CN114331500A

CN114331500A - Click rate prediction method, device, equipment, storage medium and program product

Info

Publication number: CN114331500A
Application number: CN202111500222.6A
Authority: CN
Inventors: 陈昊
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2021-12-09
Filing date: 2021-12-09
Publication date: 2022-04-12

Abstract

The application discloses a click rate prediction method, a click rate prediction device, click rate prediction equipment, a click rate storage medium and a click rate program product, and belongs to the field of artificial intelligence. The method comprises the following steps: determining N clicked with target user account₁M with click relation in first-order commodities₂A second order user account; m for determining and clicking target commodity₁N with click relation in first-order user account₂A second-order commodity; based on N₁A first order commodity and M₂Obtaining a first airspace characteristic by the second-order user account; based on M₁First order user account number and N₂Obtaining second-order commodities to obtain second airspace characteristics; based on N₁A first order commodity and M₂Clicking the time stamp of each second-order user account to obtain a first time domain characteristic; based on M₁First order user account number and N₂The click time stamp of each second-order commodity is obtained to obtain a second time domain feature; and predicting the click rate based on the first spatial domain feature, the second spatial domain feature, the first time domain feature and the second time domain feature. The method improves the accuracy of the click rate.

Description

Click rate prediction method, device, equipment, storage medium and program product

Technical Field

The present application relates to the field of artificial intelligence, and in particular, to a method, an apparatus, a device, a storage medium, and a program product for predicting a click rate.

Background

With the development of the internet e-commerce platform, whether a decision mode for pushing a target commodity for a target user is frequently available or not is judged based on the click rate of the target user on the target commodity, and the click rate is the probability of the target user clicking the target commodity.

In the related art, the target user click rate on the target commodity is predicted according to feature intersection between the target user and each feature domain of the target commodity. For example, the target user's probability of clicking on an advertisement for a luxury (which is characterized by a high price and a low volume) is predicted based on the target user's age and occupation, and a decision is made whether to push the advertisement to the target user based on the predicted click-through rate.

However, in the related technology, only the target user and each feature domain of the target commodity are relied on for prediction, the click rate obtained through prediction is not accurate, and the method for predicting the click rate is more accurate.

Disclosure of Invention

The application provides a click rate prediction method, a click rate prediction device, equipment, a storage medium and a program product, which can improve the accuracy of the predicted click rate. The technical scheme is as follows:

according to an aspect of the present application, there is provided a method for predicting a click rate, the method including:

determining N clicked with target user account₁M with click relation among first-order commodities₂A second order user account; and determining and clicking M of the target commodity₁N with click relation between first-order user accounts₂A second-order commodity; n is a radical of₁、M₂、M₁And N₂Are all positive integers;

based on N₁Features of a first order item and M₂Obtaining a first airspace characteristic by the characteristics of the second-order user accounts; and based on M₁Characteristics of first order user account and N₂Obtaining a second airspace characteristic according to the characteristics of the second-order commodities;

based on N₁A first order commodity and M₂Points of second order user accountClicking the timestamp to obtain a first time domain characteristic; based on M₁First order user account number and N₂The click time stamp of each second-order commodity is used for obtaining a second time domain characteristic;

and predicting the probability of the target user account clicking the target commodity based on the first spatial domain feature, the second spatial domain feature, the first time domain feature and the second time domain feature.

According to an aspect of the present application, there is provided a click rate prediction apparatus, including:

a determining module for determining N clicked with the target user account₁M with click relation among first-order commodities₂A second order user account; and determining and clicking M of the target commodity₁N with click relation between first-order user accounts₂A second-order commodity; n is a radical of₁、M₂、M₁And N₂Are all positive integers;

a processing module for N-based₁Features of a first order item and M₂Obtaining a first airspace characteristic by the characteristics of the second-order user accounts; and based on M₁Characteristics of first order user account and N₂Obtaining a second airspace characteristic according to the characteristics of the second-order commodities;

a processing module for further processing based on N₁A first order commodity and M₂The second-order user accounts have click time stamps to obtain first time domain characteristics; based on M₁First order user account number and N₂The click time stamp of each second-order commodity is used for obtaining a second time domain characteristic;

and the prediction module is used for predicting the probability of the target commodity clicked by the target user account based on the first spatial domain feature, the second spatial domain feature, the first time domain feature and the second time domain feature.

In an optional embodiment, the determining module is further configured to determine N clicked with the target user account according to the target user node corresponding to the target user account₁N corresponding to first-order commodities₁A first order commodity node, and N₁M with click relation among first-order commodities₂M corresponding to second-order user account₂And constructing a first interactive subgraph of the target user side by taking the click timestamp as an edge weight between the target user node and the first-order commodity node and taking the click timestamp as an edge weight between the first-order commodity node and the second-order user node.

In an optional embodiment, the determining module is further configured to determine, according to the target product node corresponding to the target product, M clicked on the target product₁M corresponding to first-order user account₁A first order user node, and M₁N with click relation between first-order user accounts₂N corresponding to second-order commodities₂And constructing a second interactive subgraph of the target commodity side by taking the click timestamp as the edge weight between the target commodity node and the first-order user node and taking the click timestamp as the edge weight between the first-order user node and the second-order commodity node.

In an optional embodiment, the processing module is further configured to assign N to the first graph neural network₁Features of a first order commodity node and M₂And transmitting the characteristics of the second-order user nodes to the target user node, and aggregating the characteristics with the characteristics of the target user node to obtain first airspace characteristics.

In an alternative embodiment, the processing module is further configured to apply M to the neural network of the second graph₁Characteristics of a first order user node and N₂And transmitting the characteristics of the second-order commodity nodes to the target commodity nodes, and aggregating the characteristics of the second-order commodity nodes with the characteristics of the target commodity nodes to obtain second airspace characteristics.

In an optional embodiment, the processing module is further configured to perform two aggregation processes through the first graph neural network, and use a feature vector of the target user node obtained through the second aggregation as a first spatial domain feature vector;

during the k-th polymerization:

will N₁N of first order commodity node₁Carrying out dimension-by-dimension mean value aggregation on the k-1 th characteristic vector to obtain a k-first-order commodity characteristic vector, and then carrying out feature vector aggregation on the k-first-order commodity characteristic vector and the k-1 th characteristic direction of the target user nodeSplicing the quantities to obtain the kth characteristic vector of the target user node;

and, for N₁The p-th one of the first-order commodity nodes, and M_2pM of second-order user nodes_2pCarrying out dimensionality-by-dimensionality mean value aggregation on the k-1 th feature vectors to obtain k-th second-order user feature vectors, splicing the k-th second-order user feature vectors with the k-1 th feature vectors of the first-order commodity nodes to obtain the k-th feature vectors of the first-order commodity nodes, wherein the value of k is 1 or 2, and the value of p is 1-N₁；

Wherein the content of the first and second substances,

in an optional embodiment, the processing module is further configured to perform two times of aggregation processes through the second graph neural network, and use a feature vector of the target commodity node obtained through the second aggregation as a second spatial domain feature vector;

during the j-th polymerization:

will M₁M of first order user nodes₁The j-1 th feature vectors are subjected to dimension-by-dimension mean aggregation to obtain a j-1 th-order user feature vector, and then the j-1 th-order user feature vector and the j-1 th feature vector of the target commodity node are spliced to obtain a j-th feature vector of the target commodity node;

and, for M₁The q-th one of the first order user nodes, N_2qN of second-order commodity nodes_2qThe j-1 th feature vectors are subjected to dimensionality mean value aggregation to obtain a j second-order commodity feature vector, then the j second-order commodity feature vector is spliced with the j-1 th feature vector of the first-order user node to obtain a j th feature vector of the first-order user node, the value of j is 1 or 2, and the value of q is 1-M₁；

Wherein the content of the first and second substances,

in an optional embodiment, the processing module is further configured to sort N sorted by click timestamp₁One-step commodity festivalInputting the initial characteristics of the points into a first recurrent neural network to obtain the sum N₁A first time domain sub-feature corresponding to the first-order commodity node; m sorted by click timestamp₂Inputting the initial characteristics of the second-order user nodes into a first recurrent neural network to obtain the sum M₂A first time domain sub-feature corresponding to each second-order user node; and obtaining the first time domain characteristic by aggregating the two first time domain sub-characteristics.

In an optional embodiment, the processing module is further configured to sort the M sorted by the click timestamp₁Inputting the initial characteristics of the first-order user nodes into a second recurrent neural network to obtain the sum M₁A second time domain sub-feature corresponding to the first-order user node; n sorted by click timestamp₂Inputting the initial characteristics of the second-order commodity nodes into a second recurrent neural network to obtain the sum N₂A second time domain sub-feature corresponding to each second-order commodity node; and obtaining a second time domain characteristic by aggregating the two second time domain sub-characteristics.

In an alternative embodiment, the recurrent neural network is an episodic memory network LSTM.

In an optional embodiment, the processing module is further configured to sort N sorted by click timestamp₁Inputting the initial characteristic vector of each first-order commodity node into a long-time and short-time memory network, and taking the characteristic vector after the last first-order commodity node is updated as N₁And the first-order commodity node corresponds to a first time domain sub-feature vector.

In an optional embodiment, the processing module is further configured to sort the M sorted by the click timestamp₂Inputting initial feature vector of each second-order user node into a long-time and short-time memory network, and taking the feature vector updated by the last second-order user node as M₂And the first time domain sub-feature vectors correspond to the second-order user nodes.

In an optional embodiment, the processing module is further configured to splice the two first time-domain sub-feature vectors to obtain a first time-domain feature vector.

In an alternative embodiment, the second recurrent neural network is an long-term memory network LSTM.

In an optional embodiment, the processing module is further configured to sort the M sorted by the click timestamp₁Inputting the initial characteristic vector of each first-order user node into a long-time and short-time memory network, and taking the updated characteristic vector of the last first-order user node as M₁And the second time domain sub-feature vector corresponding to the first-order user node.

In an optional embodiment, the processing module is further configured to sort N sorted by click timestamp₂Inputting the initial characteristic vector of each second-order commodity node into a long-time and short-time memory network, and taking the characteristic vector after the last second-order commodity node is updated as N₂And the second time domain sub-feature vectors correspond to the second-order commodity nodes.

In an optional embodiment, the processing module is further configured to splice the two second time-domain sub-feature vectors to obtain a second time-domain feature vector.

In an optional embodiment, the initial features of the first-order commodity nodes are obtained by aggregating the features and the time features of the first-order commodity nodes; the initial characteristics of the second-order user nodes are obtained by aggregating the characteristics and the time characteristics of the second-order user nodes.

In an optional embodiment, the initial features of the first-order user nodes are obtained by aggregating the features and the time features of the first-order user nodes; the initial characteristics of the second-order commodity nodes are obtained by aggregating the characteristics and the time characteristics of the second-order commodity nodes.

In an optional embodiment, the prediction module is further configured to splice the first space domain feature vector and the first time domain feature vector to obtain a feature vector of the target user side; and splicing the second space domain feature vector and the second time domain feature vector to obtain the feature vector of the target commodity side.

In an optional embodiment, the prediction module is further configured to splice the feature vector of the target user side and the feature vector of the target commodity side to obtain an intermediate feature vector.

In an optional embodiment, the prediction module is further configured to predict the probability that the target user account clicks the target commodity through the multi-layer perceptron.

In an optional embodiment, the determining module is further configured to construct an interactive bipartite graph, where the interactive bipartite graph includes a plurality of user nodes and a plurality of commodity nodes, connect edges between the user nodes and the commodity nodes when a click relationship exists between the user nodes and the commodity nodes, and set an edge right between the user nodes and the commodity nodes as a click timestamp.

In an optional embodiment, the determining module is further configured to determine a target user node, a plurality of first-order candidate commodity nodes connected to the target user node, and a plurality of second-order candidate user nodes connected to the plurality of first-order candidate commodity nodes; a plurality of first-order candidate commodity nodes are sorted and sampled according to the click time stamps to obtain N₁A first-order commodity node; a plurality of second-order candidate user nodes are sequenced and sampled according to click time stamps to obtain M₂And the second-order user nodes.

In an optional embodiment, the determining module is further configured to determine a target commodity node, a plurality of first-order candidate user nodes connected to the target commodity node, and a plurality of second-order candidate commodity nodes connected to the plurality of first-order candidate user nodes; a plurality of first-order candidate user nodes are sequenced and sampled according to click time stamps to obtain M₁A first-order user node; a plurality of second-order candidate commodity nodes are sequenced and sampled according to the click time stamps to obtain N₂And each second-order commodity node.

According to an aspect of the present application, there is provided a computer device including: a processor and a memory, the memory storing a computer program that is loaded and executed by the processor to implement the click-through rate prediction method as described above.

According to another aspect of the present application, there is provided a computer-readable storage medium storing a computer program loaded and executed by a processor to implement the prediction method of click-through rate as described above.

According to another aspect of the application, a computer program product or computer program is provided, comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to execute the method for predicting the click rate.

The beneficial effects brought by the technical scheme provided by the embodiment of the application at least comprise:

by mining the target user side (target user account number, N)₁A first order commodity and M₁First order user account) and target commodity side (target commodity, M)₁First order user account number and N₂Second-order merchandise), not only explicit information (information between the target user account and the first-order merchandise, information between the target merchandise and the first-order user account), but also implicit information (information between the first-order merchandise and the second-order user account, information between the first-order user account and the second-order merchandise) is focused on from a global perspective. And a symmetrical information attention mode is adopted, so that the click rate is prevented from being predicted one-sidedly, and the effectiveness and the reliability of the predicted click rate are improved.

And the predicted click rate is more accurate by fully paying attention to the historical click information between the commodity and the user.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, 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 application, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.

FIG. 1 is a schematic diagram of a computer system provided in an exemplary embodiment of the present application;

FIG. 2 is a flow chart of a method for predicting click through rate provided by an exemplary embodiment of the present application;

FIG. 3 is a flowchart of a method for constructing a first interactive subgraph and a second interactive subgraph provided by an exemplary embodiment of the present application;

FIG. 4 is a schematic diagram of an interaction bipartite graph provided by an exemplary embodiment of the present application;

FIG. 5 is a schematic diagram of a first interaction sub-graph provided by an exemplary embodiment of the present application;

FIG. 6 is a schematic diagram of a second interaction sub-graph provided by an exemplary embodiment of the present application;

FIG. 7 is a flow chart of a method of computing a first spatial signature and a second spatial signature provided by an exemplary embodiment of the present application;

FIG. 8 is a flow chart of a method of computing a first time-domain feature provided by an exemplary embodiment of the present application;

FIG. 9 is a schematic diagram of an aggregation process of two first time domain sub-features provided by an exemplary embodiment of the present application;

FIG. 10 is a flow chart of a method of computing a second time-domain feature provided by an exemplary embodiment of the present application;

FIG. 11 is a schematic diagram of an aggregation process of two second time-domain sub-features provided by an exemplary embodiment of the present application;

FIG. 12 is a flow chart of a method for predicting click-through rate as provided by another exemplary embodiment of the present application;

FIG. 13 is a block diagram of a click rate prediction device according to an exemplary embodiment of the present application;

fig. 14 is a block diagram of a computer device according to an exemplary embodiment of the present application.

Detailed Description

To make the objects, technical solutions and advantages of the present application more clear, embodiments of the present application will be described in further detail below with reference to the accompanying drawings.

First, terms referred to in the embodiments of the present application are briefly described:

graph network: in the embodiment of the present application, a graph network refers to data stored in the form of a graph, and is also referred to as graph data, a graph model, a graph representation, and graph structure data. The graph network comprises at least one node and at least one edge, each node is provided with corresponding characteristics, the edge is used for representing the connection relation between different nodes, and each edge is also provided with edge weight for representing the connection information between different nodes.

Graph neural network: the method refers to a general name of a model applied by a neural network on a graph network, wherein the graph neural network comprises a graph convolution neural network, a graph attention network and the like. The graph neural network is used for predicting the class of the graph according to the structural features of the graph. In particular, the graph neural network may include one or more feature extraction layers. The feature extraction layer is, for example, Graph Convolution Layers (GCL). The feature extraction layer is used for extracting structural features of the graph. For example, if the two graphs are isomorphic, then the structural features of the graphs of the two graphs after passing through the feature extraction layer will be similar. If the two graphs are heterogeneous, the structural features of the graphs of the two graphs will be different after passing through the feature extraction layer. Thus, the graph neural network is able to map graph structures with homogeneous properties into the same representation domain and output the same classes.

RNN (current Neural Networks, Recurrent Neural Networks): RNNs are a class of neural networks used to process sequence data in which the output of a current position of a sequence is also related to the output of a previous position. RNN relies on a neural network with fixed weights, external inputs and internal states, which can be viewed as behavioural dynamics about the internal states with the weights and external inputs as parameters.

LSTM (Long Short Term Memory, Long duration Memory network): LSTM is a time-recursive neural network for processing and predicting significant events in a time series that are spaced or delayed for a relatively long time, a special RNN.

FIG. 1 illustrates a schematic diagram of a computer system provided by an exemplary embodiment of the present application. As shown in fig. 1, a predicted network of click through rates (with the click through rate prediction method provided in an exemplary embodiment of the present application running) is obtained by a training device 101, and the predicted network of click through rates is sent to a user device 102, where the user device 102 can use the predicted network of click through rates.

The training device 101 and the using device 102 may be computer devices with machine learning capability, for example, the computer devices may be terminals or servers.

Optionally, the training device 101 and the using device 102 may be the same computer device, or the training device 101 and the using device 102 may be different computer devices. Also, when training device 101 and using device 102 are different devices, training device 101 and using device 102 may be the same type of device, such as training device 101 and using device 102 may both be servers; alternatively, training device 101 and user device 102 may be different types of devices, such as training device 101 being a server and user device 102 being a terminal. The server may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing basic cloud computing services such as a cloud service, a cloud database, cloud computing, a cloud function, cloud storage, a Network service, cloud communication, a middleware service, a domain name service, a security service, a CDN (Content Delivery Network), a big data and artificial intelligence platform. The terminal may be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart television, a vehicle-mounted terminal, a wearable device, a smart sound box, and the like, but is not limited thereto. The terminal and the server may be directly or indirectly connected through wired or wireless communication, and the application is not limited herein.

In the related art of the application, two possible prediction modes exist for predicting the click rate of the target user account for clicking the target commodity.

The first prediction mode: and performing cross fusion to predict according to the target user account and each characteristic domain of the target commodity, wherein the prediction mode is simple and has a certain effect, but more characteristic domain information is needed and only explicit information existing between the target user account and the target commodity is mined.

The second prediction method: and predicting the probability of the target commodity clicked by the target user account according to the interest information of the target user account contained in the historical behavior sequence of the target user account. The method implicitly includes historical click information for the target user account, but the historical click information is still underutilized.

Compared with the two prediction modes, the click rate prediction method provided by the application is more accurate in result and better in effect.

To improve the accuracy of the predicted click through rate, fig. 2 shows a flowchart of a click through rate prediction method provided by an exemplary embodiment of the present application, which is exemplified by applying the method to the user device 102 shown in fig. 1, and the method includes:

step 220, determine the N clicked with the target user account₁M with click relation among first-order commodities₂A second order user account; and determining and clicking M of the target commodity₁N with click relation between first-order user accounts₂A second-order commodity;

wherein N is₁、M₂、M₁And N₂Are all positive integers.

The target user account number is as follows: in the present application, the click rate refers to the probability that the target user account clicks the target commodity, that is, the target user account is the target account of the present application. N clicked by target user account₁The first-order commodities are the commodity clusters clicked by the target user account in the first preset historical time period. And N₁M with click relation among first-order commodities₂Each second-order user account is that N clicks within a second preset historical time period₁A second-order user account cluster for the first-order commodity.

Optionally, the first preset historical time period is the same as the second preset historical time period. At this time, M₂The second-order user accounts include the target user account.

Illustratively, the "user account 11" (target user account) has clicked through 3 items of merchandise (N) such as a cup, keyboard, and mouse within the last seven days₁First order merchandise), wherein the cup was clicked (M) by 4 user accounts, user account 11, user account 12, user account 13, and user account 14, within the last seven days₂₁A second-order user account), the keyboard is atThe last seven days are clicked by 3 user accounts including the user account 21 (which is the same account as the user account 11), the user account 22 and the user account 23 (M)₂₂Second order user account), the mouse was clicked on by 2 user accounts (M) in the past five days (user account 31, which is the same account as user account 11), user account 32₂₃Second order user account number), M₂₁Second order user account number, M₂₂Second order user account number and M₂₃A second-order user account forms M₂A second order user account.

Target goods: in the present application, the click rate refers to the probability that the target user account clicks the target commodity, that is, the target commodity is the target commodity of the present application. M clicked through target commodity₁And the first-order user accounts are first-order user account clusters of the target commodity clicked in a third preset historical time period. And M₁N with click relation between first-order user accounts₂Each second-order commodity, namely, the commodity is processed by M within a fourth preset historical time period₁And the second-order commodity cluster clicked by the first-order user account.

Optionally, the third preset history time period is the same as the fourth preset history time period. At this time, N₂Each second-order item contains a target item.

Illustratively, the notebook computer (target commodity) is divided into 2 user accounts (M) by the user account 01 and the user account 02 in seven days₁First order user account), user account 01 has clicked on two items (N) of a notebook computer and a pen within the last seven days₂₁Second-order merchandise), the user account 02 has clicked on two merchandise (N) including a laptop and a chair within the last seven days₂₂Second order commodity), N₂₁Second order commodity and N₂₂A second-order commodity constitutes N₂A second-order commodity.

It will be appreciated that at N₁The first-order item may or may not contain the target item, M₁The first order user account may or may not include the target user account.

For a detailed description of step 220, please refer to "for step 220" below.

Step 240, based on N₁Features of a first order item and M₂Obtaining a first airspace characteristic by the characteristics of the second-order user accounts; and based on M₁Characteristics of first order user account and N₂Obtaining a second airspace characteristic according to the characteristics of the second-order commodities;

characteristics of the user account: optionally, the characteristics of the user account can be obtained according to the identity information, the historical purchase condition, the credit score and other information input by the user. In the present application, the feature of the click time stamp between the user account and the target product is used as the time feature, and the following description will be made specifically.

The characteristics of the commodity are as follows: optionally, the characteristics of the commodity can be obtained according to the information of the price, the historical sales volume, the inventory, the sales speed and the like of the commodity. In the present application, the feature of the click time stamp of the target product clicked by the user account is referred to as a time feature, and a specific description will be given below.

Spatial domain characteristics: in the present application, a feature obtained by fusing a feature based on a plurality of user accounts and a feature of a plurality of commodities is referred to as an airspace feature.

For a detailed description of step 240, please refer to "for step 240" below.

Step 260, based on N₁A first order commodity and M₂The second-order user accounts have click time stamps to obtain first time domain characteristics; based on M₁First order user account number and N₂The click time stamp of each second-order commodity is used for obtaining a second time domain characteristic;

click timestamp: which refers to the point in time when the user account clicks on the item (or the point in time when the item is clicked by the user account). Optionally, the granularity of the click timestamp is at least one of year, quarter, month, week, day, and time. For example, the user account "haha" clicked on xxx brand skin care product (schematically, this time day granularity) at 6/12/2021.

Time domain characteristics: in the present application, a feature obtained by fusing click timestamps between a plurality of user accounts and a plurality of commodities is referred to as a time domain feature.

For a detailed description of step 260, please refer to "for step 260" below.

And step 280, predicting the probability of the target commodity clicked by the target user account based on the first spatial domain feature, the second spatial domain feature, the first time domain feature and the second time domain feature.

Based on the obtained first spatial domain feature, the second spatial domain feature, the first time domain feature and the second time domain feature, the device 102 is used for predicting the probability of clicking the target commodity by the target user account.

For a detailed description of step 280, please refer to "for step 280" below.

In summary, the target user side (target user account number, N) is mined₁A first order commodity and M₁First order user account) and target commodity side (target commodity, M)₁First order user account number and N₂Second-order merchandise), not only explicit information (information between the target user account and the first-order merchandise, information between the target merchandise and the first-order user account), but also implicit information (information between the first-order merchandise and the second-order user account, information between the first-order user account and the second-order merchandise) is focused on from a global perspective. And a symmetrical information attention mode is adopted, so that the click rate is prevented from being predicted one-sidedly, and the effectiveness and the reliability of the predicted click rate are improved.

For example, from the perspective of the target user, implicit information between the second-order user account and the first-order merchandise may be used to analyze the interest level of the second-order user account in the first-order merchandise, and overall, the obscuration includes a rank of the interest level of the target user account between competing accounts, which can be fed back to the prediction of the click rate of the target user account in the first-order merchandise.

And the predicted click rate is more accurate by fully paying attention to the historical click information between the commodity and the user. And, the prediction network of click-through rate applying the above method is a complete end-to-end network without training or predicting in stages.

With respect to step 220: the method can be realized by constructing a first interactive subgraph at a target user side and a second interactive subgraph at a target commodity side.

FIG. 3 is a flowchart illustrating a method for constructing a first interactive subgraph and a second interactive subgraph according to an exemplary embodiment of the present application, the method comprising:

step 221, constructing an interactive bipartite graph comprising a plurality of user nodes and a plurality of commodity nodes;

the interactive bipartite graph comprises a plurality of user nodes and a plurality of commodity nodes, the user nodes and the commodity nodes are connected in an edge mode under the condition that a click relation exists between the user nodes and the commodity nodes, and the edge right between the user nodes and the commodity nodes is set to be a click timestamp. Optionally, the click relationship refers to that a click event occurs within a preset historical time period.

Schematically, fig. 4 is a schematic diagram of the interactive bipartite graph. The user node cluster 41 includes user nodes a1 to a6, and the commodity node cluster 42 includes commodity nodes B1 to B6. The user node A1 is respectively connected with the commodity node B1, the commodity node B2, the commodity node B4 and the commodity node B6;

the user node A2 is respectively connected with the commodity node B1, the commodity node B3 and the commodity node B5;

the user node A3 is respectively connected with the commodity node B1, the commodity node B2, the commodity node B4 and the commodity node B6;

the user node A4 is respectively connected with the commodity node B3, the commodity node B4 and the commodity node B5;

the user node A5 is respectively connected with the commodity node B1, the commodity node B4 and the commodity node B6;

the user node a6 is connected to the commodity node B5 and the commodity node B6, respectively.

And the edge right between the user node and the commodity node is a click timestamp. Optionally, the click timestamp is a timestamp of a latest click event, or a mean time in a time period in which the most intensive click event occurs within a preset historical time period.

In one embodiment, a matrix form may be employed

Representing an interactive bipartite graph, wherein each row represents a user node, and the total number of the user nodes is M; each column represents a commodity node, and the total number of the commodity nodes is N. The value at each position on the matrix represents the click timestamp.

Step 222, extracting a first interactive subgraph from the interactive bipartite graph;

determining a target user node, a plurality of first-order candidate commodity nodes connected with the target user node and a plurality of second-order candidate user nodes connected with the plurality of first-order candidate commodity nodes from the interactive bipartite graph; a plurality of first-order candidate commodity nodes are sorted and sampled according to the click time stamps to obtain N₁A first-order commodity node; a plurality of second-order candidate user nodes are sequenced and sampled according to click time stamps to obtain M₂And the second-order user nodes.

Illustratively, the number of the multiple first-order candidate commodity nodes is determined to be s₁A1 is to₁The first-order candidate commodity nodes are sorted from near to far according to the time stamp, and s is obtained₁Dividing the first-order candidate commodity node into N₁Stage, the number of candidate commodity nodes in each stage is s₁/N₁Optionally, a first-order candidate commodity node is randomly extracted from each stage to serve as a first-order commodity node, and based on the first-order candidate commodity node, N is obtained₁A first order commodity node.

Similarly, M can be obtained from a plurality of second-order candidate user nodes₂And the second-order user nodes.

Next, a target user node based, N, will be introduced₁A first order commodity node and M₂And each second-order user node constructs a first interactive subgraph.

According to the target user node corresponding to the target user account and the N clicked by the target user account₁N corresponding to first-order commodities₁A first order commodity node, and N₁M with click relation among first-order commodities₂Two isM corresponding to the order user account₂And constructing a first interactive subgraph of the target user side by taking the click timestamp as an edge weight between the target user node and the first-order commodity node and taking the click timestamp as an edge weight between the first-order commodity node and the second-order user node.

Schematically, fig. 5 shows a schematic diagram of a first interaction sub-diagram. Wherein:

the target user node 51 includes a target user node u. The first order commodity nodes 520 include a first order commodity node v₁₁And a first order commodity node v₁₂. The second order user node 530 includes a second order user node u₂₁Second order user node u₂₂Second order user node u₂₃And a second order user node u₂₄。

Target user node u and first-order commodity node v₁₁And a first order commodity node v₁₂Connecting edges;

first order commodity node v₁₁Respectively with second-order user nodes u₂₁And a second order user node u₂₂Connecting edges;

first order commodity node v₁₂Respectively with second-order user nodes u₂₃And a second order user node u₂₄And connecting the edges.

In one embodiment, G may be employed_uA first interactive sub-graph representing the target user side, using I_uRepresents N₁A set of first-order commodity nodes, using U_uRepresents M₂A set of second order user nodes, then:

G_u＝{u，I_u，U_u}；

and step 223, extracting a second interactive subgraph from the interactive bipartite graph.

Determining a target commodity node, a plurality of first-order candidate user nodes connected with the target commodity node and a plurality of second-order candidate commodity nodes connected with the plurality of first-order candidate user nodes from the interactive bipartite graph; sequencing and sampling a plurality of first-order user commodity nodes according to click time stamps to obtain M₁A first-order user node; a plurality of second-order candidate commodity nodes are sequenced and sampled according to the click time stampsTo N₂And each second-order commodity node.

Illustratively, the number of the plurality of first-order candidate user nodes is determined to be s₂A1 is to₂The first-order candidate user nodes are sorted from near to far according to the time stamp, and s is obtained₂Division of first order candidate user nodes into M₁Stage, the number of candidate user nodes in each stage is s₂/M₁Optionally, a first-order candidate user node is randomly extracted from each segment to serve as a first-order user node, and based on the first-order candidate user node, M is obtained₁A first order user node.

Similarly, N can be obtained from a plurality of second-order candidate commodity nodes₂And each second-order commodity node.

Next, a node based on the target commodity, M, will be described₁A first order user node and N₂And constructing a second interactive subgraph by the second-order commodity nodes.

According to the target commodity node corresponding to the target commodity and the M clicking the target commodity₁M corresponding to first-order user account₁A first order user node, and M₁N with click relation between first-order user accounts₂N corresponding to second-order commodities₂And constructing a second interactive subgraph of the target commodity side by taking the click timestamp as the edge weight between the target commodity node and the first-order user node and taking the click timestamp as the edge weight between the first-order user node and the second-order commodity node.

Schematically, fig. 6 shows a schematic diagram of a second interactive subgraph. Wherein:

the target commodity node 61 includes a target commodity node v. First order user nodes 620 includes first order user nodes u₁₁And first order user node u₁₂. The second order commodity nodes 630 include a second order commodity node v₂₁Second order commodity node v₂₂Second order commodity node v₂₃And second order commodity node v₂₄。

Target commodity node v and first-order user node u₁₁And first order user node u₁₂Connecting edges;

first order user node u₁₁Respectively with second-order commodity nodes v₂₁And second order commodity node v₂₂Connecting edges;

first order user node u₁₂Respectively with second-order commodity nodes v₂₃And second order commodity node v₂₄And connecting the edges.

In one embodiment, G may be employed_vA second interactive subgraph representing the target commodity side and adopting U_vRepresents M₁A set of first-order user nodes, using I_vRepresents N₂A set of second order commodity nodes, then:

G_v＝{v,U_v,I_v}；

in summary, by constructing the first interactive subgraph and the second interactive subgraph, the first spatial domain feature, the second spatial domain feature, the first time domain feature and the second time domain feature are calculated based on the first interactive subgraph and the second interactive subgraph, and the complex calculation process involved in the application can be simplified through the graph network.

With respect to step 240: optionally, step 240 may be further implemented by constructing the completed first interactive subgraph and the second interactive subgraph based on the method embodiment shown in fig. 3.

Fig. 7 is a flowchart illustrating a method for calculating a first spatial signature and a second spatial signature according to an exemplary embodiment of the present application, the method comprising:

step 241, through the first graph neural network, dividing N₁Features of a first order commodity node and M₂The characteristics of the second-order user nodes are transmitted to the target user node and aggregated with the characteristics of the target user node to obtain first airspace characteristics;

optionally, the first Graph neural Network is any one of GCN (Graph Convolutional neural Network), Graph sage (Graph neural Network) and GAT (Graph Attention Network).

In one embodiment, through the first graph neural network, two aggregation processes are executed, and the feature vector of the target user node obtained by the second aggregation is used as the first spatial feature vector.

During the kth polymerization (k having a value of 1 or 2): comprises N₁The individual first-order commodity node transmits the characteristics to the target user node, and, M₂A second order user node towards N₁The first-order commodity node transfer characteristics comprise two subprocesses.

In the first sub-process, first, N is₁N of first order commodity node₁Carrying out dimension-by-dimension mean aggregation on the k-1 characteristic vectors to obtain a k-first-order commodity characteristic vector, and then splicing the k-first-order commodity characteristic vector with the k-1 characteristic vector of the target user node to obtain a k-th characteristic vector of the target user node;

illustratively, the first sub-process may be represented by the following equation:

where MEAN (. smallcircle.) is a dimension-by-dimension averaging operation, W^kAnd b^kRespectively representing a trainable parameter matrix and a trainable parameter vector, CONCAT (inverse control integration) represents vector splicing operation, sigma (inverse) represents a nonlinear activation function, and a Relu function is usually taken; n (u) is a first order commodity node set of target user nodes u,

for the above-mentioned first-order k commodity feature vectors,

the k-1 th feature vector of the target user node,

for the kth feature vector of the target user node,

and the k-1 characteristic vector of a certain node in the first-order commodity node set is obtained.

In the second sub-process, for N₁The p-th one of the first-order commodity nodes, and M_2pM of second-order user nodes_2pCarrying out dimensionality-by-dimensionality mean value aggregation on the k-1 th feature vectors to obtain k-th second-order user feature vectors, splicing the k-th second-order user feature vectors with the k-1 th feature vectors of the first-order commodity nodes to obtain k-th feature vectors of the first-order commodity nodes, wherein the value of p is 1-N₁；

Wherein the content of the first and second substances,

similarly, the second sub-process can also use the equations (1) and (2) similar to the above, and the description is omitted here.

Through the two polymerization processes, the preparation of M₂The characteristics (implicit information) of the second-order user nodes are transmitted to the target user node, and the obtained characteristics of the target user node are used as first airspace characteristics.

Step 242, through the second graph neural network, connecting M₁Characteristics of a first order user node and N₂And transmitting the characteristics of the second-order commodity nodes to the target commodity nodes, and aggregating the characteristics of the second-order commodity nodes with the characteristics of the target commodity nodes to obtain second airspace characteristics.

Optionally, the second graph neural network is any one of GCN, GraphSAGE and GAT.

In one embodiment, through the second graph neural network, two times of aggregation processes are executed, and the feature vector of the target commodity node obtained through the second time of aggregation is used as a second spatial domain feature vector.

During the jth polymerization (j having a value of 1 or 2): comprises M₁The first order user node transmits the characteristics to the target commodity node, and, N₂Second order commodity node direction M₁The first-order user node transfers features in two sub-processes.

In the first sub-process, first, M is₁M of first order user nodes₁A (j) th1, carrying out dimension-by-dimension mean aggregation on the characteristic vectors to obtain a jth first-order user characteristic vector, and splicing the jth first-order user characteristic vector with a jth-1 characteristic vector of a target commodity node to obtain a jth characteristic vector of the target commodity node;

where MEAN (. smallcircle.) is a dimension-by-dimension averaging operation, W^jAnd b^jRespectively representing a trainable parameter matrix and a trainable parameter vector, CONCAT (inverse control integration) represents vector splicing operation, sigma (inverse) represents a nonlinear activation function, and a Relu function is usually taken; m (v) a first-order user node set of target commodity nodes v,

for the above-mentioned first order j user feature vectors,

the j-1 th feature vector of the target commodity node,

the j-th feature vector of the target commodity node,

is the j-1 th characteristic vector of a certain node in the first-order user node set.

In the second sub-process, for M₁The q-th one of the first order user nodes, N_2qN of second-order commodity nodes_2qThe j-1 th feature vector is subjected to dimensionality mean value polymerization to obtain a j second-order commodity feature vector, and then the j second-order commodity feature vector and oneSplicing the jth-1 characteristic vector of the first-order user node to obtain the jth characteristic vector of the first-order commodity node, wherein the value of q is 1-M₁；

Wherein the content of the first and second substances,

similarly, the second sub-process can also use equations (3) and (4) similar to the above equations, and will not be described herein.

After the two polymerization processes, N can be obtained₂The characteristics (implicit information) of the second-order commodity nodes are transmitted to the target commodity nodes, and the obtained characteristics of the target commodity nodes serve as second airspace characteristics.

It is understood that the order between the

steps

241 and 242 may be interchanged, i.e., the step 242 is executed first and then the step 241 is executed.

In summary, the features of the first-order node and the features of the second-order node are transferred to the target node through the graph neural network, and are aggregated with the features of the target node to obtain the spatial domain features. The method focuses on the global characteristics of the complete target user account-target commodity, provides a method for calculating the airspace characteristics, and further ensures the validity and reliability of the predicted click rate.

With respect to step 260: optionally, step 260 may be further implemented by constructing the completed first interactive sub-graph and second interactive sub-graph based on the method embodiment shown in fig. 3.

Fig. 8 is a flowchart illustrating a method for calculating a first time-domain feature according to an exemplary embodiment of the present application. The method comprises the following steps:

step 261, sort N by click timestamp₁Inputting the initial characteristics of the first-order commodity nodes into a first recurrent neural network to obtain the sum N₁A first time domain sub-feature corresponding to the first-order commodity node;

illustratively, N will be ordered by click time stamp₁Inputting the initial characteristic vector of each first-order commodity node into a long-time and short-time memory network, and taking the characteristic vector after the last first-order commodity node is updated as N₁A first time domain sub-feature vector corresponding to the first-order commodity node;

in one embodiment, the initial features of the first order commodity nodes are aggregated from the features and time features of the first order commodity nodes. The time characteristic is a characteristic of a click timestamp between a first-order commodity node and a target user node, and an initial feature vector of the first-order commodity node can be expressed as:

h_i＝CONCAT(x_i,time_i)； (5)

wherein x is_iIs a feature vector, time, of a first-order commodity node i itself_iThe time characteristic corresponding to the first-order commodity node i.

A long-time memory network for: setting the length of the ordered node sequence as T, and setting the current position as T e [1, T ∈]Aiming at the t-th node, based on the updated feature vector h of the t-1-th node_t-1And the initial feature vector x of the t-th node_tObtaining a first intermediate vector f by a forgetting gate_t(ii) a Based on h_t-1And x_tObtaining a second intermediate vector i through the input gate_t(ii) a Based on h_t-1And x_tObtaining candidate cell status

Based on f_tT-1 th node cell status C_t-1、i_tAnd

obtaining the cell state C of the t node_t(ii) a Based on h_t-1And x_tThe third intermediate vector o is obtained through the output gate_t(ii) a Based on C_tAnd o_tTo obtain the updated characteristic vector h of the t-th node_t。

Forget the door:

f_t＝sigmoid(W_f·CONCAT(h_t-1,x_t)+b_f)； (6)

an input gate:

i_t＝sigmoid(W_i·CONCAT(h_t-1,x_t)+b_i)； (7)

candidate cellular states

Cell State C_t：

Output gate o_t：

o_t＝sigmoid(W_o·CONCAT(h_t-1,x_t)+b_o)； (10)

Calculating the updated feature vector h of the t-th node_t：

h_t＝o_t*tanh(C_t)； (11)

Wherein, the above W_fAnd b_fIs a trainable parameter matrix and trainable parameter vector in a forgetting gate, W_iAnd b_iIs a trainable parameter matrix and a trainable parameter vector of the input gate, W_cAnd b_cIs the calculation of a trainable parameter matrix and trainable parameter vector of the candidate cell state, W_oAnd b_oIs a matrix of trainable parameters and a vector of trainable parameters for the output gate, the sigmoid function maps values to [0,1]Range, CONCAT represents vector concatenation.

Schematically, fig. 9 shows an aggregation process of the first time-domain sub-feature. The first-order commodity node vector representation 522 output by the long and short term memory network is the first time domain sub-feature.

Step 262, sort M by click timestamp₂Inputting the initial characteristics of the second-order user nodes into a first recurrent neural network to obtain the sum M₂A first time domain sub-feature corresponding to each second-order user node;

schematically, willM sorted by click timestamp₂Inputting initial feature vector of each second-order user node into a long-time and short-time memory network, and taking the feature vector updated by the last second-order user node as M₂And the first time domain sub-feature vectors correspond to the second-order user nodes.

In one embodiment, the initial characteristics of the second-order user node are obtained by aggregating the characteristics of the second-order user node and the time characteristics, and the time characteristics are characteristics of the click time stamp between the second-order user node and the first-order commodity node.

The long-short term memory network is similar to that described above in step 261 and will not be described here again.

Schematically, referring to fig. 9 in combination, fig. 9 also shows another aggregation process of the first time domain sub-features. The second-order user node vector representation 532 output by the long-time and short-time memory network is another first time domain sub-feature.

At step 263, the first time domain feature is obtained by aggregating the two first time domain sub-features.

Schematically, the two first time domain sub-feature vectors are spliced to obtain a first time domain feature vector.

Wherein the content of the first and second substances,

is equal to N₁A first time domain sub-feature vector corresponding to a first-order commodity node,

is equal to M₂A first time domain sub-feature vector corresponding to each second order user node,

is a first time domain feature vector.

Next, how to aggregate the second time-domain feature of the target commodity side will be described.

Fig. 10 shows a flowchart of a method for calculating a second time-domain feature according to an exemplary embodiment of the present application. The method comprises the following steps:

step 264, sort M by click timestamp₁Inputting the initial characteristics of the first-order user nodes into a second recurrent neural network to obtain the sum M₁A second time domain sub-feature corresponding to the first-order user node;

illustratively, M ordered by click timestamp₁Inputting the initial characteristic vector of each first-order user node into a long-time and short-time memory network, and taking the updated characteristic vector of the last first-order user node as M₁And the second time domain sub-feature vector corresponding to the first-order user node.

In one embodiment, the initial characteristics of the first-order user nodes are obtained by aggregating the characteristics of the first-order user nodes and time characteristics, and the time characteristics are characteristics of click time stamps between the first-order user nodes and the target commodity nodes.

Schematically, referring to fig. 11 in combination, fig. 11 shows an aggregation process of the second time-domain sub-feature. The first-order user node vector 622 representation output by the long-time and short-time memory network is the second time-domain sub-feature.

Step 265, sort N by click timestamp₂Inputting the initial characteristics of the second-order commodity nodes into a second recurrent neural network to obtain the sum N₂A second time domain sub-feature corresponding to each second-order commodity node;

n sorted by click timestamp₂Inputting the initial characteristic vector of each second-order commodity node into a long-time and short-time memory network, and taking the characteristic vector after the last second-order commodity node is updated as N₂A second time domain sub-feature vector corresponding to each second-order commodity node;

in one embodiment, the initial features of the second-order commodity nodes are obtained by aggregating the features of the second-order commodity nodes and the time features, and the time features are the features of the click time stamps between the second-order commodity nodes and the first-order user nodes.

Schematically, referring to fig. 11 in combination, fig. 11 also shows an aggregation process of another second time-domain sub-feature. The second-order commodity node vector representation 632 output by the long-time and short-time memory network is another second time-domain sub-feature.

