprising:
acquiring an existing dynamic interaction graph constructed based on a plurality of historical interaction events; the method comprises the steps of representing a plurality of nodes of a plurality of objects, wherein each node corresponds to an implicit vector; two nodes representing two objects in the same historical interaction event are connected through a connecting edge, and the connecting edge indicates the interaction time of the same historical interaction event;
acquiring a newly added interaction event;
determining two target nodes corresponding to the newly added interaction event in the existing dynamic interaction graph, and establishing a newly added connecting edge between the two target nodes;
at least fusing the node information of the two target nodes to obtain event interaction information; wherein the node information of each target node is determined based on at least its implicit vector;
respectively updating respective implicit vectors of the two target nodes according to respective target time differences of interaction time of the latest historical interaction events where the two target nodes are respectively located and interaction time of the newly added interaction events and the event interaction information;
respectively determining respective adjacent nodes of each order of the two target nodes;
aiming at any first neighbor node in the neighbor nodes of each order, determining propagation information corresponding to the first neighbor node based on a first time difference between the interaction time of the latest historical interaction event where the first neighbor node is located and the interaction time of the newly added interaction event, a first distance between the first neighbor node and a corresponding target node, and respective updated implicit vectors of the two target nodes;
and updating the implicit vector of the first neighbor node according to the propagation information.
2. The method of claim 1, wherein the respective order neighbor nodes of any first target node comprise:
starting from the first target node, and reaching each node through a connecting edge within a preset number N; and/or the presence of a gas in the gas,
and starting from the first target node, each node which can be reached through the connecting edge and has the interaction time within a preset time range.
3. The method of claim 1, the new interactivity event having an event characteristic; the fusion of the node information of at least two target nodes to obtain event interaction information comprises:
and fusing the node information of the two target nodes and the event characteristics to obtain the event interaction information.
4. The method of claim 3, the newly added interaction event being a transaction event, the event characteristics including at least one of: transaction type, transaction amount, and transaction channel.
5. The method of claim 1, the two target nodes comprising a first target node; the updating the implicit vectors of the two target nodes respectively according to the target time difference between the interaction time of the recent historical interaction event in which the two target nodes are respectively located and the interaction time of the newly added interaction event and the event interaction information comprises:
mapping a target time difference corresponding to the first target node into a target vector with a predetermined dimension by using a first mapping algorithm;
fusing the target vector and the implicit vector of the first target node to obtain an intermediate vector;
and inputting the intermediate vector and the event interaction information into a neural network model to obtain an updated implicit vector of the first target node.
6. The method of claim 1, wherein the determining the propagation information corresponding to the first neighbor node based on a first time difference between an interaction time of a recent historical interaction event in which the first neighbor node is located and an interaction time of the newly added interaction event, a first distance between the first neighbor node and a corresponding target node, and respective updated implicit vectors of the two target nodes comprises:
mapping the first time difference into a first vector with a predetermined dimension by using a first mapping algorithm;
mapping the first distance into a second vector with a predetermined dimension by using a second mapping algorithm;
and fusing the first vector, the second vector and the updated implicit vectors of the two target nodes to obtain the propagation information corresponding to the first neighbor node.
7. The method of claim 1, the updating an implicit vector for the first neighbor node according to the propagation information, comprising:
integrating the propagation information and the implicit vector of the first neighbor node by adopting an attention mechanism to obtain an integration result;
updating the implicit vector of the first neighbor node by utilizing a multilayer perceptron based on the integration result.
8. The method of claim 7, wherein the newly added interaction event is a sample interaction event, and the sample interaction event has a corresponding sample tag;
before the fusion of the node information of at least two target nodes, the method further includes:
respectively acquiring historical propagation information received by the two target nodes, and respectively updating the implicit vectors of the two target nodes for the first time by utilizing the multilayer perceptron based on the historical propagation information;
the respectively updating the implicit vectors of the two target nodes comprises:
and respectively updating the first updated implicit vectors of the two target nodes again.
9. The method of claim 8, further comprising:
predicting the event category of the newly added interaction event by using a classifier based on the updated implicit vectors of the two target nodes respectively;
determining a predicted loss according to the sample label and the event category;
jointly training the multi-layered perceptron and the classifier according to the prediction loss.
10. The method of claim 1, further comprising:
and predicting the event category of the newly added interaction event by utilizing a classifier based on the updated implicit vectors of the two target nodes.
11. The method of claim 10, wherein predicting the event class of the newly added interactivity event based on the updated implicit vectors of the two target nodes using a classifier comprises:
respectively activating the updated implicit vectors of the two target nodes by using an activation function to obtain two embedded vectors used for representing the updated node information of the two target nodes;
and inputting the two embedded vectors into the classifier to obtain the event category of the newly added interaction event.
