CN115687791A

CN115687791A - Service information pushing method applying big data and AI analysis and online service system

Info

Publication number: CN115687791A
Application number: CN202211637345.9A
Authority: CN
Inventors: 黎晓同; 汪向阳
Original assignee: Jinan Linyu Dashu Network Technology Service Co ltd
Current assignee: Jinan Linyu Dashu Network Technology Service Co ltd
Priority date: 2022-12-20
Filing date: 2022-12-20
Publication date: 2023-02-03

Abstract

The embodiment of the application provides a service information pushing method and an online service system applying big data and AI analysis, big data analysis is carried out on user feedback big data of a target internet product, service information pushing activities aiming at associated users are triggered, corresponding internet product service information is pushed to each corresponding associated user before a next internet product interaction process is initiated based on the service information pushing activities, and therefore the reliability of the service information pushing activities can be tested by comparing service information which is pushed in a targeted mode aiming at user feedback data before the next internet product interaction process with interest point data generated later, and a basic theoretical reference basis is provided for optimization of subsequent service information pushing activities.

Description

Service information pushing method applying big data and AI analysis and online service system

Technical Field

The invention relates to the technical field of big data, in particular to a service information pushing method and an online service system applying big data and AI analysis.

Background

Currently, personalized information push can determine service information matched with a specific user and a specific scene through a personalized algorithm or technology, and the data mining and processing capacity becomes the core capacity of personalized information push different from the traditional information push. For example, conditional probability characteristics of user behaviors (used for representing the probability of iterative feedback behaviors of a certain feedback behavior in a next product interaction time sequence) are generally analyzed based on a big data mining technology at present, and then corresponding service information pushing activities are generated for personalized information pushing.

Disclosure of Invention

In order to overcome at least the above-mentioned deficiencies in the prior art, the present invention provides a service information pushing method and an online service system using big data and AI analysis.

In a first aspect, the present application provides a service information pushing method applying big data and AI analysis, which is applied to an online service system, where the online service system is in communication connection with a plurality of internet service systems, and the method includes:

big data analysis is carried out on user feedback big data of a target internet product, and service information pushing activities aiming at one or more associated users are triggered, wherein the associated users are mining targets mapped by target graph members in a feedback graph network corresponding to the user feedback big data;

based on the service information pushing activity, pushing corresponding internet product service information to each corresponding associated user before the next internet product interaction process is initiated, and analyzing the associated information between the point of interest data of the associated user in the next internet product interaction process and the internet product service information;

determining interest conversion rate of the associated user for the service information pushing activity based on the associated information between the point of interest data of the associated user in the next internet product interaction process and the internet product service information;

and generating a pushing reliability result of the service information pushing activity based on the interest conversion rate of the associated user for the service information pushing activity.

In a possible implementation manner of the first aspect, the step of performing big data analysis on user feedback big data of the target internet product and triggering a service information pushing activity for one or more associated users includes:

carrying out graph network extraction on user feedback big data of a target Internet product to obtain a corresponding feedback graph network, wherein the feedback graph network comprises graph members corresponding to feedback behaviors and product service elements and graph network links among the graph members, and the graph network links are used for indicating that feedback service relation information exists between excavation targets mapped by any two connected graph members;

extracting basic characteristic information of each graph member in the feedback graph network, and generating basic conditional probability characteristics of each graph member according to the basic characteristic information of each graph member;

aggregating the basic conditional probability features of the plurality of connected graph members of each graph member, and taking the aggregated features corresponding to each graph member as the basic aggregated conditional probability features of the plurality of connected graph members of each graph member, wherein the connected graph members are graph members located in the same graph network link with the graph members;

corresponding to each graph member, extracting basic condition confirmation features of basic aggregation conditional probability features of the graph members, splicing the basic condition confirmation features and basic feature information of the graph members, and determining progress feature information of the graph members;

and triggering service information pushing activities aiming at one or more associated users according to the progress characteristic information of one or more target graph members in the feedback graph network, wherein the associated users are mining targets mapped with the target graph members.

In a possible implementation manner of the first aspect, the extracting basic feature information of each graph member in the feedback graph network includes:

performing the following operations on each graph member respectively:

carrying out conditional probability feature decision on the initial online event features of the graph members to determine the initial conditional probability features of the graph members;

performing conditional probability feature decision on initial online event features of a plurality of connected graph members of the graph members, and determining initial conditional probability features of the plurality of connected graph members;

aggregating the initial conditional probability features of the plurality of connected graph members, and taking the aggregated features as the initial aggregated conditional probability features of the graph members;

extracting initial condition confirmation features of the initial aggregation condition probability features of the graph members, splicing the initial condition confirmation features and the initial online event features of the graph members, and determining basic feature information of the graph members.

In a possible implementation manner of the first aspect, the performing a conditional probability feature decision on the initial online event feature of the graph member and determining the initial conditional probability feature of the graph member includes:

performing relevance extraction on the basic condition field and the initial online event characteristics to determine basic condition field relevance characteristics;

performing feature derivation on the basic condition field correlation features according to basic feature derivation dimensions, and determining initial derivation features of the graph members;

performing relevance extraction on the progress condition field and the initial online event characteristics to determine the relevance characteristics of the progress condition field;

performing feature derivation on the progress condition field correlation features according to progress feature derivation dimensions, and determining candidate extension features of the graph members;

performing activation function processing on the candidate extension features, and determining the initial extension features of the graph members;

and taking the conditional probability features characterized by the initial extension features and the initial derivative features as initial conditional probability features of the graph members.

performing the following operations on each graph member respectively:

based on a Kth neural model unit in W neural model units which are sequentially connected, carrying out conditional probability feature output on loading data of the Kth neural model unit, and determining a Kth conditional probability feature; wherein, W is more than or equal to 2, K is increased in integer grade from 1, and K is more than or equal to 1 and less than or equal to W-1;

loading the Kth conditional probability feature to a K +1 neural model unit to continue outputting the conditional probability feature;

when K is 1, the loading data of the K-th neural model unit is the initial online event characteristic of the graph member, when K is more than or equal to 2 and less than or equal to W-1, the loading data of the K-th neural model unit is the K-1-th conditional probability characteristic of the K-1-th neural model unit, and when K is W-1, the K + 1-th conditional probability characteristic output by the K + 1-th neural model unit is the basic characteristic information of the graph member.

In a possible implementation manner of the first aspect, when K is 2 ≦ K ≦ W-1, the performing conditional probability feature output on the loading data of the kth neural model unit based on the kth neural model unit of the W sequentially connected neural model units, and determining the kth conditional probability feature includes:

performing the following operations based on the Kth neural model unit:

carrying out conditional probability feature decision on the K-1 conditional probability features of the graph members to determine the K-1 layer conditional probability features of the graph members, wherein the K-1 layer conditional probability features are the K-1 layer feature information of the graph members;

carrying out conditional probability feature decision on the K-1 conditional probability features of a plurality of connected graph members of the graph members to determine the K-1 layer conditional probability features of the plurality of connected graph members;

performing K-1 layer distribution aggregation on the K-1 layer conditional probability characteristics of the plurality of connected graph members, and taking the K-1 layer aggregation characteristics aiming at the graph members as the K-1 layer aggregation conditional probability characteristics of the graph members;

and extracting the K-1 layer condition confirmation feature of the K-1 layer aggregation condition probability feature of the graph member, splicing the K-1 layer condition confirmation feature and the K-1 layer feature information, and determining the K layer feature information of the graph member as the K condition probability feature.

In a possible implementation manner of the first aspect, before extracting the basic feature information of each graph member in the feedback graph network, the method further includes:

when the graph member corresponds to the feedback behavior, extracting basic behavior characteristic distribution of the feedback behavior, performing basic coding on the basic behavior characteristic distribution, determining basic coding information, and taking the basic coding information as the initial online event characteristic of the graph member;

when the graph member corresponds to the product service element, extracting the progress behavior feature distribution of the product service element, performing progress coding on the progress behavior feature distribution, determining progress coding information, and using the progress coding information as the initial online event feature of the graph member.

In a possible implementation manner of the first aspect, the aggregating the basic conditional probability features of the plurality of connected graph members of each graph member includes:

executing the following operations for each graph member respectively:

acquiring derived features of basic conditional probability features of all the connected graph members and influence parameters corresponding to all the connected graph members;

performing weighted fusion on the derived features of the basic conditional probability features of the connected graph members according to the influence parameters of the connected graph members to determine the aggregate derived features;

performing extended feature aggregation on the loaded data of an R-th aggregation unit based on the R-th aggregation unit in L aggregation units which are sequentially connected, and determining an R-th extended feature aggregation result;

wherein L is an integer greater than or equal to 2, R is increased in integer level from 1, and R is greater than or equal to 1 and less than or equal to L-1; loading the R-th extended feature aggregation result to an R + 1-th aggregation unit to continue extended feature aggregation;

when R is 1, the loaded data of the R-th aggregation unit is the extension features of the basic conditional probability features of the multiple connected graph members and the extension features of the basic conditional probability features of the graph members, when R is 2 or more and R is less than or equal to L-1, the loaded data of the R-th aggregation unit is the aggregation result of the R-1-th extension features of the R-1-th aggregation unit and the extension features of the basic conditional probability features of the multiple connected graph members, and when R is L-1, the aggregation result of the R + 1-th extension features output by the R + 1-th aggregation unit is the aggregation extension features;

and taking the conditional probability features characterized by the aggregation-derived features and the aggregation extension features as the aggregation features of the graph members.

In a possible implementation manner of the first aspect, the point of interest data of the associated user in the next internet product interaction process is determined by:

the user operation log data of the associated user in the next internet product interaction process;

loading the user operation log data into a first interest point positioning model for interest point data positioning, and determining output interest point data;

the first interest point positioning model is generated by covering model function layer information corresponding to unbalanced interest categories corresponding to full-link units of an initialized interest point positioning model by using unbalanced interest category influence factors, determining a current interest point positioning model, and performing model iterative updating on the current interest point positioning model by using template user operation data, wherein the unbalanced interest category influence factors are obtained by loading user operation training data corresponding to the unbalanced interest categories in the template user operation data into the initialized interest point positioning model for feature coding, determining user operation features and calculating according to the user operation features.

For instance, in one possible implementation of the first aspect, the method further comprises:

acquiring an initialized interest point positioning model and acquiring template user operation data, wherein the template user operation data comprises user operation training data corresponding to the unbalanced interest category; loading each user operation training data corresponding to the unbalanced interest category into the initialized interest point positioning model for feature coding, determining each user operation feature, and calculating according to each user operation feature to obtain an unbalanced interest category influence factor;

covering model function layer information corresponding to the unbalance interest categories in the initialized interest point positioning model by using the unbalance interest category influence factors, and determining a current interest point positioning model;

performing model iteration updating on the current interest point positioning model by using the template user operation data, and determining a first interest point positioning model when the model iteration termination requirement is met;

taking the first interest point positioning model as an initialization interest point positioning model, returning and loading each user operation training data corresponding to the unbalanced interest category into the initialization interest point positioning model for feature coding, determining each user operation feature, calculating according to each user operation feature to obtain an unbalanced interest category influence factor, and determining a second interest point positioning model until the iteration termination requirement of the model is met;

deploying the second interest point positioning model to an online service system, and generating a calling interface, wherein the calling interface is used for calling the second interest point positioning model to perform interest point data positioning on user operation log data;

the generation of the initialized interest point positioning model comprises the following steps:

selecting current user operation training data from the template user operation data;

loading the current user operation training data into the long-short term memory model for forward propagation calculation, and determining current interest point positioning data;

obtaining marked interest point data corresponding to the current user operation training data, and calculating a current training cost parameter value by using the current interest point positioning data and the corresponding marked interest point data;

performing back propagation calculation on the long-short term memory model by using the current training cost parameter value, determining a long-short term memory model for optimizing model functional layer information, taking the long-short term memory model for optimizing model functional layer information as the long-short term memory model, returning to the step of selecting current user operation training data from the template user operation data, and determining the initialized interest point positioning model until the model iteration termination requirement is met;

calculating to obtain an unbalanced interest category influence factor according to the user operation characteristics, wherein the method comprises the following steps:

determining target user operation characteristics corresponding to each unbalanced interest category from the user operation characteristics;

respectively calculating the averaging characteristics of the target user operation characteristics corresponding to each unbalance interest category, and determining the averaging characteristics corresponding to each unbalance interest category;

taking the averaged features corresponding to the unbalanced interest categories as unbalanced interest category influence factors corresponding to the unbalanced interest categories;

the determining a current interest point positioning model by covering model function layer information corresponding to the unbalanced interest category in the initialized interest point positioning model with the unbalanced interest category influence factor includes:

acquiring unbalance interest category influence factors corresponding to each unbalance interest category, and determining model function layer information corresponding to the unbalance interest categories corresponding to each unbalance interest category from a full-connection unit of the initialized interest point positioning model;

covering the model functional layer information corresponding to the corresponding unbalance interest categories by using the unbalance interest category influence factors corresponding to the unbalance interest categories to determine a current interest point positioning model;

the using the template user operation data to perform model iteration updating on the current interest point positioning model, and when a model iteration termination requirement is met, determining a first interest point positioning model, including:

determining target user operation training data from the template user operation data, loading the target user operation training data to the current interest point positioning model for forward propagation calculation, and determining training interest point positioning data;

obtaining target interest point data corresponding to the target user operation training data, and calculating an initial training cost parameter value based on the training interest point positioning data and the target interest point data;

determining the number of user operation training data corresponding to the target interest point data from the template user operation data, calculating target weight corresponding to the target interest point data according to the number of the user operation training data corresponding to the target interest point data, and calculating a target training cost parameter value based on the target weight and the initial training cost parameter value;

performing back propagation calculation on the current interest point positioning model by using the target training cost parameter value, determining an interest point positioning model of optimization model functional layer information, taking the interest point positioning model of the optimization model functional layer information as the current interest point positioning model, returning to determine target user operation training data from the template user operation data, loading the target user operation training data to the current interest point positioning model for forward propagation calculation, and determining the training interest point positioning data until a model iteration termination requirement is met, and determining the first interest point positioning model; the calculating the target weight corresponding to the target interest point data according to the number of the user operation training data corresponding to the target interest point data comprises the following steps:

acquiring a set super-parameter, calculating the ratio of the number of user operation training data corresponding to the set super-parameter and the target interest point data, and determining the target weight corresponding to the target interest point data;

or obtaining the template user operation data total amount, calculating the ratio of the template user operation data total amount to the user operation training data amount corresponding to the target interest point data, and determining the target weight corresponding to the target interest point data.

In a second aspect, an embodiment of the present application further provides a service information pushing system applying big data and AI analysis, where the service information pushing system applying big data and AI analysis includes an online service system and multiple internet service systems in communication connection with the online service system;

the online service system is used for:

performing big data analysis on user feedback big data of a target internet product, and triggering service information pushing activities aiming at one or more associated users, wherein the associated users are mining targets mapped by target graph members in a feedback graph network corresponding to the user feedback big data;

In any aspect, big data analysis is performed on user feedback big data of a target internet product, service information pushing activities for one or more associated users are triggered, corresponding internet product service information is pushed to each corresponding associated user before a next internet product interaction process is initiated based on the service information pushing activities, associated information between interest point data of the associated user in the next internet product interaction process and internet product service information is analyzed, interest conversion rate of the associated user for the service information pushing activities is determined based on the associated information between the interest point data of the associated user in the next internet product interaction process and the internet product service information, and therefore a pushing reliability result of the service information pushing activities is generated, reliability of the service information pushing activities can be tested through comparison of the service information which is pushed specifically for the user feedback data before the next internet product interaction process, and basic theoretical reference basis is provided for optimization of subsequent service information pushing activities.

Drawings

Fig. 1 is a schematic flow chart of a service information pushing method applying big data and AI analysis according to an embodiment of the present invention.