Step 266, a second time domain feature is obtained by aggregating the two second time domain sub-features.

Schematically, the two second time domain sub-feature vectors are spliced to obtain a second time domain feature vector.

It should be noted that the execution sequence of the above steps 261 to 263 and the execution sequence of the steps 264 to 266 may be interchanged, that is, the steps 264 to 266 are executed first, and then the steps 261 to 263 are executed.

In summary, through the recurrent neural network, the characteristics of the click timestamp in each-order commodity (or user account) are sufficiently concerned, a method for calculating the time domain characteristics is provided, and the accuracy of the predicted click rate is ensured.

With respect to step 280:

FIG. 12 is a flowchart illustrating a method for predicting click-through rate according to an exemplary embodiment of the present application. The method comprises the following steps:

step 281, splicing the first airspace characteristic vector and the first time domain characteristic vector to obtain a characteristic vector of a target user side; splicing the second space domain feature vector and the second time domain feature vector to obtain a feature vector of the target commodity side;

the first spatial domain feature vector is obtained in step 241, the second spatial domain feature vector is obtained in step 242, the first time domain feature vector is obtained in

steps

261, 262, and 263, and the second time domain feature vector is obtained in

steps

264, 265, and 266.

In one embodiment, the first spatial domain feature vector is spliced with the first time domain feature vector to obtain a feature vector of the target user side, which is schematically,

wherein Emb_uA feature vector representing the target user side,

representing a first spatial feature vector of the spatial domain,

representing a first time domain feature vector, W_rIs a matrix of trainable parameters.

In the same way, the feature vector Emb of the target commodity side can be obtained_v。

282, splicing the feature vector of the target user side and the feature vector of the target commodity side to obtain an intermediate feature vector;

illustratively, the intermediate feature vector can be obtained by:

F＝CONCAT(Emb_u，Emb_v)； (14)

f denotes the intermediate feature vector.

And 283, predicting the probability of the target commodity clicked by the target user account through the intermediate feature vector by the multilayer perceptron.

Illustratively, the probability is obtained by:

α＝σ(MLP(F))； (15)

where σ (·) represents the sigmoid function to map the output to a value between [0,1], α represents the final predicted click rate, and MLP is a multi-layered perceptron.

In summary, the above method provides a method for predicting click rate based on the first spatial domain feature, the second spatial domain feature, the first time domain feature and the second time domain feature, and further ensures the implementation of the whole scheme.

In one embodiment, the server predicts the click rate of a game player on a certain character skin through a prediction network of the click rate, and recommends to display the character skin on a home page of a game mall of the player if the predicted click rate is higher. And the server sends the display instruction to the client, and the client makes a response for displaying the character skin.

In one embodiment, the server predicts the click rate of the reader on the information of a certain aspect through the prediction network of the click rate, and recommends and displays the relevant information of the aspect on the browser of the reader if the predicted click rate is higher. The server sends the instruction for displaying the related information to the terminal of the reader, and the terminal displays the related information.

In one embodiment, the server predicts the click rate of the consumer on a certain category of commodities through a prediction network of the click rate, and recommends related commodities of the category on the e-commerce platform if the click rate is higher. The server sends the instruction for displaying the related commodity to the terminal of the consumer, and the terminal displays the related commodity.

Fig. 13 is a block diagram illustrating a structure of a click rate prediction apparatus according to an exemplary embodiment of the present application, where the apparatus includes:

a determining module 1301, configured to determine N clicked with the target user account₁M with click relation among first-order commodities₂A second order user account; and determining and clicking M of the target commodity₁N with click relation between first-order user accounts₂A second-order commodity; n is a radical of₁、M₂、M₁And N₂Are all positive integers;

a processing module 1302 for N-based₁Features of a first order item and M₂Obtaining a first airspace characteristic by the characteristics of the second-order user accounts; and based on M₁Characteristics of first order user account and N₂Obtaining a second airspace characteristic according to the characteristics of the second-order commodities;

a processing module 1302 for further basing N₁A first order commodity and M₂The second-order user accounts have click time stamps to obtain first time domain characteristics; based on M₁First order user account number and N₂The click time stamp of each second-order commodity is used for obtaining a second time domain characteristic;

and the predicting module 1303 is used for predicting the probability of the target commodity clicked by the target user account based on the first spatial domain feature, the second spatial domain feature, the first time domain feature and the second time domain feature.

In an optional embodiment, the determining module 1301 is further configured to determine, according to a target user node corresponding to the target user account, N clicked with the target user account₁N corresponding to first-order commodities₁A first order commodity node, and N₁M with click relation among first-order commodities₂M corresponding to second-order user account₂And constructing a first interactive subgraph of the target user side by taking the click timestamp as an edge weight between the target user node and the first-order commodity node and taking the click timestamp as an edge weight between the first-order commodity node and the second-order user node.

In an optional embodiment, the determining module 1301 is further configured to determine, according to a target product node corresponding to the target product, M clicked on the target product₁M corresponding to first-order user account₁A first order user node, and M₁N with click relation between first-order user accounts₂N corresponding to second-order commodities₂And constructing a second interactive subgraph of the target commodity side by taking the click timestamp as the edge weight between the target commodity node and the first-order user node and taking the click timestamp as the edge weight between the first-order user node and the second-order commodity node.

In an alternative embodiment, the processing module 1302 is further configured to apply N to the first graph neural network₁Features of a first order commodity node and M₂And transmitting the characteristics of the second-order user nodes to the target user node, and aggregating the characteristics with the characteristics of the target user node to obtain first airspace characteristics.

In an alternative embodiment, the processing module 1302 is further configured to apply M to the neural network of the second graph₁Characteristics of a first order user node and N₂And transmitting the characteristics of the second-order commodity nodes to the target commodity nodes, and aggregating the characteristics of the second-order commodity nodes with the characteristics of the target commodity nodes to obtain second airspace characteristics.

In an optional embodiment, the processing module 1302 is further configured to perform two aggregation processes through the first graph neural network, and use the feature vector of the target user node obtained by the second aggregation as the first spatial domain feature vector;

during the k-th polymerization:

will N₁N of first order commodity node₁Carrying out dimension-by-dimension mean aggregation on the k-1 characteristic vectors to obtain a k-first-order commodity characteristic vector, and then splicing the k-first-order commodity characteristic vector with the k-1 characteristic vector of the target user node to obtain a k-th characteristic vector of the target user node;

Wherein the content of the first and second substances,

in an optional embodiment, the processing module 1302 is further configured to execute two aggregation processes through the second graph neural network, and use the feature vector of the target commodity node obtained through the second aggregation as a second spatial domain feature vector;

during the j-th polymerization:

and, for M₁The q-th one of the first order user nodes, N_2qN of second-order commodity nodes_2qThe j-1 th feature vectors are subjected to dimensionality mean value polymerization to obtain j second-order commodity feature vectors, and then the j second-order commodity feature vectors and the first order are usedSplicing the j-1 th feature vector of the user node to obtain the j-th feature vector of the first-order user node, wherein the value of j is 1 or 2, and the value of q is 1 to M₁；

Wherein the content of the first and second substances,

in an alternative embodiment, the processing module 1302 is further configured to sort N sorted by the click timestamp₁Inputting the initial characteristics of the first-order commodity nodes into a first recurrent neural network to obtain the sum N₁A first time domain sub-feature corresponding to the first-order commodity node; m sorted by click timestamp₂Inputting the initial characteristics of the second-order user nodes into a first recurrent neural network to obtain the sum M₂A first time domain sub-feature corresponding to each second-order user node; and obtaining the first time domain characteristic by aggregating the two first time domain sub-characteristics.

In an alternative embodiment, the processing module 1302 is further configured to sort the M sorted by the click timestamp₁Inputting the initial characteristics of the first-order user nodes into a second recurrent neural network to obtain the sum M₁A second time domain sub-feature corresponding to the first-order user node; n sorted by click timestamp₂Inputting the initial characteristics of the second-order commodity nodes into a second recurrent neural network to obtain the sum N₂A second time domain sub-feature corresponding to each second-order commodity node; and obtaining a second time domain characteristic by aggregating the two second time domain sub-characteristics.

In an alternative embodiment, the processing module 1302 is further configured to sort N sorted by the click timestamp₁Inputting the initial characteristic vector of each first-order commodity node into a long-time and short-time memory network, and taking the characteristic vector after the last first-order commodity node is updated as N₁And the first-order commodity node corresponds to a first time domain sub-feature vector.

In an alternative embodiment, the processing module 1302 is further configured to sort the M sorted by the click timestamp₂Second order user nodeInputting the initial feature vector into a long-time and short-time memory network, and taking the feature vector updated by the last second-order user node as M₂And the first time domain sub-feature vectors correspond to the second-order user nodes.

In an optional embodiment, the processing module 1302 is further configured to splice the two first time-domain sub-feature vectors to obtain a first time-domain feature vector.

In an alternative embodiment, the processing module 1302 is further configured to sort the M sorted by the click timestamp₁Inputting the initial characteristic vector of each first-order user node into a long-time and short-time memory network, and taking the updated characteristic vector of the last first-order user node as M₁And the second time domain sub-feature vector corresponding to the first-order user node.

In an alternative embodiment, the processing module 1302 is further configured to sort N sorted by the click timestamp₂Inputting the initial characteristic vector of each second-order commodity node into a long-time and short-time memory network, and taking the characteristic vector after the last second-order commodity node is updated as N₂And the second time domain sub-feature vectors correspond to the second-order commodity nodes.

In an optional embodiment, the processing module 1302 is further configured to splice the two second time-domain sub-feature vectors to obtain a second time-domain feature vector.

In an optional embodiment, the prediction module 1303 is further configured to splice the first spatial domain feature vector and the first time domain feature vector to obtain a feature vector of the target user side; and splicing the second space domain feature vector and the second time domain feature vector to obtain the feature vector of the target commodity side.

In an optional embodiment, the prediction module 1303 is further configured to splice the feature vector of the target user side and the feature vector of the target product side to obtain an intermediate feature vector.

In an optional embodiment, the predicting module 1303 is further configured to predict the probability that the target user account clicks the target product through the multi-layer perceptron.

In an optional embodiment, the determining module 1301 is further configured to construct an interactive bipartite graph, where the interactive bipartite graph includes a plurality of user nodes and a plurality of commodity nodes, connect edges between the user nodes and the commodity nodes when a click relationship exists between the user nodes and the commodity nodes, and set an edge right between the user nodes and the commodity nodes as a click timestamp.

In an optional embodiment, the determining module 1301 is further configured to determine a target user node, a plurality of first order candidate commodity nodes connected to the target user node, and a plurality of second order candidate user nodes connected to the plurality of first order candidate commodity nodes; a plurality of first-order candidate commodity nodes are sorted and sampled according to the click time stamps to obtain N₁A first-order commodity node; a plurality of second-order candidate user nodes are sequenced and sampled according to click time stamps to obtain M₂And the second-order user nodes.

In an optional embodiment, the determining module 1301 is further configured to determine a target commodity node, a plurality of first-order candidate user nodes connected to the target commodity node, and a plurality of second-order candidate commodity nodes connected to the plurality of first-order candidate user nodes; a plurality of first-order candidate user nodes are sequenced and sampled according to click time stamps to obtain M₁A first-order user node; a plurality of second-order candidate commodity nodes are sequenced and sampled according to the click time stamps to obtain N₂And each second-order commodity node.

In summary, the above device excavates the target user side (target user account number, N)₁A first order commodity and M₁First order user account) and target commodity side (target commodity, M)₁First order user account number and N₂Second-order merchandise), not only explicit information (information between the target user account and the first-order merchandise, information between the target merchandise and the first-order user account), but also implicit information (information between the first-order merchandise and the second-order user account, information between the first-order user account and the second-order merchandise) is focused on from a global perspective. And a symmetrical information attention mode is adopted, so that the click rate is prevented from being predicted one-sidedly, and the effectiveness and the reliability of the predicted click rate are improved.

FIG. 14 is a schematic diagram illustrating a configuration of a computer device, according to an example embodiment. The computer apparatus 1400 includes a Central Processing Unit (CPU) 1401, a system Memory 1404 including a Random Access Memory (RAM) 1402 and a Read-Only Memory (ROM) 1403, and a system bus 1405 connecting the system Memory 1404 and the Central Processing Unit 1401. The computer device 1400 also includes a basic Input/Output system (I/O system) 1406 that facilitates transfer of information between devices within the computer device, and a mass storage device 1407 for storing an operating system 1413, application programs 1414, and other program modules 1415.

The basic input/output system 1406 includes a display 1408 for displaying information and an input device 1409, such as a mouse, keyboard, etc., for user input of information. Wherein the display 1408 and input device 1409 are both connected to the central processing unit 1401 via an input-output controller 1410 connected to the system bus 1405. The basic input/output system 1406 may also include an input/output controller 1410 for receiving and processing input from a number of other devices, such as a keyboard, mouse, or electronic stylus. Similarly, input-output controller 1410 also provides output to a display screen, a printer, or other type of output device.

The mass storage device 1407 is connected to the central processing unit 1401 through a mass storage controller (not shown) connected to the system bus 1405. The mass storage device 1407 and its associated computer device-readable media provide non-volatile storage for the computer device 1400. That is, the mass storage device 1407 may include a computer device readable medium (not shown) such as a hard disk or Compact disk-Only Memory (CD-ROM) drive.

Without loss of generality, the computer device readable media may comprise computer device storage media and communication media. Computer device storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer device readable instructions, data structures, program modules or other data. Computer device storage media includes RAM, ROM, Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), CD-ROM, Digital Video Disk (DVD), or other optical, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices. Of course, those skilled in the art will appreciate that the computer device storage media is not limited to the foregoing. The system memory 1404 and mass storage device 1407 described above may collectively be referred to as memory.

The computer device 1400 may also operate as a remote computer device connected to a network through a network, such as the internet, in accordance with various embodiments of the present disclosure. That is, the computer device 1400 may be connected to the network 1411 through the network interface unit 1412 that is coupled to the system bus 1405, or may be connected to other types of networks or remote computer device systems (not shown) using the network interface unit 1412.

The memory further includes one or more programs, which are stored in the memory, and the central processing unit 1401 implements all or part of the steps of the infection tendency prediction method by executing the one or more programs.

The present application further provides a computer-readable storage medium, in which at least one instruction, at least one program, a set of codes, or a set of instructions is stored, and the at least one instruction, the at least one program, the set of codes, or the set of instructions is loaded and executed by a processor to implement the method for predicting click-through rate provided by the above method embodiments.

A computer program product or computer program is provided that includes computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to enable the computer device to execute the click rate prediction method provided by the method embodiment.

The above-mentioned serial numbers of the embodiments of the present application are merely for description and do not represent the merits of the embodiments.

It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program instructing relevant hardware, where the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.

The above description is only exemplary of the present application and should not be taken as limiting, as any modification, equivalent replacement, or improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for predicting click rate, the method comprising:

based on the N₁Characteristics of a first-order item and said M₂Obtaining a first airspace characteristic by the characteristics of the second-order user accounts; and based on said M₁Characteristics of first order user accounts and said N₂Obtaining a second airspace characteristic according to the characteristics of the second-order commodities;

based on the N₁A first order commodity and said M₂The second-order user accounts have click time stamps to obtain first time domain characteristics; based on the M₁A first order user account number and said N₂The click time stamp of each second-order commodity is used for obtaining a second time domain characteristic;

2. The method of claim 1, wherein the determining of N clicked with the target user account₁M with click relation among first-order commodities₂A second-order user account comprising:

according to the target user node corresponding to the target user account, the N clicked by the target user account₁N corresponding to first-order commodities₁A first order commodity node, and said N₁The M of the click relationship between first-order commodities₂M corresponding to second-order user account₂The click timestamp is used as an edge weight between the target user node and the first-order commodity node, the click timestamp is used as an edge weight between the first-order commodity node and the second-order user node, and a first interaction subgraph of a target user side is constructed;

the M of the target commodity is determined and clicked₁N with click relation between first-order user accounts₂A second-order commodity, comprising:

according to the target commodity node corresponding to the target commodity and the clicked target merchantSaid M of (D)₁M corresponding to first-order user account₁A first order user node, and the M₁The N of the click relation exists between first-order user accounts₂N corresponding to second-order commodities₂And constructing a second interactive subgraph of the target commodity side by taking the click timestamp as the edge weight between the target commodity node and the first-order user node and taking the click timestamp as the edge weight between the first-order user node and the second-order commodity node.

3. The method of claim 2, wherein the basing on the N is based on the N₁Characteristics of a first-order item and said M₂Obtaining a first airspace characteristic by the characteristics of the second-order user account, wherein the first airspace characteristic comprises the following steps:

the N is divided into N by a first graph neural network₁Characteristics of a first order commodity node and said M₂The characteristics of the second-order user nodes are transmitted to the target user node and aggregated with the characteristics of the target user node to obtain the first airspace characteristics;

the base is based on the M₁Characteristics of first order user accounts and said N₂And obtaining a second airspace characteristic by the characteristics of the second-order commodity, wherein the second airspace characteristic comprises the following steps:

passing the M through a second graph neural network₁Characteristics of a first order user node and said N₂And transmitting the characteristics of the second-order commodity nodes to the target commodity nodes, and aggregating the characteristics of the second-order commodity nodes with the characteristics of the target commodity nodes to obtain the second airspace characteristics.

4. The method of claim 3, wherein said N is mapped to said first graph neural network₁Characteristics of a first order commodity node and said M₂The transmitting the characteristics of the second-order user nodes to the target user node, and aggregating the characteristics with the characteristics of the target user node to obtain the first airspace characteristics, including:

executing two times of aggregation processes through the first graph neural network, and taking the feature vector of the target user node obtained by the second aggregation as a first space domain feature vector;

during the k-th polymerization:

the N is₁N of first order commodity node₁Carrying out dimension-by-dimension mean aggregation on the k-1 characteristic vectors to obtain a k-first-order commodity characteristic vector, and then splicing the k-first-order commodity characteristic vector with the k-1 characteristic vector of the target user node to obtain a k-th characteristic vector of the target user node;

and, for said N₁The p-th one of the first-order commodity nodes, and M_2pM of each second-order user node_2pCarrying out dimensionality-by-dimensionality mean value aggregation on the k-1 th feature vectors to obtain k-th second-order user feature vectors, splicing the k-th second-order user feature vectors and the k-1 th feature vectors of the first-order commodity nodes to obtain the k-th feature vectors of the first-order commodity nodes, wherein the value of k is 1 or 2, and the value of p is 1 to N₁；

Wherein the content of the first and second substances,

5. the method of claim 3, wherein the M is mapped to the second graph neural network by the second graph neural network₁Characteristics of a first order user node and said N₂The feature of each second-order commodity node is transmitted to the target commodity node, and is aggregated with the feature of the target commodity node to obtain the second airspace feature, which includes:

executing two times of aggregation processes through the second graph neural network, and taking the feature vector of the target commodity node obtained by the second aggregation as a second airspace feature vector;

during the j-th polymerization:

the M is added₁M of first order user nodes₁Carrying out dimension-by-dimension mean value aggregation on the j-1 th feature vector to obtain a j-first-order user feature vector, and then carrying out dimension-by-dimension mean value aggregation on the j-first-order user feature vector and the targetSplicing the jth-1 characteristic vector of the target commodity node to obtain the jth characteristic vector of the target commodity node;

and, for said M₁The q-th one of the first order user nodes, N_2qN of each second-order commodity node_2qCarrying out dimensionality-to-dimensionality mean value aggregation on the j-1 th feature vectors to obtain j second-order commodity feature vectors, splicing the j second-order commodity feature vectors and the j-1 th feature vectors of the first-order user nodes to obtain the j-th feature vectors of the first-order user nodes, wherein the value of j is 1 or 2, and the value of q is 1-M₁；

Wherein the content of the first and second substances,

6. the method of any of claims 2 to 5, wherein said N is based on₁A first order commodity and said M₂The click time stamp of each second-order user account obtains a first time domain characteristic, and the method comprises the following steps:

the N sorted according to the click timestamp₁Inputting the initial characteristics of the first-order commodity nodes into a first recurrent neural network to obtain the initial characteristics of the first-order commodity nodes and the N₁A first time domain sub-feature corresponding to the first-order commodity node; the M ordered according to the click timestamp₂Inputting the initial characteristics of the second-order user nodes into the first recurrent neural network to obtain the initial characteristics of the second-order user nodes and the M₂A first time domain sub-feature corresponding to each second-order user node; obtaining the first time domain feature by aggregating the two first time domain sub-features;

the base is based on the M₁A first order user account number and said N₂Obtaining a second time domain feature from the click timestamp of each second-order commodity, including:

the M ordered according to the click timestamp₁Inputting the initial characteristics of the first-order user nodes into a second recurrent neural network to obtain the initial characteristics of the first-order user nodes and the M₁A second time domain sub-feature corresponding to the first-order user node; will be arranged according to the click timestampSaid N after the sequence₂Inputting the initial characteristics of the second-order commodity nodes into the second recurrent neural network to obtain the initial characteristics of the second-order commodity nodes and the N₂A second time domain sub-feature corresponding to each second-order commodity node; and obtaining the second time domain feature by aggregating the two second time domain sub-features.

7. The method of claim 6, wherein the first recurrent neural network is a long-term memory network (LSTM);

based on the N₁A first order commodity and said M₂The click time stamp of each second-order user account obtains a first time domain characteristic, and the method comprises the following steps:

the N sorted according to the click timestamp₁Inputting the initial feature vector of each first-order commodity node into the long-time and short-time memory network, and taking the updated feature vector of the last first-order commodity node as the feature vector of the first-order commodity node and the N₁A first time domain sub-feature vector corresponding to the first-order commodity node;

the M ordered according to the click timestamp₂Inputting the initial feature vector of each second-order user node into the long-time and short-time memory network, and taking the updated feature vector of the last second-order user node as the feature vector of the M₂A first time domain sub-feature vector corresponding to each second-order user node;

and splicing the two first time domain sub-feature vectors to obtain a first time domain feature vector.

8. The method of claim 6, wherein the second recurrent neural network is a long-term memory network (LSTM);

the M ordered according to the click timestamp₁Inputting the initial characteristic vector of each first-order user node into the long-time and short-time memory network, and taking the updated characteristic vector of the last first-order user node as the characteristic vectorIs with the M₁A second time domain sub-feature vector corresponding to the first-order user node;

the N sorted according to the click timestamp₂Inputting the initial feature vector of each second-order commodity node into the long-time and short-time memory network, and taking the feature vector updated by the last second-order commodity node as the feature vector corresponding to the N₂A second time domain sub-feature vector corresponding to each second-order commodity node;

and splicing the two second time domain sub-feature vectors to obtain a second time domain feature vector.

9. The method of claim 6,

the initial characteristics of the first-order commodity nodes are obtained by aggregating the characteristics and the time characteristics of the first-order commodity nodes; the initial characteristics of the second-order user nodes are obtained by aggregating the characteristics and the time characteristics of the second-order user nodes;

the initial characteristics of the first-order user nodes are obtained by aggregating the characteristics and the time characteristics of the first-order user nodes; and the initial characteristics of the second-order commodity nodes are obtained by aggregating the characteristics and the time characteristics of the second-order commodity nodes.

10. The method according to any one of claims 1 to 5, wherein predicting the probability of the target user clicking on the target commodity based on the first spatial domain feature, the second spatial domain feature, the first time domain feature, and the second spatial domain feature comprises:

splicing the first airspace feature vector and the first time domain feature vector to obtain a feature vector of a target user side; splicing the second space domain feature vector and the second time domain feature vector to obtain a feature vector of the target commodity side;

splicing the feature vector of the target user side and the feature vector of the target commodity side to obtain an intermediate feature vector;

and predicting the probability of the target commodity clicked by the target user account through the intermediate feature vector by a multilayer perceptron.

11. The method of claim 2, further comprising:

constructing an interactive bipartite graph, wherein the interactive bipartite graph comprises a plurality of user nodes and a plurality of commodity nodes, the user nodes and the commodity nodes are connected in an edge mode under the condition that the click relationship exists between the user nodes and the commodity nodes, and the edge right between the user nodes and the commodity nodes is set as the click timestamp;

determining the target user node, a plurality of first-order candidate commodity nodes connected with the target user node and a plurality of second-order candidate user nodes connected with the first-order candidate commodity nodes; sorting and sampling the plurality of first-order candidate commodity nodes according to the click time stamps to obtain the N₁A first-order commodity node; sequencing and sampling a plurality of second-order candidate user nodes according to the click time stamp to obtain the M₂A second-order user node;

determining the target commodity node, a plurality of first-order candidate user nodes connected with the target commodity node, and a plurality of second-order candidate commodity nodes connected with the plurality of first-order candidate user nodes; the M is obtained by sequencing and sampling the plurality of first-order candidate user nodes according to the click time stamp₁A first-order user node; the plurality of second-order candidate commodity nodes are sequenced and sampled according to the click time stamps to obtain the N₂And each second-order commodity node.

12. An apparatus for predicting click rate, the apparatus comprising:

a processing module for processing the data based on the N₁Characteristics of a first-order item and said M₂Obtaining a first airspace characteristic by the characteristics of the second-order user accounts; and based on said M₁Characteristics of first order user accounts and said N₂Obtaining a second airspace characteristic according to the characteristics of the second-order commodities;

the processing module is further configured to base on the N₁A first order commodity and said M₂The second-order user accounts have click time stamps to obtain first time domain characteristics; based on the M₁A first order user account number and said N₂The click time stamp of each second-order commodity is used for obtaining a second time domain characteristic;

and the prediction module is used for predicting the probability of clicking the target commodity by the target user account based on the first spatial domain feature, the second spatial domain feature, the first time domain feature and the second time domain feature.

13. A computer device, characterized in that the computer device comprises: a processor and a memory, the memory storing a computer program that is loaded and executed by the processor to implement the prediction method of click rate according to any one of claims 1 to 11.

14. A computer-readable storage medium, storing a computer program, which is loaded and executed by a processor to implement the prediction method of click rate according to any one of claims 1 to 11.

15. A computer program product, characterized in that it stores a computer program for execution by a computer device comprising a processor and a memory, said processor loading and executing said computer program to implement the prediction method of click rate according to any one of claims 1 to 11.