12. An apparatus for processing an interaction event, comprising:
the acquisition unit is used for acquiring an existing dynamic interaction diagram constructed based on a plurality of historical interaction events; the method comprises the steps of representing a plurality of nodes of a plurality of objects, wherein each node corresponds to an implicit vector; two nodes representing two objects in the same historical interaction event are connected through a connecting edge, and the connecting edge indicates the interaction time of the same historical interaction event;
the acquiring unit is further used for acquiring a newly added interaction event;
a determining unit, configured to determine two target nodes corresponding to the newly added interaction event in the existing dynamic interaction graph, and establish a newly added connection edge between the two target nodes;
the fusion unit is used for fusing at least the node information of the two target nodes to obtain event interaction information; wherein the node information of each target node is determined based on at least its implicit vector;
the updating unit is used for respectively updating the implicit vectors of the two target nodes according to the target time difference between the interaction time of the recent historical interaction event where the two target nodes are respectively located and the interaction time of the newly added interaction event and the event interaction information;
the determining unit is further configured to determine respective neighboring nodes of each order of the two target nodes respectively;
the determining unit is further configured to determine, for any first neighbor node in the neighbor nodes of each order, propagation information corresponding to the first neighbor node based on a first time difference between an interaction time of a recent historical interaction event where the first neighbor node is located and an interaction time of the newly added interaction event, a first distance between the first neighbor node and a corresponding target node, and respective updated implicit vectors of the two target nodes;
the updating unit is further configured to update the implicit vector of the first neighbor node according to the propagation information.
13. The apparatus of claim 12, wherein the respective order neighbor nodes of any first target node comprise:
starting from the first target node, and reaching each node through a connecting edge within a preset number N; and/or the presence of a gas in the gas,
and starting from the first target node, each node which can be reached through the connecting edge and has the interaction time within a preset time range.
14. The apparatus of claim 12, the new interactivity event having an event characteristic; the fusion unit is specifically configured to:
and fusing the node information of the two target nodes and the event characteristics to obtain the event interaction information.
15. The apparatus of claim 14, the newly added interaction event being a transaction event, the event characteristics including at least one of: transaction type, transaction amount, and transaction channel.
16. The apparatus of claim 12, the two target nodes comprising a first target node; the update unit is specifically configured to:
mapping a target time difference corresponding to the first target node into a target vector with a predetermined dimension by using a first mapping algorithm;
fusing the target vector and the implicit vector of the first target node to obtain an intermediate vector;
and inputting the intermediate vector and the event interaction information into a neural network model to obtain an updated implicit vector of the first target node.
17. The apparatus according to claim 12, wherein the determining unit is specifically configured to:
mapping the first time difference into a first vector with a predetermined dimension by using a first mapping algorithm;
mapping the first distance into a second vector with a predetermined dimension by using a second mapping algorithm;
and fusing the first vector, the second vector and the updated implicit vectors of the two target nodes to obtain the propagation information corresponding to the first neighbor node.
18. The apparatus according to claim 12, wherein the updating unit is further specifically configured to:
integrating the propagation information and the implicit vector of the first neighbor node by adopting an attention mechanism to obtain an integration result;
updating the implicit vector of the first neighbor node by utilizing a multilayer perceptron based on the integration result.
19. The apparatus of claim 18, the newly added interactivity event being a sample interactivity event, the sample interactivity event having a corresponding sample tag;
the obtaining unit is further configured to obtain historical propagation information received by each of the two target nodes, and based on the historical propagation information, update the implicit vectors of each of the two target nodes by using the multilayer perceptron for the first time;
the update unit is specifically configured to:
and respectively updating the first updated implicit vectors of the two target nodes again.
20. The apparatus of claim 19, further comprising:
the prediction unit is used for predicting the event type of the newly added interaction event by utilizing a classifier based on the respectively updated implicit vectors of the two target nodes;
the determining unit is further configured to determine a predicted loss according to the sample label and the event category;
and the training unit is used for jointly training the multilayer perceptron and the classifier according to the prediction loss.
21. The apparatus of claim 12, further comprising:
and the predicting unit is used for predicting the event type of the newly added interaction event by utilizing a classifier based on the updated implicit vectors of the two target nodes.
22. The apparatus of claim 21, the prediction unit to:
respectively activating the updated implicit vectors of the two target nodes by using an activation function to obtain two embedded vectors used for representing the updated node information of the two target nodes;
and inputting the two embedded vectors into the classifier to obtain the event category of the newly added interaction event.