Detailed Description

The architecture of the service information push system 10 applying big data and AI analysis according to an embodiment of the present invention is described below, and the service information push system 10 applying big data and AI analysis may include an online service system 100 and an internet service system 200 communicatively connected to the online service system 100. In the service information pushing system 10, the online service system 100 and the internet service system 200 may cooperate to execute a service information pushing method applying big data and AI analysis described in the following method embodiments, and the detailed description of the method embodiments may be referred to in the following steps of the online service system 100 and the internet service system 200.

The service information pushing method applying the big data and the AI analysis provided in this embodiment may be executed by the online service system 100, and the service information pushing method applying the big data and the AI analysis is described in detail below with reference to fig. 1.

The Process110 performs big data analysis on user feedback big data of a target internet product, and triggers service information pushing activities for one or more associated users, where the associated users are mining targets mapped by target graph members in a feedback graph network corresponding to the user feedback big data.

In this embodiment, the target internet product may be any internet online product applied to the online service system 100, for example, an electronic commerce service product, and the user feedback big data may refer to behavior data recorded in a process of product interaction performed by each user for the target internet product. By performing big data analysis on user feedback big data of a target internet product, subsequent iteration feedback behaviors of related associated users can be analyzed and predicted, and then service information pushing activities for one or more associated users can be predicted and triggered.

The Process120 pushes corresponding internet product service information to each corresponding associated user before the next internet product interaction Process is initiated based on the service information pushing activity, and analyzes the associated information between the point of interest data of the associated user in the next internet product interaction Process and the internet product service information.

In this embodiment, before each corresponding associated user initiates the next internet product interaction flow, the corresponding internet product service information may be pushed based on the service information pushing activity in advance, and then the associated information between the point of interest data of the associated user in the next internet product interaction flow and the internet product service information is analyzed, for example, the weight, such as the coverage ratio, of the point of interest data of the associated user in the next internet product interaction flow in the internet product service information is analyzed.

The Process130 determines an interest conversion rate of the associated user for the service information pushing activity based on the associated information between the point of interest data of the associated user in the next internet product interaction Process and the internet product service information.

In this embodiment, the Process120 may determine a weight of the point of interest data of the associated user in the next internet product interaction Process in the internet product service information, and then determine an interest conversion rate of the associated user for the service information push activity based on a preset mapping relationship between the weight and the interest conversion rate.

And the Process140 generates a pushing reliability result of the service information pushing activity based on the interest conversion rate of the associated user for the service information pushing activity.

For example, the interest conversion rate of the associated user for the service information pushing activity may be directly used as the pushing reliability result of the service information pushing activity, or the interest conversion rate of the associated user for the service information pushing activity may be compared with a set interest conversion rate, and a pushing reliability result of the service information pushing activity is generated based on the comparison result, where the pushing reliability result of the service information pushing activity is output when the interest conversion rate is greater than or equal to the set interest conversion rate, or the pushing reliability result of the service information pushing activity is output when the interest conversion rate is not greater than the set interest conversion rate. Therefore, a basic theoretical reference basis can be provided for optimization of subsequent service information pushing activities.

Based on the steps, big data analysis is carried out on user feedback big data of a target internet product, service information pushing activities for one or more associated users are triggered, corresponding internet product service information is pushed to each corresponding associated user before a next internet product interaction process is initiated based on the service information pushing activities, associated information between interest point data of the associated user in the next internet product interaction process and internet product service information is analyzed, interest conversion rate of the associated user for the service information pushing activities is determined based on the associated information between the interest point data of the associated user in the next internet product interaction process and the internet product service information, and therefore pushing reliability results of the service information pushing activities are generated, reliability of the service information pushing activities can be tested through comparison of the service information which is pushed in a targeted mode aiming at the user feedback data before the next internet product interaction process, and basic theoretical reference basis is provided for optimization of subsequent service information pushing activities.

In some exemplary designs, the Process110 may be implemented by the embodiments described below.

In the Process101, graph network extraction is performed on the user feedback big data of the target internet product, and a corresponding feedback graph network is obtained.

For example, the feedback graph network comprises graph members corresponding to each feedback behavior and each product service element, and graph network links among the graph members, wherein the graph network links are used for indicating that feedback service relation information exists between mining targets mapped by two connected graph members. The feedback behavior refers to an interactive behavior generated by a user in the process of interacting with a target internet product (such as a shopping behavior in the process of live broadcasting of products for online e-commerce products). The product service element can refer to one active interactive behavior or passive interactive behavior of a target internet product, such as product live behavior of an online e-commerce product.

In the Process102, the basic feature information of each graph member in the feedback graph network is obtained, and the basic conditional probability feature of each graph member is generated according to the basic feature information of each graph member.

For example, in addition to the above-described direct use of the initial online event feature as the basic feature information, the basic feature information may be acquired by executing STEP1 to STEP 4.

STEP1: and carrying out conditional probability feature decision on the initial online event features of the graph members to determine the initial conditional probability features of the graph members.

In some exemplary design ideas, before generating the basic feature information of the basic conditional probability features of each graph member in the feedback graph network, when the graph member corresponds to a feedback behavior, for example, the graph member is a graph member corresponding to a shopping feedback behavior, a basic behavior feature distribution of the feedback behavior is obtained, and the basic behavior feature distribution is discrete data. Since the basic behavior feature distribution is discrete data, basic encoding needs to be performed on the basic behavior feature distribution, basic encoding information is determined, and the basic encoding information is used as the initial online event feature of the graph member of the feedback behavior. When the graph member corresponds to the product service element, because the distribution of the progress behavior characteristics is also discrete data, the distribution of the progress behavior characteristics needs to be subjected to progress coding, progress coding information is determined, the progress coding information is used as the initial online event characteristics of the graph member of the product service element, and the progress coding process is similar to the basic coding process and is not repeated.

When the graph members of the feedback graph network are initialized, the initial online event characteristics of each graph member v in the feedback graph network are obtained. When the graph member v corresponds to the feedback behavior, the initial online event feature is used for representing the online event feature of the feedback behavior, and when the graph member v corresponds to the product service element, the initial online event feature is used for representing the online event feature of the product service element.

For example, the initial online event feature may be subjected to mapping processing, an initial conditional probability feature corresponding to the initial online event feature is determined, and it can be clarified through the above description that the initial online event feature may characterize various information of the graph member, but there may be a problem of incomplete feature quantity only through feature output of the initial online event feature, so that the feature expression capability may be improved by introducing a conditional probability feature (which is used for representing a probability of an iterative feedback behavior of a certain feedback behavior in a next round of product interaction time sequence) to characterize information of the graph member, that is, the initial conditional probability feature is obtained according to the initial online event feature and is used for characterizing another data form of the information of the graph member. When the initial conditional probability features are gaussian distributions, the parameters of the initial conditional probability features include derivative features and extension features, the initial derivative features and the initial extension features can be obtained according to the initial online event features, and the conditional probability features characterized by the initial extension features and the initial derivative features are used as the initial conditional probability features of the graph members.

For example, for the initial derived features of the initial conditional probability features, performing relevance extraction on a basic condition field and initial online event features to determine the relevance features of the basic condition field, wherein the basic condition field is known condition data obtained through training; and performing feature derivation on the associated features of the basic condition field according to the basic feature derivation dimensions, determining the initial derivation features of the graph members, wherein the basic feature derivation dimensions are known condition data obtained through training, and the initial derivation features are used as the derivation features corresponding to the initial condition probability features. The above description embodies a process of mapping the initial online event features to the initial derivative features, and the initial derivative features related to the initial online event features can be obtained by linear processing according to the basic condition field and the basic feature derivative dimensions, so that the information characterization capability of the initial conditional probability features characterized by the initial derivative features and the initial extended features on the graph members is improved.

For example, for the initial extension feature of the initial conditional probability feature, performing relevance extraction on a progress condition field and an initial online event feature, and determining a progress condition field relevance feature, wherein the progress condition field is known condition data obtained through training; carrying out feature derivation on the association features of the progress condition field according to the progress feature derivation dimensions, and determining candidate extension features of the image members, wherein the progress feature derivation dimensions are known condition data obtained through training; and performing activation function processing on the candidate extension features, determining the initial extension features of the graph members, and taking the initial extension features as the extension features corresponding to the basic conditional probability features. The above description embodies a process of mapping the initial online event features to the initial extended features, and candidate extended features related to the initial online event features can be obtained by linear processing according to the progress condition field and the derivative dimensions of the progress features, and the accuracy of feature expression of the initial extended features to the graph members can be made stronger by activating function processing. And taking the conditional probability characteristics characterized by the initial extension characteristics and the initial derivative characteristics as initial conditional probability characteristics of the graph members.

STEP2: and carrying out conditional probability feature decision on the initial online event features of a plurality of connected graph members of the graph members, and determining the initial conditional probability features of the plurality of connected graph members.

The implementation mode of determining the basic conditional probability characteristics of the graph members by performing conditional probability characteristic decision on the initial online event characteristics of the graph members in the STEP1 is similar to the implementation mode of determining the basic conditional probability characteristics of a plurality of connected graph members by performing conditional probability characteristic decision on the initial online event characteristics of two connected graph members in the STEP2, and corresponding basic conditional probability characteristics need to be acquired corresponding to the initial online event characteristics of each connected graph member.

The vector can be converted into the conditional probability characteristic through the conditional probability characteristic decision, the conditional probability characteristic can be a Gaussian conditional probability characteristic, a multi-center Gaussian distribution and the like, and the diversity of the graph member information can be displayed through the conditional probability characteristic, so that the characterization capability of the subsequent progress characteristic information is improved.

STEP3: and aggregating the initial conditional probability features of the plurality of connected graph members, and taking the aggregated features aiming at the graph members as the initial aggregated conditional probability features of the graph members.

And when the number of the connected graph members is one, taking the initial conditional probability characteristic of the connected graph members as the initial aggregation conditional probability characteristic of the graph members.

STEP4: and acquiring initial condition confirmation characteristics of the initial aggregation condition probability characteristics of the graph members, splicing the initial condition confirmation characteristics and the initial online event characteristics of the graph members, and determining basic characteristic information of the graph members.

In some exemplary design concepts, the STEP102 can obtain the basic feature information of each graph member in the feedback graph network by executing STEPs 1021 to STEP1022 corresponding to each graph member.

In the Process1021, based on a Kth neural model unit in W sequentially connected neural model units, conditional probability feature output is carried out on loading data of the Kth neural model unit, and a Kth conditional probability feature is determined.

The AI model comprises W cascaded neural model units, the value range of W satisfies 2-W, K is increased in an integer order from 1, and the value range of K satisfies 1-W-1.

In the Process1022, the kth conditional probability feature is loaded to the K +1 th neural model unit to continue the conditional probability feature output.

And when K is 1, the loading data of the K-th neural model unit is the initial online event characteristic of the graph member, when K is more than or equal to 2 and less than or equal to W-1, the loading data of the K-th neural model unit is the K-1-th conditional probability characteristic of the K-1-th neural model unit, and when K is W-1, the K + 1-th conditional probability characteristic output by the K + 1-th neural model unit is the basic characteristic information of the graph member. The information representation capability of the basic characteristic information can be effectively improved through an iteration mode, and the follow-up accurate completion of the service information pushing activity is facilitated.

When K is 1, the above-mentioned embodiment of determining the K-th conditional probability feature by performing the conditional probability feature output on the load data of the K-th neural model unit based on the K-th neural model unit among the W sequentially connected neural model units may refer to the above description of the non-iterative method, and it is necessary to input the basic feature information output from STEP4 as the layer 1 feature information to the 2-th neural model unit instead of using the basic feature information output from STEP4 as the layer 1 feature information, and to acquire the layer 1 conditional probability feature again from the layer 1 feature information.

In some exemplary design concepts, when K is 2 ≦ K ≦ W-1, the conditional probability feature output is performed on the loaded data of the kth neural model unit based on the kth neural model unit of the W sequentially connected neural model units, and the kth conditional probability feature is determined, and STEP5-STEP8 may be performed by the kth neural model unit.

In the Process5, the conditional probability feature decision is carried out on the K-1 conditional probability features of the graph members, and the K-1 layer conditional probability features of the graph members are determined.

The K-1 conditional probability feature is the K-1 layer feature information of the graph member.

In the Process6, the conditional probability feature decision is carried out on the K-1 conditional probability features of a plurality of connected graph members of the graph members, and the K-1 layer conditional probability features of the plurality of connected graph members are determined.

The implementation of STEP5 is similar to that of STEP6, requiring the K-1-th conditional probability features corresponding to each connectivity graph member to obtain the corresponding K-1-th layer conditional probability features.

In the Process7, the K-1 layer conditional probability characteristics of the multiple connected graph members are subjected to K-1 layer distribution aggregation, and the K-1 layer aggregation characteristics of the graph members are taken as the K-1 layer aggregation conditional probability characteristics of the graph members.

And when the number of the connected graph members is one, taking the K-1 layer conditional probability characteristic of the connected graph members as the K-1 layer aggregation conditional probability characteristic of the graph members.

And in the Process8, acquiring the K-1 layer condition confirmation characteristics of the K-1 layer aggregation condition probability characteristics of the graph members, splicing the K-1 layer condition confirmation characteristics and the K-1 layer characteristic information, and determining the K layer characteristic information of the graph members as the K condition probability characteristics.

For example, the K-1 layer condition confirmation characteristics of the graph members v are generated according to the K-1 layer aggregation condition probability characteristics of the graph members v, the K-1 layer condition confirmation characteristics are subject to the K-1 layer aggregation condition probability characteristics, the K-1 layer condition confirmation characteristics and the K-1 layer characteristic information of the graph members v output by the K-1 layer network are used for splicing, and the K-1 layer characteristic information of the graph members v output by the K-1 layer network is determined.

The Process103 aggregates the basic conditional probability features of the plurality of connected graph members of each graph member, and sets the aggregated feature corresponding to each graph member as the basic aggregated conditional probability feature of the plurality of connected graph members of each graph member.

And when the number of the connected graph members is 1, taking the basic conditional probability characteristics of the connected graph members as basic aggregation conditional probability characteristics.

In some exemplary design considerations, STEP103 is specifically implemented by STEPs 1031 to STEP1033 executed corresponding to respective graph members.

And in the Process1031, performing derived feature aggregation on the basic conditional probability features of the multiple connected graph members of the graph members, and determining aggregated derived features.

In some exemplary design ideas, the above aggregation of the derivative features on the basic conditional probability features of a plurality of connected graph members to determine the aggregated derivative features can be implemented by the following technical solutions: acquiring derivative characteristics of basic conditional probability characteristics of each connected graph member and influence parameters corresponding to each connected graph member; and performing weighted fusion on the derived features of the basic conditional probability features of the connected graph members according to the influence parameters of the connected graph members to determine the aggregation derived features.

The influence parameter is the weight distributed to each derived feature when the derived features of the multiple basic conditional probability features are subjected to weighted summation, and the basic conditional probability features of the multiple connected graph members can be effectively aggregated from the dimension of the derived features through the weighted fusion of the derived features of the multiple basic conditional probability features, so that the derived features of the basic aggregated conditional probability features can effectively represent the information of the multiple connected graph members.

In Process1032, the extended feature aggregation is performed on the basic conditional probability features of the multiple connected graph members of the graph members, and the aggregated extended feature is determined.

In some exemplary design ideas, the STEP1032 performs extension feature aggregation on the basic conditional probability features of a plurality of connected graph members, and determines an aggregation extension feature, which can be implemented by the following STEPs 5 to 6:

in the Process5, based on an R-th aggregation unit in the L aggregation units connected in sequence, the extended feature aggregation is performed on the loaded data of the R-th aggregation unit, and an R-th extended feature aggregation result is determined.

Wherein the value range of L is more than or equal to 2, R is increased in an integer order from 1, and the value range of R is more than or equal to 1 and less than or equal to L-1.

In the Process6, the R-th extended feature aggregation result is loaded to the R + 1-th aggregation unit to continue the extended feature aggregation.

When R is 1, the loaded data of the R-th aggregation unit is the expansion characteristics of the basic conditional probability characteristics of a plurality of connected graph members and the expansion characteristics of the basic conditional probability characteristics of the graph members, when R is 2-L, the loaded data of the R-th aggregation unit is the aggregation result of the R-1-th expansion characteristics of the R-1-th aggregation unit and the expansion characteristics of the basic conditional probability characteristics of the plurality of connected graph members, and when R is L-1, the aggregation result of the R + 1-th expansion characteristics output by the R + 1-th aggregation unit is the aggregation expansion characteristics.

The expansion features of the basic conditional probability features are aggregated in an iterative mode, the basic conditional probability features of a plurality of connected graph members can be effectively aggregated from the dimension of the expansion features, and the aggregation expansion features are determined, so that the expansion features of the basic aggregation conditional probability features can effectively represent the information of the plurality of connected graph members.

Process1033 uses as the aggregation feature of the graph members the conditional probability features characterized by the aggregation derived features and the aggregation extended features.

Aggregation derived features can be obtained by aggregating derived features of the basic conditional probability features of the plurality of connected graph members, and aggregation extended features can be obtained by aggregating extended features of the basic conditional probability features of the plurality of connected graph members, so that the aggregation extended features and the aggregation features (basic aggregation conditional probability features) characterized by the aggregation derived features can effectively and variously represent information of the connected graph members.

In the Process104, the basic condition confirmation features of the basic aggregation conditional probability features of the graph members are obtained corresponding to the graph members, and the basic condition confirmation features and the basic feature information of the graph members are spliced to determine the progress feature information of the graph members.

In some exemplary design ideas, the STEP104 splices the basic condition confirmation features and the basic feature information of the graph members to determine the progress feature information of the graph members, and can be implemented by the following technical solutions: splicing the basic condition confirmation features and the basic feature information of the graph members to determine a splicing vector; and mapping the splicing vectors to determine the progress characteristic information of the graph members.

In the Process105, according to the progress characteristic information of one or more target graph members in the feedback graph network, service information push activities for one or more associated users are triggered.

Based on the above processes 101-105, in this embodiment, by obtaining a feedback graph network capable of representing information of a feedback service relationship between a product service element and a feedback behavior, information of the product service element and the feedback behavior can be efficiently and comprehensively obtained by obtaining the feedback graph network, basic feature information of a basic conditional probability feature of each graph member in the feedback graph network is generated, information diversity of the graph members can be represented, basic conditional probability features of multiple connected graph members of each graph member are aggregated, a basic aggregation conditional probability feature is determined, a complete conditional probability feature is propagated between the graph members and the graph members instead of a vector generated according to the conditional probability feature, loss of graph member information is avoided, a basic condition confirmation feature of the basic aggregation conditional probability feature of the graph members is spliced with the basic feature information of the graph members, progress feature information of the graph members is determined, and the progress feature information can accurately represent information of the graph members in the feedback graph network, thereby improving accuracy of subsequent service information pushing activities.

In some exemplary design concepts, reference may be made to the following embodiments regarding the generation process of the point of interest data of the associated user in the next internet product interaction process.

The Process1312 obtains the initialized interest point location model and obtains template user operation data, where the template user operation data includes user operation training data corresponding to the unbalanced interest categories.

The initialized interest point positioning model refers to an interest point positioning model obtained by performing model iteration updating by using an AI training algorithm according to user operation training data in advance, and is used for identifying interest points of the user operation data. The user operation training data are user operation data which are collected and provided with marked interest point data, and the marked interest point data are used for indicating the interest point data corresponding to the user operation training data. The AI training algorithm may be an iterative neural network algorithm. The template user operation data refers to user operation training data with unbalanced interest category distribution, and the unbalanced interest category refers to an interest category corresponding to unbalanced interest point positioning data of user operation training data in different interest categories.

For example, the online service system acquires the initialized point of interest positioning model, may directly acquire the point of interest positioning model from a third party, and uses the acquired point of interest positioning model as the initialized point of interest positioning model, where the third party may be a service party providing the point of interest positioning model. Meanwhile, the online service system acquires template user operation data and acquires user operation training data corresponding to the unbalanced interest categories from the template user operation data. In some exemplary design ideas, the online service system may first obtain template user operation data, perform model iteration updating using a deep neural network algorithm according to the template user operation data, and obtain an initialized point of interest positioning model when training is completed.

And the Process1314 is used for loading the user operation training data corresponding to the unbalanced interest categories into the initialized interest point positioning model for feature coding to obtain the user operation features, and calculating to obtain the unbalanced interest category influence factors based on the user operation features.

The user operation characteristics are used for representing characteristics corresponding to the user operation training data, and the user operation training data corresponding to each unbalanced interest category has corresponding user operation characteristics. The unbalanced interest category influence factors are obtained according to user operation characteristics corresponding to the unbalanced interest categories and are used for representing the learning effect of the model on the unbalanced interest categories, and each unbalanced interest category has a corresponding unbalanced interest category influence factor.

For example, the online service system loads each user operation training data corresponding to each unbalanced interest category into the initialized interest point positioning model for feature coding, so as to obtain each user operation feature corresponding to each unbalanced interest category, then performs inter-feature calculation on each user operation feature corresponding to each unbalanced interest category to generate a calculation result, and uses the calculation result as an unbalanced interest category influence factor, where the inter-feature calculation may be calculating an averaged feature of each user operation feature, and uses the averaged feature as the unbalanced interest category influence factor. Or combining the operation characteristics of each user, and taking the combined characteristics as the influence factors of the unbalanced interest categories. The user operation characteristics can also be sequenced to obtain the user operation characteristics at the middle position, and the user operation characteristics at the middle position are used as the influence factors of the unbalanced interest categories.

And the Process1316, covering the model function layer information corresponding to the unbalanced interest category in the initialized interest point positioning model by using the unbalanced interest category influence factor, and obtaining the current interest point positioning model.

The model function layer information refers to variable parameters inside the initialized point of interest positioning model and is obtained when the training of the initialized point of interest positioning model is completed. The model function layer information corresponding to the unbalanced interest category refers to model function layer information corresponding to the unbalanced interest category in the initialized interest point positioning model. The current interest point positioning model is obtained by covering model function layer information corresponding to the unbalanced interest categories in the initialized interest point positioning model by using all unbalanced interest category influence factors.

For example, when the online service system obtains the unbalanced interest category influence factor, it first determines the model function layer information corresponding to the unbalanced interest category in the initialized interest point location model, then deletes the model function layer information corresponding to the unbalanced interest category, writes the unbalanced interest category influence factor into the model function layer information, and ends the coverage of the unbalanced interest category influence factor, and when all the unbalanced interest categories corresponding to the unbalanced interest category end the coverage of the unbalanced interest category, obtains the current interest point location model.

And the Process1318 performs model iteration updating on the current interest point positioning model by using the template user operation data, and obtains a first interest point positioning model when the model iteration termination requirement is met.

In some exemplary design concepts, the obtained first point of interest location model may be deployed in an online service system, and a call interface of the first point of interest location model is generated, so that a user can use the call interface to call the first point of interest location model from the online service system to perform point of interest data location on operation behavior data to be identified, that is, deploy the first point of interest location model through the online service system.

Based on the steps, the user operation characteristics of each user operation training data corresponding to the unbalanced interest category are identified through the initialized interest point positioning model, so that unbalanced interest category influence factors are obtained, the unbalanced interest category influence factors are used as model function layer information corresponding to the unbalanced interest category in the initialized interest point positioning model, then model iteration updating is further carried out to obtain the first interest point positioning model, the contribution of the user operation training data corresponding to the unbalanced interest category to the interest category expression can be more prominent, so that the user operation training data corresponding to the unbalanced interest category can be fully learned, and therefore, the accuracy of identification of the unbalanced interest category is improved for the obtained first interest point positioning model, and meanwhile, the user operation training data corresponding to the unbalanced interest category are only covered, and other model function layer information is not covered, so that the user operation training data corresponding to other non-interest categories are not influenced in the interest point positioning effect of the unbalanced interest categories.

In some exemplary design considerations, after the processing 1318, that is, after performing model iteration updating on the current interest point location model by using the template user operation data, and when a model iteration termination requirement is met, obtaining the first interest point location model, the method further includes:

and the Process1322 judges whether the model iteration termination requirement is met, executes the Process1324a when the model iteration termination requirement is met, executes the Process1324b when the model iteration termination requirement is not met, and returns to the Process1314 to execute. And the Process1324a obtains a second interest point positioning model. And the Process1324b is used for taking the first interest point positioning model as an initialized interest point positioning model, returning and loading each user operation training data corresponding to the unbalanced interest category into the initialized interest point positioning model for feature coding to obtain each user operation feature, and calculating to obtain the influence factor of the unbalanced interest category based on each user operation feature.

For example, when the online service system obtains the first interest point location model, it further determines whether the model iteration termination requirement is satisfied, where the iteration refers to iteratively executing the Process1314, the Process1316, and the Process1318, each time the iteration is completed to the Process1318, it determines whether the model iteration termination requirement is satisfied, and when the model iteration termination requirement is satisfied, the corresponding first interest point location model is used as the second interest point location model. When the model iteration termination requirement is not met, the corresponding first interest point positioning model is used as an initialization interest point positioning model, and the execution is restarted from the Process 1314.

In some exemplary design considerations, after obtaining the second point of interest localization model, the method further includes:

and deploying the second interest point positioning model into the online service system, and generating a calling interface, wherein the calling interface is used for calling the second interest point positioning model to perform interest point data positioning on the user operation log data.

In some exemplary design concepts, initializing the generation of the point of interest localization model includes the following steps:

and obtaining template user operation data, loading the template user operation data into the long-term and short-term memory model for model iteration updating, and obtaining an initialized interest point positioning model when the requirement of model iteration termination is met.

For example, the online service system acquires template user operation data, performs model iteration updating on the long-term and short-term memory model by using the template user operation data, and takes the last obtained interest point positioning model as an initialized interest point positioning model when the requirement of model iteration termination is met.

In some exemplary design ideas, obtaining template user operation data, loading the template user operation data into a long-term and short-term memory model for model iteration updating, and obtaining an initialized interest point positioning model when a model iteration termination requirement is met, includes:

the Process1331 selects the current user operation training data from the template user operation data.

And the Process1332 loads the current user operation training data into the long-short term memory model for forward propagation calculation to obtain the current interest point positioning data.

For example, the online service system selects current user operation training data from the template user operation data, and loads the current user operation training data into the long-term and short-term memory model for forward propagation calculation to obtain current interest point positioning data output by the model.

And the Process1333 is used for obtaining the marking interest point data corresponding to the operation training data of the current user and calculating the current training cost parameter value by using the positioning data of the current interest point and the corresponding marking interest point data.

The marked interest point data is real interest point data corresponding to current user operation training data and is set.

For example, the online service system obtains the labeled interest point data corresponding to the current user operation training data, and calculates the training cost parameter value by using a loss function according to the current interest point positioning data and the corresponding labeled interest point data.

And the Process1334 performs back propagation calculation on the long-short term memory model by using the current training cost parameter value to obtain the long-short term memory model for optimizing the functional layer information of the model.

And the Process1335 judges whether the model iteration termination requirement is met, executes the Process1336a when the model iteration termination requirement is met, and executes the Process1336b and returns to the Process1332 to execute when the model iteration termination requirement is not met.

And a Process1336a obtaining the initialized interest point positioning model.

And the Process1336b takes the long-short term memory model for optimizing the model function layer information as the long-short term memory model, and returns the step of selecting the current user operation training data from the template user operation data for execution.

For example, the online service system determines whether the model iteration termination requirement is met, that is, whether the model function layer information of the long-term and short-term memory model is optimized, and when the model iteration termination requirement is met, that is, the model function layer information is optimized, an initialized point of interest positioning model is obtained. And when the model iteration termination requirement is not met, the model function layer information is not optimized, the long-short term memory model for optimizing the model function layer information is used as the long-short term memory model, the step of selecting the current user operation training data from the template user operation data is returned to execute the step of continuing to optimize the model function layer information, and after multiple times of optimization, the model iteration termination requirement is met, so that the final initialized interest point positioning model is obtained.

In the embodiment, the training of the long-term and short-term memory model is carried out through the template user operation data, and when the training is completed, the initialized interest point positioning model is obtained, so that the use is convenient.

In some exemplary design concepts, the calculating of the unbalanced interest category influence factor based on the operation characteristics of each user includes the steps of: determining target user operation characteristics corresponding to each unbalanced interest category from each user operation characteristic; respectively calculating the average characteristics of the target user operation characteristics corresponding to each unbalanced interest category to obtain the average characteristics corresponding to each unbalanced interest category; and taking the average characteristic corresponding to each unbalance interest category as an influence factor of the unbalance interest category corresponding to each unbalance interest category.

The target user operation characteristics refer to user operation characteristics corresponding to the unbalanced interest categories, and different unbalanced interest categories have different user operation characteristics. The average feature is obtained by calculating the sum of all target user operation features corresponding to the unbalanced interest category, and then calculating the ratio of the sum of all target user operation features to the number of all target user operation features after obtaining the number of all target user operation features, for example, the vector obtained by performing feature weighting calculation on all target user operation features corresponding to the unbalanced interest category includes (1, 1), (2, 2) and (3, 3), and is (6, 6). If the number of the operation features of the target user is 3, the average feature is (2, 2).

For example, the online service system determines all target user operation features corresponding to each unbalanced interest category from all user operation features, calculates an averaging feature corresponding to each unbalanced interest category according to all target user operation features corresponding to each unbalanced interest category, and then uses the calculated averaging feature corresponding to each unbalanced interest category as an unbalanced interest category influence factor corresponding to each interest category. For example, when there are all target user operation features corresponding to 10 unbalanced interest categories, the average features of all target user operation features corresponding to the 10 unbalanced interest categories are respectively calculated, and then unbalanced interest category influence factors corresponding to the 10 unbalanced interest categories are obtained.

Based on the steps, the target user operation characteristics corresponding to the unbalanced interest categories are determined from the user operation characteristics, then the averaging characteristics of the target user operation characteristics corresponding to the unbalanced interest categories are respectively calculated, the unbalanced interest category influence factors corresponding to the unbalanced interest categories are obtained, and the obtained unbalanced interest category influence factors can be more accurate.

In some exemplary design considerations, the Process1316, using the unbalanced interest category influence factor to cover model function layer information corresponding to the unbalanced interest category in the initialized interest point location model, obtains the current interest point location model, and includes:

and acquiring unbalance interest category influence factors corresponding to the unbalance interest categories, and determining model function layer information corresponding to the unbalance interest categories from the full-connection unit of the initialized interest point positioning model. And covering the model functional layer information corresponding to the corresponding unbalance interest category by using the unbalance interest category influence factor corresponding to each unbalance interest category to obtain the current interest point positioning model.

The full-connection unit of the initialized interest point positioning model is used for calculating the confidence degree of the interest point data according to the input features. The model function layer information corresponding to the unbalanced interest categories refers to model function layer information used when the confidence coefficient of the unbalanced interest categories is calculated in a full-connection unit of the initialized interest point positioning model.

For example, the unbalance interest category influence factor corresponding to each unbalance interest category is obtained, and the model function layer information corresponding to the unbalance interest category corresponding to each unbalance interest category is determined from the full-connection unit of the initialized interest point positioning model. And then, respectively covering the model function layer information corresponding to the corresponding unbalance interest category by using the unbalance interest category influence factor corresponding to each unbalance interest category, and obtaining the current interest point positioning model when the covering is finished. For example, if the number of the interest categories that can be identified by the initialized interest point location model is 100, and 25 of the interest categories are unbalanced interest categories, 25 unbalanced interest category influence factors that are unbalanced interest categories are obtained, then model function layer information corresponding to the 25 unbalanced interest categories is determined from the full-connection unit of the initialized interest point location model, and then the model function layer information corresponding to the unbalanced interest categories of the consistent unbalanced interest categories is covered as the unbalanced interest category influence factors, so that the current interest point location model is obtained.

In some exemplary design concepts, the process of unbalanced interest category impact factor coverage may include the following steps:

the Process1341 obtains the template user operation data of the unbalanced interest category I.

And the Process1342 loads the template user operation data of the unbalanced interest category I into the initialized interest point positioning model for feature coding to obtain each user operation feature corresponding to the unbalanced interest category I.

And the Process1343 calculates an unbalanced interest category influence factor based on each user operation characteristic corresponding to the unbalanced interest category I.

The Process1344 covers, by using the unbalanced interest category influence factor, model function layer information corresponding to the unbalanced interest category corresponding to the full-connection unit of the initialized interest point positioning model.

The Process1345 determines whether or not the processing of the unbalance interest category is completed, and executes the Process1346 when the processing is completed, and executes the Process1341 when the processing is not completed.

And a Process1346, obtaining a current interest point positioning model.

For example, the online service system first obtains all template user operation data of an unbalanced interest category, obtains user operation characteristics corresponding to all template user operation data of the unbalanced interest category through initializing the interest point positioning model, then calculates an averaging characteristic of each user operation characteristic to obtain an unbalanced interest category influence factor, directly covers model function layer information corresponding to an unbalanced interest category corresponding to a full connection unit of the initialized interest point positioning model with the unbalanced interest category influence factor, then judges whether the model function layer information corresponding to the unbalanced interest category is completely covered, and when the whole coverage is finished, indicates that the unbalanced interest category is completely processed to obtain the current interest point positioning model. And when the coverage is not completely finished, the step of acquiring all the template user operation data of the next unbalance interest category is continuously executed until all the unbalance interest categories are processed.

In the embodiment, the model function layer information corresponding to the unbalanced interest category is covered one by one until the whole coverage is finished, so that the current interest point positioning model is obtained, errors in the coverage process can be prevented, and the accuracy of the coverage is improved.

In some exemplary design considerations, the Process1318, performing model iterative update on the current interest point location model by using the template user operation data, and obtaining the first interest point location model when a model iteration termination requirement is met, includes:

the Process1351 determines target user operation training data from the template user operation data, and loads the target user operation training data to the current interest point positioning model for forward propagation calculation to obtain training interest point positioning data;

and the Process1352 acquires target interest point data corresponding to the operation training data of the target user, and calculates an initial training cost parameter value according to the training interest point positioning data and the target interest point data.

The target user operation training data refers to user operation training data randomly determined from the template user operation data, and the user operation training data may be user operation training data corresponding to an unbalanced interest category or user operation training data corresponding to a normal interest category. The training interest point positioning data refers to interest categories obtained by performing interest point data positioning on the training data operated by the target user. The target interest point data is real interest point data corresponding to the operation training data of the target user, and the real interest point data is set.

For example, the online service system randomly selects target user operation training data from the template user operation data, loads the target user operation training data to the current interest point positioning model for forward propagation calculation to obtain training interest point positioning data, obtains target interest point data corresponding to the selected target user operation training data, and then calculates an initial training cost parameter value between the training interest point positioning data and the target interest point data through a loss function.

The Process1353 determines the number of user operation training data corresponding to the target interest point data from the template user operation data, calculates the target weight corresponding to the target interest point data based on the number of user operation training data corresponding to the target interest point data, and calculates the target training cost parameter value according to the target weight and the initial training cost parameter value.

The number of the user operation training data refers to the number of the user operation training data corresponding to the target interest point data included in the template user operation data. And setting corresponding interest point data for each user operation training data in the template user operation data. The target weight refers to a template weight corresponding to the target interest point data. The template weight corresponding to each point of interest data is different.

For example, the online service system searches all user operation training data corresponding to the target interest point data from the template user operation data, performs quantity statistics to obtain the quantity of the user operation training data, calculates a target weight corresponding to the target interest point data by using the quantity of the user operation training data, and performs weighted calculation on the initial training cost parameter value by using the target weight to obtain a target training cost parameter value.

In some exemplary design ideas, calculating a target weight corresponding to target interest point data based on a user operation training data amount corresponding to the target interest point data includes: and acquiring set hyper-parameters, and calculating the ratio of the number of user operation training data corresponding to the set hyper-parameters target interest point data to obtain the target weight corresponding to the target interest point data. The setting of the super-parameter means flexible setting based on experience in advance.

In some exemplary design ideas, calculating a target weight corresponding to target interest point data based on the number of user operation training data corresponding to the target interest point data includes: and acquiring the total amount of the template user operation data, and calculating the ratio of the total amount of the template user operation data to the amount of the user operation training data corresponding to the target interest point data to obtain the target weight corresponding to the target interest point data. The total amount of the template user operation data refers to the amount of all user operation training data in the template user operation data. For example, the online service system calculates the total amount of the template user operation data, and calculates the ratio of the total amount of the template user operation data to the amount of the user operation training data corresponding to the target interest point data to obtain the target weight corresponding to the target interest point data.

And the Process1354 performs back propagation calculation on the current interest point positioning model by using the target training cost parameter value to obtain the interest point positioning model for optimizing the functional layer information of the model.

And the Process1355 judges whether the preset training completion condition is met, executes the Process1356a when the requirement of the model iteration termination is met, and executes the Process1356b when the preset training completion condition is not met.

And the Process1356b takes the interest point positioning model of the information of the optimized model function layer as a current interest point positioning model, returns to determine target user operation training data from the template user operation data, and loads the target user operation training data to the current interest point positioning model for forward propagation calculation to obtain the training interest point positioning data.

And the Process1356a obtains a first interest point positioning model.

For example, when a target training cost parameter value is obtained, the target training cost parameter value is used to perform back propagation calculation on the current point of interest positioning model, so as to obtain the point of interest positioning model of the optimized model functional layer information. At the moment, the online service system judges whether a preset training completion condition is met, obtains a first interest point positioning model when the model iteration termination requirement is met, takes the interest point positioning model for optimizing the model functional layer information as the current interest point positioning model when the preset training completion condition is not met, and returns to the Process1352 for execution.

Based on the steps, an initial training cost parameter value is calculated according to training interest point positioning data and target interest point data, then the number of user operation training data corresponding to the target interest point data is determined from template user operation data, target weight corresponding to the target interest point data is calculated based on the number of the user operation training data corresponding to the target interest point data, a target training cost parameter value is calculated according to the target weight and the initial training cost parameter value, the target training cost parameter value is used for carrying out back propagation calculation on a current interest point positioning model, an interest point positioning model of optimizing model functional layer information is obtained, when the requirement of iteration termination of the model is met, the first interest point positioning model is obtained, and the accuracy of positioning of the interest point aiming at the unbalanced interest class can be improved through the obtained first interest point positioning model.

In some exemplary design ideas, the obtaining of the second interest point localization model based on the foregoing steps specifically includes the following steps:

and the Process1361 acquires the initialized interest point positioning model and acquires template user operation data, wherein the template user operation data comprises user operation training data corresponding to the unbalanced interest categories.

The Process1362 obtains user operation training data corresponding to the unbalanced interest category I from the template user operation data, loads the user operation training data corresponding to the unbalanced interest category I into the initialized interest point positioning model for feature coding to obtain user operation features corresponding to the unbalanced interest category I,

and the Process1363 is used for calculating an averaging characteristic based on each user operation characteristic corresponding to the unbalanced interest category I and taking the averaging characteristic as an unbalanced interest category influence factor corresponding to the unbalanced interest category I.

The Process1364 covers the model function layer information corresponding to the unbalanced interest category corresponding to the full-connection unit of the initialized interest point positioning model with the unbalanced interest category influence factor.

Process1365 determines whether the processing of the unbalanced interest category is finished, executes Process1366 when the processing is finished, and returns to Process1364 when the processing is not finished.

A Process1366, obtaining a current interest point positioning model,

and the Process1367 is used for performing model iteration updating on the current interest point positioning model by using the template user operation data, and obtaining a first interest point positioning model when the model iteration termination requirement is met.

And the Process1368 judges whether the model iteration termination requirement is met, executes the Process1369a when the model iteration termination requirement is met, and executes the Process1369b and returns to the Process1364 to execute when the model iteration termination requirement is not met.

And the Process1369b takes the first interest point positioning model as an initialized interest point positioning model and returns to the Process1364 for execution.

And the Process1369a obtains a second interest point positioning model.

The training of the interest point positioning model is completed through the training framework, and the efficiency of obtaining the second interest point positioning model is improved.

In some exemplary design concepts, another embodiment of the present application may further include the following steps:

the Process1371 acquires user operation log data.

And a Process1372, which loads user operation log data into a first interest point positioning model to perform interest point data positioning, so as to obtain output interest point data, wherein the first interest point positioning model is generated by covering model function layer information corresponding to unbalanced interest categories corresponding to a full connection unit of an initialized interest point positioning model by using unbalanced interest category influence factors, the current interest point positioning model is obtained, model iteration updating is performed on the current interest point positioning model by using template user operation data, and the unbalanced interest category influence factors are obtained by loading user operation training data corresponding to the unbalanced interest categories in the template user operation data into the initialized interest point positioning model to perform feature coding, so as to obtain user operation features, and the user operation features are calculated on the basis of the user operation features.

The user operation log data refers to user operation data which needs to be subjected to interest category identification. The interest point data refers to the interest category of the user operation log data obtained by identification.

For example, the online service system may pre-deploy the point of interest localization model trained using the above-described point of interest localization model training method. For example, the first point of interest location model is deployed to an online service system, then when the online service system obtains user operation log data sent by a user terminal, the user operation log data is input into the first point of interest location model to perform point of interest location, output point of interest data is obtained, and the point of interest data is returned to the user terminal for display. In some exemplary design considerations, a second point of interest localization model may also be deployed to the online service system. And carrying out interest point data positioning on the user operation log data by using the second interest point positioning model to obtain output interest point data, and returning the interest point data to the user terminal for displaying.

In some exemplary design ideas, the point of interest positioning model obtained by training the point of interest positioning model training method can be deployed in the terminal, and the user operation log data locally stored in the terminal is directly acquired to perform point of interest positioning, so that the efficiency of obtaining the point of interest data through identification is improved.

By obtaining the user operation log data and loading the user operation log data into the first interest point positioning model for interest point data positioning, the output interest point data is obtained, and the accuracy of obtaining the unbalanced interest category can be improved.

For some possible implementations, the online service system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

The processor 110 may perform various appropriate actions and processes by a program stored in the machine-readable storage medium 120, such as program instructions related to a service information push method applying big data and AI analysis described in the foregoing embodiments. The processor 110, the machine-readable storage medium 120, and the communication unit 140 perform signal transmission through the bus 130.

In particular, the processes described in the above exemplary flow diagrams may be implemented as computer software programs, according to embodiments of the present invention. For example, embodiments of the invention include a computer program product comprising a computer program embodied on a computer-readable medium, the computer program comprising program code for performing the method illustrated in the flow chart. In such an embodiment, the computer program may be downloaded and installed from a network via the communication unit 140, and when executed by the processor 110, performs the above-described functions defined in the methods of the embodiments of the present invention.

Yet another embodiment of the present invention further provides a computer-readable storage medium, in which computer-executable instructions are stored, and the computer-executable instructions are executed by a processor to implement the service information pushing method applying big data and AI analysis according to any one of the above embodiments.

Yet another embodiment of the present invention further provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the computer program implements the service information pushing method for applying big data and AI analysis according to any one of the above embodiments.

The foregoing is only an alternative implementation of some implementation scenarios in this application, and it should be noted that it is also within the scope of the present application for a person of ordinary skill in the art to adopt other similar implementation means based on the technical idea of the present application without departing from the technical idea of the present application.

Claims

1. A service information pushing method applying big data and AI analysis, the method comprising:

2. The service information pushing method applying big data and AI analysis as claimed in claim 1, wherein the step of big data analysis of user feedback big data of target internet products to trigger service information pushing activities for one or more associated users comprises:

and triggering service information pushing activities aiming at one or more associated users according to the progress characteristic information of one or more target graph members in the feedback graph network, wherein the associated users are mining targets mapped by the target graph members.

3. The method according to claim 2, wherein the extracting the basic feature information of each graph member in the feedback graph network comprises:

performing the following operations on each graph member respectively:

4. The method for pushing service information by applying big data and AI analysis as claimed in claim 3, wherein said making a conditional probability feature decision on the initial online event features of the graph members and determining the initial conditional probability features of the graph members comprises:

performing relevance extraction on the progress condition field and the initial online event characteristic to determine a progress condition field relevance characteristic;

performing feature derivation on the associated features of the progress condition field according to a progress feature derivation dimension, and determining candidate extension features of the graph members;

performing activation function processing on the candidate extension features, and determining initial extension features of the graph members;

and taking the conditional probability feature characterized by the initial extended feature and the initial derivative feature as the initial conditional probability feature of the graph member.

5. The method according to claim 2, wherein the extracting the basic feature information of each graph member in the feedback graph network comprises:

performing the following operations on each graph member respectively:

based on a Kth neural model unit in W sequentially connected neural model units, performing conditional probability feature output on loading data of the Kth neural model unit to determine a Kth conditional probability feature; wherein, W is more than or equal to 2, K is increased progressively in integer grade from 1, and K is more than or equal to 1 and less than or equal to W-1;

when K is equal to 1, the loading data of the K-th neural model unit is the initial online event characteristic of the graph member, when K is equal to or more than 2 and is equal to or less than K and is equal to or less than W-1, the loading data of the K-th neural model unit is the K-1-th conditional probability characteristic of the K-1-th neural model unit, and when K is equal to W-1, the K + 1-th conditional probability characteristic output by the K + 1-th neural model unit is the basic characteristic information of the graph member.

6. The method as claimed in claim 5, wherein when K is 2 ≤ K ≤ W-1, the determining the kth conditional probability feature by performing conditional probability feature output on the loaded data of the kth neural model unit based on the kth neural model unit of the W sequentially-connected neural model units comprises:

performing the following operations based on the Kth neural model unit:

7. The service information pushing method applying big data and AI analysis as claimed in claim 3 or 5, wherein before extracting the basic feature information of each graph member in the feedback graph network, the method further comprises:

when the graph member corresponds to the feedback behavior, extracting basic behavior feature distribution of the feedback behavior, performing basic coding on the basic behavior feature distribution, determining basic coding information, and taking the basic coding information as the initial online event feature of the graph member;

when the graph member corresponds to the product service element, extracting the progress behavior feature distribution of the product service element, performing progress coding on the progress behavior feature distribution, determining progress coding information, and taking the progress coding information as the initial online event feature of the graph member.

8. The method according to claim 2, wherein the aggregating the basic conditional probability characteristics of the plurality of connected graph members of each graph member comprises:

executing the following operations for each graph member respectively:

performing weighted fusion on the derived features of the basic conditional probability features of the connected graph members according to the influence parameters of the connected graph members to determine aggregate derived features;

based on an R aggregation unit in L aggregation units which are sequentially connected, performing extended feature aggregation on the loaded data of the R aggregation unit, and determining an R extended feature aggregation result;

9. The service information pushing method applying big data and AI analysis as claimed in any of claims 1-8, wherein the point of interest data of the associated user in the next Internet product interaction process is determined by:

the first interest point positioning model is generated by covering model function layer information corresponding to an unbalanced interest category corresponding to a full-link unit of an initialized interest point positioning model by using an unbalanced interest category influence factor, determining a current interest point positioning model, and performing model iterative updating on the current interest point positioning model by using template user operation data, wherein the unbalanced interest category influence factor is obtained by loading each user operation training data corresponding to the unbalanced interest category in the template user operation data into the initialized interest point positioning model for feature coding, determining each user operation feature and calculating according to each user operation feature.

10. An online service system comprising a processor and a memory for storing a computer program capable of running on the processor, wherein the processor is configured to execute the service information push method of applying big data and AI analysis of any one of claims 1 to 9 when running the computer program.