23. A computer-readable storage medium, on which a computer program is stored, wherein the computer program, when executed in a computer, causes the computer to perform the method of any of claims 1-11.
24. A computing device comprising a memory and a processor, wherein the memory has stored therein executable code that when executed by the processor implements the method of any of claims 1-11.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011409575.0A CN112541129B (en) | 2020-12-06 | 2020-12-06 | Method and device for processing interaction event |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011409575.0A CN112541129B (en) | 2020-12-06 | 2020-12-06 | Method and device for processing interaction event |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112541129A true CN112541129A (en) | 2021-03-23 |
| CN112541129B CN112541129B (en) | 2023-05-23 |
Family
ID=75016006
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011409575.0A Active CN112541129B (en) | 2020-12-06 | 2020-12-06 | Method and device for processing interaction event |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112541129B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113987280A (en) * | 2021-10-27 | 2022-01-28 | 支付宝(杭州)信息技术有限公司 | Method and device for training graph model aiming at dynamic graph |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110598847A (en) * | 2019-08-15 | 2019-12-20 | 阿里巴巴集团控股有限公司 | Method and device for processing interactive sequence data |
| CN110689110A (en) * | 2019-08-28 | 2020-01-14 | 阿里巴巴集团控股有限公司 | Method and device for processing interaction event |
| US20200327394A1 (en) * | 2019-04-12 | 2020-10-15 | Optum, Inc. | Contextually optimizing routings for interactions |
| CN111949892A (en) * | 2020-08-10 | 2020-11-17 | 浙江大学 | Multi-relation perception temporal interaction network prediction method |
-
2020
- 2020-12-06 CN CN202011409575.0A patent/CN112541129B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200327394A1 (en) * | 2019-04-12 | 2020-10-15 | Optum, Inc. | Contextually optimizing routings for interactions |
| CN110598847A (en) * | 2019-08-15 | 2019-12-20 | 阿里巴巴集团控股有限公司 | Method and device for processing interactive sequence data |
| CN110689110A (en) * | 2019-08-28 | 2020-01-14 | 阿里巴巴集团控股有限公司 | Method and device for processing interaction event |
| CN111949892A (en) * | 2020-08-10 | 2020-11-17 | 浙江大学 | Multi-relation perception temporal interaction network prediction method |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113987280A (en) * | 2021-10-27 | 2022-01-28 | 支付宝(杭州)信息技术有限公司 | Method and device for training graph model aiming at dynamic graph |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112541129B (en) | 2023-05-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US10565527B2 (en) | Predicting the next application that you are going to use on aviate | |
| CN110598847B (en) | Method and device for processing interactive sequence data | |
| Cai et al. | A noise-immune LSTM network for short-term traffic flow forecasting | |
| WO2019114423A1 (en) | Method and apparatus for merging model prediction values, and device | |
| TW202008264A (en) | Method and apparatus for recommendation marketing via deep reinforcement learning | |
| CN111210008B (en) | Method and device for processing interactive data by using LSTM neural network model | |
| Chen et al. | Exploiting Web service geographical neighborhood for collaborative QoS prediction | |
| US10728104B2 (en) | Time-dependent network embedding | |
| US10970350B2 (en) | Method and apparatus for processing user interaction sequence data | |
| CN110909942B (en) | Method and system for training model and method and system for predicting sequence data | |
| CN112085293B (en) | Method and device for training interactive prediction model and predicting interactive object | |
| Pereira et al. | A metamodel for estimating error bounds in real-time traffic prediction systems | |
| US11621892B2 (en) | Temporal-based network embedding and prediction | |
| CN110689110B (en) | Method and device for processing interaction event | |
| Fathalla et al. | An LSTM-based distributed scheme for data transmission reduction of IoT systems | |
| CN116627630A (en) | Resource demand prediction model training, demand prediction and resource scheduling method and system | |
| CN112541129A (en) | Method and device for processing interaction event | |
| da Silva et al. | Online machine learning for auto-scaling in the edge computing | |
| Hayashi et al. | Sequential network change detection with its applications to ad impact relation analysis | |
| CN113449176A (en) | Recommendation method and device based on knowledge graph | |
| CN112085279B (en) | Method and device for training interactive prediction model and predicting interactive event | |
| CN115860856A (en) | Data processing method and device, electronic equipment and storage medium | |
| Xu et al. | A meta-path-based recurrent model for next poi prediction with spatial and temporal contexts | |
| CN114510627A (en) | Object pushing method and device, electronic equipment and storage medium | |
| CN111382846B (en) | Method and device for training neural network model based on transfer learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| REG | Reference to a national code |
Ref country code: HK Ref legal event code: DE Ref document number: 40048341 Country of ref document: HK |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |