CN115587250A

CN115587250A - User interest analysis method for big data mining and cloud AI deployment system

Info

Publication number: CN115587250A
Application number: CN202211269026.7A
Authority: CN
Inventors: 周刚
Original assignee: Individual
Current assignee: Individual
Priority date: 2022-05-25
Filing date: 2022-05-25
Publication date: 2023-01-10
Also published as: CN114840764B; CN114840764A

Abstract

The embodiment of the invention provides a user interest analysis method and a cloud AI deployment system for big data mining, which can be used for carrying out network weight optimization on a target training interest analysis model based on prior model loading source data, namely outputting a first prior user interest point thermodynamic diagram according to a first target training interest analysis sub-network in the target training interest analysis model, outputting a second prior user interest point thermodynamic diagram according to a second target training interest analysis sub-network in the target training interest analysis model, carrying out network weight optimization on the target training interest analysis model based on the first prior user interest point thermodynamic diagram and the second prior user interest point thermodynamic diagram, determining a calibrated user interest analysis model, and improving the precision of predicted user interest prediction thermodynamic prediction according to the calibrated user interest analysis model obtained by carrying out network weight optimization on the target training interest analysis model.

Description

User interest analysis method for big data mining and cloud AI deployment system

The application is a divisional application of Chinese application with the application number of 202210572038.0, the application date of 2022, 05 and 25, entitled "big data mining method for service user interest analysis and cloud AI deployment system".

Technical Field

The invention relates to the technical field of big data, in particular to a user interest analysis method and a cloud AI deployment system for big data mining.

Background

With the rapid development of the internet and information technology, network services such as electronic commerce, online service and transaction and the like are more and more popularized, and a large amount of information is gathered to form massive large data. Currently, through a big data technology and an AI analysis technology, a user interest point can be analyzed according to a behavior data record of a user on a related online service platform before, and then different contents are recommended to the user according to different user interest points, and such an information acquisition mode is called a personalized recommendation mode.

However, in the related art, only the user interest points are combined to push the page content, and the pertinence of the page content is poor.

Disclosure of Invention

In order to overcome at least the above disadvantages in the prior art, the present invention is directed to a user interest analysis method and a cloud AI deployment system for big data mining.

In a first aspect, an embodiment of the present invention provides a big data mining method for service of user interest analysis, which is applied to a cloud-based AI deployment system, and the method includes:

carrying out big data mining on the big data of the online behavior activities of the candidate users, and determining a user interest prediction thermodynamic diagram of the corresponding candidate users aiming at the big data of the online behavior activities of the candidate users;

determining a sequence of trusted user interest points of the candidate user based on the user interest prediction thermodynamic diagram;

analyzing the user demand information of the candidate user flow to each trusted user interest point in the trusted user interest point sequence;

and based on the user demand information of each credible user interest point, referring corresponding page content information in the online page associated with the candidate user.

In a second aspect, an embodiment of the present invention further provides a big data mining system for user interest analysis, where the big data mining system for user interest analysis includes a cloud AI deployment system and a plurality of cloud AI deployment systems in communication connection with the cloud AI deployment system;

the cloud AI deployment system is configured to:

analyzing the user demand information of the candidate user flow to each credible user interest point in the credible user interest point sequence;

By adopting the embodiment scheme of any one aspect, the online behavior activity big data of the candidate user is mined to determine the user interest prediction thermodynamic diagram of the corresponding candidate user for the online behavior activity big data of the candidate user, the credible user interest point sequence and the credible user interest point sequence of the candidate user are determined based on the user interest prediction thermodynamic diagram, the user demand information of the candidate user flowing to each credible user interest point in the credible user interest point sequence is analyzed, and the corresponding page content information is quoted in the online page associated with the candidate user based on the user demand information of each credible user interest point, so that the page content is pushed by combining the credible user interest points and the user demand information flowing to the credible user interest points on the basis of user interest prediction, and the pertinence of the page content can be improved.

Drawings

Fig. 1 is a schematic flow chart of a big data mining method serving for user interest analysis according to an embodiment of the present invention.

Detailed Description

The architecture of the big data mining system 10 for user interest analysis according to an embodiment of the present invention is described below, where the big data mining system 10 for user interest analysis may include a cloud AI deployment system 100 and a cloud AI deployment system 200 communicatively connected to the cloud AI deployment system 100. The cloud AI deployment system 100 and the cloud AI deployment system 200 in the big data mining system 10 for analyzing user interests may cooperatively perform the big data mining method for analyzing user interests described in the following method embodiments, and specific steps of the cloud AI deployment system 100 and the cloud AI deployment system 200 may refer to the detailed description of the following method embodiments.

The big data mining method for user interest analysis provided in this embodiment may be executed by the cloud AI deployment system 100, and the big data mining method for user interest analysis is described in detail below with reference to fig. 1.

The Process100 performs big data mining on the big data of the online behavior activity of the candidate user, and determines the user interest prediction thermodynamic diagram of the corresponding candidate user for the big data of the online behavior activity of the candidate user.

In this embodiment, the user interest prediction thermodynamic diagram may be used to represent interest confidence of the candidate user for each target interest point in the online behavior activity big data of the candidate user.

And the Process200 determines a credible user interest point sequence of the candidate user based on the user interest prediction thermodynamic diagram.

In this embodiment, the trusted user interest point sequence may be used to represent a user interest point sequence whose interest confidence is greater than a preset confidence.

And the Process300 analyzes the user demand information of the candidate user flowing to each trusted user interest point in the trusted user interest point sequence.

For example, flow direction path information of the candidate user flowing to each trusted user interest point in the trusted user interest point sequence may be determined from the online behavior activity big data of the candidate user (for example, the flow direction path information may include each trigger requirement information in a process that the candidate user accesses the trusted user interest point, such as a path formed by search phrase information), and trigger requirement information whose duration is longer than a preset time is obtained from the flow direction path information, and the user requirement information of the candidate user flowing to each trusted user interest point is determined.

The Process400 refers to corresponding page content information in the online page associated with the candidate user based on the user demand information of each trusted user interest point.

After the user demand information of each trusted user interest point is obtained, initial page content information related to each trusted user interest point can be searched, and then page content information is further screened out from the initial page content information based on the user demand information of each trusted user interest point to be quoted in an online page related to the candidate user.

Based on the steps, by carrying out big data mining on the online behavior activity big data of the candidate user, determining a user interest prediction thermodynamic diagram of the corresponding candidate user aiming at the online behavior activity big data of the candidate user, determining a credible user interest point sequence and a credible user interest point sequence of the candidate user based on the user interest prediction thermodynamic diagram, analyzing user demand information of the candidate user flowing to each credible user interest point in the credible user interest point sequence, and referring to corresponding page content information in an online page associated with the candidate user based on the user demand information of each credible user interest point, so that page content is pushed by combining the credible user interest point and the user demand information flowing to the credible user interest point on the basis of user interest prediction, and the pertinence of the page content can be improved.

An exemplary embodiment of the Process100 can be seen in the following.

The Process101 obtains the calibrated user interest analysis model and the candidate user corresponding to the online behavior activity big data of the candidate user, and generates model loading source data based on the candidate user and the online behavior activity big data of the candidate user.

For example, the cloud-based AI deployment system may obtain a calibrated user interest analysis model. The calibrated user interest analysis model comprises a basic user interest analysis sub-network and an extended user interest analysis sub-network. The basic user interest analysis sub-network and the extended user interest analysis sub-network are obtained by performing AI training according to a plurality of user interest dimensions respectively. In some embodiments, the cloud AI deployment system may obtain a candidate user corresponding to the online behavior activity big data of the candidate user, and use a user tag of the candidate user, a data tag of the online behavior activity big data of the candidate user, and a online behavior activity data set of the user corresponding to the candidate user as model loading source data for calibrating the user interest analysis model.

Wherein, the user label can be used for representing the candidate user, and the data label can be used for representing the online behavior activity data of the user. User tags may include, but are not limited to, a user ID, registration information (e.g., age, gender), and a user profile of the candidate user, among others; the data tags may include, but are not limited to, a data category (i.e., a data statistics category) of the online behavioral activity data of the user, an online behavioral activity ID of the user, and the like

The online behavior activity data set of the user comprises an active online behavior activity data set of the user and a passive online behavior activity data set of the user. The user active online behavior activity data set comprises user online behavior activity data of a candidate user with active activity in a preset stage, and the user passive online behavior activity data set comprises user online behavior activity data of the candidate user with passive activity in the preset stage. The passive activities may include low-span passive activities and high-span passive activities, and the preset period may be any period of time, for example, one month. For example, the user-passive online behavioral activity dataset may include a low-span user-passive online behavioral activity dataset and a high-span user-passive online behavioral activity dataset, the low-span user-passive online behavioral activity dataset may include user online behavioral activity data for which the candidate user has a low-span passive activity, and the high-span user-passive online behavioral activity dataset may include user online behavioral activity data for which the candidate user has a high-span passive activity.

The Process102 is used for performing behavior preference variable mining on the model loading source data according to the basic user interest analysis sub-network, determining interest influence connected variables corresponding to the user preference connected variables and the user interest dimensions respectively, and determining basic user interest point thermodynamic diagrams corresponding to the user interest dimensions based on the user preference connected variables and the interest influence connected variables;

for example, the cloud-based AI deployment system can load model loading source data into the base user interest analysis subnetwork. The basic user interest analysis sub-network comprises a first variable loading unit, a plurality of user preference analysis units and preference influence analysis units corresponding to a plurality of user interest dimensions. In some embodiments, the cloud AI deployment system may perform feature embedding aggregation on the model loading source data according to the first variable loading unit, and determine the first user preference member variable of the candidate user generated by the first variable loading unit for the online behavior activity big data of the candidate user. In some embodiments, the cloud AI deployment system may load the first user preference member variable to the plurality of user preference analysis units, perform connectivity node analysis on the first user preference member variable according to the plurality of user preference analysis units, and determine the user preference connectivity variable generated by each user preference analysis unit. Wherein, the user preference analysis unit may be an interest prediction unit. In some embodiments, the cloud AI deployment system may load the first user preference member variable to the plurality of preference influence analysis units, perform preference influence analysis on the first user preference member variable according to the plurality of preference influence analysis units, and determine the interest influence connected variable generated by each preference influence analysis unit. In some embodiments, the cloud AI deployment system may determine a base user interest point thermodynamic diagram corresponding to a plurality of user interest dimensions based on a plurality of user preference connectivity variables and a plurality of interest influence connectivity variables.

The user preference analysis unit may be an interest prediction unit, and the user preference analysis unit may perform interest prediction on the first user preference member variable. For example, the user preference analysis unit may also be any other model, which is not limited in this application.

On one hand, the preference influence analysis unit is light-weight, and the user preference analysis unit is common to all user interest dimensions, so that the preference influence analysis unit has advantages in calculation amount and parameter amount; on the other hand, the preference influence analysis unit of each user interest dimension realizes the selective utilization of the user preference analysis unit according to different final output weights. The preference influence analysis units of different user interest dimensions can learn the patterns of different combined user preference analysis units, so that the basic user interest analysis sub-network takes into account the relevance and difference of capturing the user interest dimensions.

When the plurality of user interest dimensions include an active user interest dimension and a passive user interest dimension, the cloud AI deployment system may generate a first active interest support corresponding to the active user interest dimension and a first passive interest support corresponding to the passive user interest dimension according to the base user interest analysis subnetwork. For example, when the plurality of user interest dimensions include an active user interest dimension, a low-span passive user interest dimension, and a high-span passive user interest dimension, the cloud AI deployment system may generate a first active interest support corresponding to the active user interest dimension, a first low-span passive interest support corresponding to the low-span passive user interest dimension, and a first high-span passive interest support corresponding to the high-span passive user interest dimension according to the base user interest analysis subnetwork.

The Process103 is used for respectively loading the model loading source data to a plurality of user interest prediction units in the expanded user interest analysis sub-network and acquiring expanded user interest point thermodynamic diagrams corresponding to a plurality of user interest dimensions generated by the expanded user interest analysis sub-network on the basis of user interest prediction data generated by each user interest prediction unit;

similarly, the cloud AI deployment system may load the model loading source data to the plurality of user interest prediction units in the extended user interest analysis sub-network, perform feature embedding aggregation on the model loading source data according to the plurality of user interest prediction units in the extended user interest analysis sub-network, generate user interest prediction data generated by each user interest prediction unit based on features obtained by the feature embedding aggregation, and generate an extended user interest point thermodynamic diagram corresponding to a plurality of user interest dimensions based on the user interest prediction data generated by each user interest prediction unit.

When the plurality of user interest dimensions include an active user interest dimension and a passive user interest dimension, the cloud AI deployment system may generate a second active interest support degree corresponding to the active user interest dimension and a second passive interest support degree corresponding to the passive user interest dimension according to the extended user interest analysis subnetwork. For example, when the plurality of user interest dimensions include an active user interest dimension, a low-span passive user interest dimension, and a high-span passive user interest dimension, the cloud AI deployment system may generate a third active interest support corresponding to the active user interest dimension, a second low-span passive interest support corresponding to the low-span passive user interest dimension, a third low-span passive interest support corresponding to the low-span passive user interest dimension, and a second high-span passive interest support corresponding to the high-span passive user interest dimension according to the extended user interest analysis subnetwork.

And the Process104 fuses the basic user interest point thermodynamic diagrams and the expanded user interest point thermodynamic diagrams in the calibrated user interest analysis model, and determines the user interest prediction thermodynamic diagrams of the candidate users for the online behavior activity big data of the candidate users.

When the multiple user interest dimensions comprise active user interest dimensions and passive user interest dimensions, the cloud AI deployment system can perform interest confidence coefficient mean value conversion on the first active interest support degree and the second active interest support degree in the calibrated user interest analysis model, and determine an active user interest prediction thermodynamic diagram of the candidate user for the candidate user online behavior activity big data. In some embodiments, the cloud-based AI deployment system may perform interest confidence mean transformation on the first passive interest support and the second passive interest support, and determine a passive user interest prediction thermodynamic diagram of the candidate user for the online behavior activity big data of the candidate user. The user interest predictive thermodynamic diagrams comprise an active user interest predictive thermodynamic diagram and a passive user interest predictive thermodynamic diagram; in other words, the active user interest predictive thermodynamic diagrams and the passive user interest predictive thermodynamic diagrams may be collectively referred to as user interest predictive thermodynamic diagrams.

For example, when the plurality of user interest dimensions include an active user interest dimension, a low-span passive user interest dimension, and a high-span passive user interest dimension, the cloud AI deployment system may perform interest confidence mean transformation on the first active interest support and the third active interest support in the calibrated user interest analysis model to determine an active user interest prediction thermodynamic diagram of the candidate user for the candidate user online behavior activity big data. In some embodiments, the cloud AI deployment system may perform interest confidence mean transformation on the first low-span passive interest support, the second low-span passive interest support, and the third low-span passive interest support to determine a low-span passive user interest prediction thermodynamic diagram of the candidate user for the online behavior activity big data of the candidate user. In some embodiments, the cloud-based AI deployment system may perform interest confidence mean transformation on the first high-span passive interest support and the second high-span passive interest support, and determine a high-span passive user interest prediction thermodynamic diagram of the candidate user for the online behavior activity big data of the candidate user. The user interest prediction thermodynamic diagrams comprise an active user interest prediction thermodynamic diagram, a low-span passive user interest prediction thermodynamic diagram and a high-span passive user interest prediction thermodynamic diagram; in other words, the active user interest predictive thermodynamic diagrams, the low-span passive user interest predictive thermodynamic diagrams, and the high-span passive user interest predictive thermodynamic diagrams may be collectively referred to as user interest predictive thermodynamic diagrams.

For example, the calibrated user interest analysis model further comprises an interest fusion unit. The number of the basic user interest analysis sub-networks is multiple, and the multiple basic user interest analysis sub-networks are used for outputting multiple basic user interest point thermodynamic diagrams; the number of the extended user interest analysis subnetworks is multiple, and the multiple extended user interest analysis subnetworks are used for outputting multiple extended user interest point thermodynamic diagrams. In this way, the cloud AI deployment system can perform interest confidence coefficient mean transformation on the multiple basic user interest point thermodynamic diagrams in the calibrated user interest analysis model, and determine the first user average interest point thermodynamic diagrams corresponding to the multiple basic user interest analysis sub-networks. In some embodiments, the cloud AI deployment system may perform interest confidence mean transformation on the multiple extended user interest point thermodynamic diagrams, and determine second user average interest point thermodynamic diagrams corresponding to the multiple extended user interest analysis subnetworks. In some embodiments, the cloud-based AI deployment system may fuse the first user average point of interest thermodynamic diagram and the second user average point of interest thermodynamic diagram to determine a fused point of interest thermodynamic diagram. In some embodiments, the cloud AI deployment system can load the fusion interest point thermodynamic diagrams to the interest fusion unit, perform fusion learning on a first user average interest point thermodynamic diagram and a second user average interest point thermodynamic diagram in the fusion interest point thermodynamic diagrams according to the interest fusion unit, and determine a user interest prediction thermodynamic diagram of a candidate user for the candidate user online behavior activity big data.

When the plurality of user interest dimensions include active user interest dimensions and passive user interest dimensions, the cloud AI deployment system can perform interest confidence mean transformation on the plurality of first active interest support degrees in the calibrated user interest analysis model, and determine first active user interest point average thermodynamic diagrams corresponding to the plurality of basic user interest analysis sub-networks. In some embodiments, the cloud AI deployment system may perform interest confidence mean transformation on the plurality of second active interest support degrees, and determine a second active user interest point average thermodynamic diagram corresponding to the plurality of extended user interest analysis subnetworks. In some embodiments, the cloud AI deployment system may fuse the first active user point-of-interest average thermodynamic diagram and the second active user point-of-interest average thermodynamic diagram to determine a fused active user point-of-interest thermodynamic diagram. In some embodiments, the cloud AI deployment system may load the fused active user interest point thermodynamic diagrams to the interest fusion unit, perform fusion learning on the first active user interest point average thermodynamic diagram and the second active user interest point average thermodynamic diagram in the fused active user interest point thermodynamic diagram according to the interest fusion unit, and determine an active user interest prediction thermodynamic diagram of the candidate user for the candidate user online behavior activity big data. When the plurality of user interest dimensions include an active user interest dimension and a passive user interest dimension, the cloud AI deployment system may refer to the description of fusion of the first active interest support and the second active interest support, which is not described herein again.

For example, when the plurality of user interest dimensions include an active user interest dimension, a low-span passive user interest dimension, and a high-span passive user interest dimension, an implementation manner in which the cloud AI deployment system fuses the first active interest support degree and the third active interest support degree, an implementation manner in which the first low-span passive interest support degree, the second low-span passive interest support degree, and the third low-span passive interest support degree are fused, and an implementation manner in which the first high-span passive interest support degree and the second high-span passive interest support degree are fused may be referred to as a description of fusion of the first active interest support degree and the third active interest support degree when the plurality of user interest dimensions include the active user interest dimension and the passive user interest dimension, which is not described herein again.

The cloud AI deployment system can fuse the basic user interest point thermodynamic diagrams generated by the basic user interest analysis sub-network and the expanded user interest point thermodynamic diagrams generated by the expanded user interest analysis sub-network, integrates the advantages of the basic user interest analysis sub-network and the expanded user interest analysis sub-network, and improves the accuracy and the robustness of multi-task modeling, so that the accuracy of predicting the user interest prediction thermodynamic diagrams is improved.

The basic user interest analysis sub-networks corresponding to the two user interest dimensions and the basic user interest analysis sub-networks corresponding to the three user interest dimensions are different, so that the first active interest support degree corresponding to the active user interest dimension in the two user interest dimensions and the first active interest support degree corresponding to the active user interest dimension in the three user interest dimensions are different. The active user interest prediction thermodynamic diagrams corresponding to the active user interest dimensions in the two user interest dimensions are different from the active user interest prediction thermodynamic diagrams corresponding to the active user interest dimensions in the three user interest dimensions. In other words, in the embodiment of the present application, a first active interest support degree corresponding to an active user interest dimension of the three user interest dimensions may be referred to as a fourth active interest support degree.

The cloud AI deployment system can acquire the basic user interest point thermodynamic diagrams of the basic user interest analysis sub-network and the extended user interest point thermodynamic diagrams of the extended user interest analysis sub-network, and further performs model integration on the basic user interest analysis sub-network and the extended user interest analysis sub-network, namely, the basic user interest point thermodynamic diagrams of the basic user interest analysis sub-network and the extended user interest point thermodynamic diagrams of the extended user interest analysis sub-network are fused.

Therefore, the online scenes corresponding to the online behavior activity big data of the candidate users can be split to determine multiple user interest dimensions, multi-network training of the basic user interest analysis sub-network and the extended user interest analysis sub-network is respectively carried out according to the multiple user interest dimensions, the characteristics that the basic user interest analysis sub-network is suitable for being small in correlation requirement among the multiple user interest dimensions and the extended user interest analysis sub-network is suitable for being large in correlation requirement among the multiple user interest dimensions are comprehensively utilized, and the accuracy of the predicted user interest prediction thermodynamic diagram is improved.

In some embodiments, another embodiment may further include the following processes 1211 to 1217, and the processes 1211 to 1217 are a specific embodiment of the aforementioned Process 102.

The Process1211 loads the model loading source data to a basic user interest analysis sub-network;

the basic user interest analysis sub-network comprises a first variable loading unit, a plurality of user preference analysis units and preference influence analysis units corresponding to a plurality of user interest dimensions.

The Process1212 is used for performing feature embedding aggregation on the model loading source data according to the first variable loading unit, and determining a first user preference member variable of the candidate user generated by the first variable loading unit for the online behavior activity big data of the candidate user;

the implementation manner of performing feature embedding aggregation on model loading source data by the cloud AI deployment system according to the first variable loading unit may refer to the foregoing description of the Process102, and is not described here again.

The Process1213 loads the first user preference member variables to the multiple user preference analysis units, performs connected node analysis on the first user preference member variables according to the multiple user preference analysis units, and determines user preference connected variables generated by the user preference analysis units respectively;

wherein the plurality of user interest dimensions comprise active user interest dimensions and passive user interest dimensions; the plurality of preference influence analyzing units comprise an active preference influence analyzing unit indicated by an active user interest dimension and a passive preference influence analyzing unit indicated by a passive user interest dimension; the plurality of interest-influence connected variables includes an active interest-influence connected variable and a passive interest-influence connected variable.

The Process1214 loads the first user preference member variable to the active preference influence analysis unit, performs preference influence analysis on the first user preference member variable according to the active preference influence analysis unit, and determines an active interest influence connected variable generated by the active preference influence analysis unit;

the Process1215 loads the first user preference member variable to the passive preference influence analysis unit, performs preference influence analysis on the first user preference member variable according to the passive preference influence analysis unit, and determines a passive interest influence connected variable generated by the passive preference influence analysis unit;

the active interest influence connected variable comprises active interest influence connected coefficients respectively corresponding to the user preference analysis units; the passive interest influence communication variables comprise passive interest influence communication coefficients respectively corresponding to the user preference analysis units; the basic user interest analysis sub-network further comprises a first active interest prediction unit and a first passive interest prediction unit; the basic user interest point thermodynamic diagram comprises a first active interest support degree and a first passive interest support degree.

A Process1216, configured to perform interest influence connectivity processing on user preference connectivity variables generated by each user preference analysis unit based on active interest influence connectivity coefficients corresponding to each user preference analysis unit, determine a target active preference feature, load the target active preference feature to a first active interest prediction unit, perform interest prediction on the target active preference feature according to the first active interest prediction unit, determine a first active interest prediction variable, and obtain a first active interest support corresponding to an active user interest dimension based on the first active interest prediction variable;

the cloud AI deployment system can take the product of the active interest influence communication coefficient corresponding to each user preference analysis unit and the user preference communication variable generated by each user preference analysis unit as the active weight fusion feature corresponding to each user preference analysis unit, and further accumulate the active weight fusion features corresponding to each user preference analysis unit to determine the target active preference feature.

The first active interest predictor variable is composed of a first active interest predictor variable value and a second active interest predictor variable value, the first active interest predictor variable value can represent that the online behavior activity big data of the candidate user has the interest prediction value of the active activity, and the second active interest predictor variable value can represent that the online behavior activity big data of the candidate user does not have the interest prediction value of the active activity. Therefore, the cloud-based AI deployment system can use the first active interest predictor variable value as a first active interest support degree of the candidate user for the online behavior activity big data of the candidate user.

The Process1217 is used for performing interest influence communication processing on the user preference connected variables generated by the user preference analysis units respectively based on the passive interest influence communication coefficients corresponding to the user preference analysis units respectively, determining a target passive preference feature, loading the target passive preference feature to a first passive interest prediction unit, performing interest prediction on the target passive preference feature according to the first passive interest prediction unit, determining a first passive interest prediction variable, and obtaining a first passive interest support corresponding to the passive user interest dimension based on the first passive interest prediction variable.

The cloud AI deployment system can take the product of the passive interest influence communication coefficient corresponding to each user preference analysis unit and the user preference communication variable generated by each user preference analysis unit as the passive weight fusion feature corresponding to each user preference analysis unit, and further accumulate the passive weight fusion features corresponding to each user preference analysis unit to determine the target passive preference feature.

The first passive interest prediction variable is formed by a first passive interest prediction variable value and a second passive interest prediction variable value, the first passive interest prediction variable value can represent an interest prediction value that the online behavior activity big data of the candidate user has passive activity, and the second passive interest prediction variable value can represent an interest prediction value that the online behavior activity big data of the candidate user does not have passive activity. Therefore, the cloud-based AI deployment system can use the first passive interest predictor variable value as the first passive interest support degree of the candidate user for the online behavior activity big data of the candidate user.

The embodiments of Process1211-Process1217 may be, for example: the first user preference member variable may be a feature generated by a first variable loading unit in the basic user interest analysis sub-network, and the basic user interest analysis sub-network may include a plurality of user preference analysis units, for convenience of understanding, the number of the user preference analysis units is N for an example, and the N user preference analysis units may specifically include: a user preference analysis unit U1, a user preference analysis unit U2, \8230, and a user preference analysis unit UN, where N may be a positive integer.

The cloud AI deployment system can load first user preference member variables into a user preference analysis unit U1 and a user preference analysis unit U2, \8230, determine the user preference analysis unit U1 and the user preference analysis unit U2, \8230, and generate user preference connected variables respectively by the user preference analysis unit UN; the cloud AI deployment system can load the first user preference member variables to the active preference influence analysis unit and the passive preference influence analysis unit respectively, and determine active interest influence connected variables generated by the active preference influence analysis unit and passive interest influence connected variables generated by the passive preference influence analysis unit.

The cloud AI deployment system can perform interest influence communication processing on user preference connected variables respectively generated by a user preference analysis unit UN and determine target active preference characteristics corresponding to the active interest influence connected coefficients on the basis of active interest influence connected coefficients in the active interest influence connected variables, wherein the user preference analysis unit U1 and the user preference analysis units U2 and 8230are performed on the basis of the active interest influence connected coefficients in the active interest influence connected variables; the cloud AI deployment system can perform interest influence communication processing on the user preference connected variables generated by the user preference analysis unit UN and the user preference analysis unit U1 and the user preference analysis unit U2, \8230basedon the passive interest influence communication coefficients in the passive interest influence communication variables, and determine target passive preference characteristics corresponding to the passive interest influence communication coefficients. In some embodiments, the cloud AI deployment system may obtain a first active interest prediction variable corresponding to the target active preference feature based on the first active interest prediction unit, and determine a first active interest support degree based on the first active interest prediction variable; the cloud AI deployment system can obtain a first passive interest prediction variable corresponding to the target passive preference feature based on the first passive interest prediction unit, and determine a first passive interest support degree based on the first passive interest prediction variable. The first active interest support degree and the first passive interest support degree can be collectively referred to as a basic user interest point thermodynamic diagram.

The first active interest support degree may indicate that the online activity big data of the candidate users generated by the base user interest analysis sub-network has an interest prediction value of the active activities performed by the candidate users, and the first passive interest support degree may indicate that the online activity big data of the candidate users generated by the base user interest analysis sub-network has an interest prediction value of the passive activities performed by the candidate users.

Based on the above steps, in this embodiment, when the multiple user interest dimensions are active user interest dimensions and passive user interest dimensions, the analysis subnetwork may output, according to the basic user interest, active interest influence connected variables corresponding to the active user interest dimensions, passive interest influence connected variables corresponding to the passive user interest dimensions, and multiple user preference connected variables, further generate, based on the active interest influence connected variables and the multiple user preference connected variables, first active interest support degrees of the candidate users for the online behavior activity big data, and generate, based on the passive interest influence connected variables and the multiple user preference connected variables, first passive interest support degrees of the candidate users for the online behavior activity big data, thereby improving accuracy of predicting the first active interest support degrees and the first passive interest support degrees.

In some embodiments, another embodiment may include the following processes 1221 to 1229, and the processes 1221 to 1229 are a specific embodiment of the Process 102.

A Process1221, which loads the model loading source data to a basic user interest analysis sub-network;

The Process1222, performing feature embedding aggregation on the model loading source data according to the first variable loading unit, and determining a first user preference member variable of the candidate user generated by the first variable loading unit for the online behavior activity big data of the candidate user;

the implementation manner of the cloud AI deployment system performing feature embedding aggregation on the model loading source data according to the first variable loading unit may refer to the foregoing description of the Process102, and is not described herein again.

The Process1223 loads the first user preference member variables to the plurality of user preference analysis units respectively, performs connected node analysis on the first user preference member variables according to the plurality of user preference analysis units respectively, and determines user preference connected variables generated by each user preference analysis unit respectively;

the Process1224 loads the first user preference member variable to the active preference influence analysis unit, performs preference influence analysis on the first user preference member variable according to the active preference influence analysis unit, and determines an active interest influence connected variable generated by the active preference influence analysis unit;

the passive preference influence analysis unit comprises a low-span passive preference influence analysis unit and a high-span passive preference influence analysis unit; the passive interest-influencing force connected variables include a low span passive interest-influencing force connected variable and a high span passive interest-influencing force connected variable.

The Process1225 loads the first user preference member variable to the low-span passive preference influence analysis unit, performs preference influence analysis on the first user preference member variable according to the low-span passive preference influence analysis unit, and determines a low-span passive interest influence connected variable generated by the low-span passive preference influence analysis unit;

the Process1226 loads the first user preference member variable to the high-span passive preference influence analysis unit, performs preference influence analysis on the first user preference member variable according to the high-span passive preference influence analysis unit, and determines a high-span passive interest influence connected variable generated by the high-span passive preference influence analysis unit;

a Process1227, which performs interest influence communication processing on user preference communication variables generated by each user preference analysis unit based on active interest influence communication coefficients corresponding to each user preference analysis unit, determines a target active preference feature, loads the target active preference feature to a first active interest prediction unit, performs interest prediction on the target active preference feature according to the first active interest prediction unit, determines a first active interest prediction variable, and obtains a first active interest support degree corresponding to an active user interest dimension based on the first active interest prediction variable;

the cloud AI deployment system can take the product of the active interest influence communication coefficient corresponding to each user preference analysis unit and the user preference communication variable generated by each user preference analysis unit as the active weight fusion characteristic corresponding to each user preference analysis unit, and further accumulate the active weight fusion characteristics corresponding to each user preference analysis unit to determine the target active preference characteristic.

The low-span passive interest influence communication variables comprise low-span passive interest influence communication coefficients respectively corresponding to the user preference analysis units; the high-span passive interest influence connected variable comprises high-span passive interest influence connected coefficients respectively corresponding to the user preference analysis units; the target passive preference features comprise target low-span passive preference features and target high-span passive preference features; the first passive interest prediction unit comprises a first low-span passive interest prediction unit and a high-span passive interest prediction unit (namely a first high-span passive interest prediction unit); the first passive interest predictor variables include a first low-span passive interest predictor variable and a high-span passive interest predictor variable (i.e., a first high-span passive interest predictor variable); the first passive interest support degree comprises a first low-span passive interest support degree and a first high-span passive interest support degree; the passive user interest dimensions include a low-span passive user interest dimension and a high-span passive user interest dimension.

A Process1228, which performs interest influence communication processing on user preference connected variables generated by each user preference analysis unit based on low-span passive interest influence communication coefficients corresponding to each user preference analysis unit, determines a target low-span passive preference feature, loads the target low-span passive preference feature to a first low-span passive interest prediction unit, performs interest prediction on the target low-span passive preference feature according to the first low-span passive interest prediction unit, determines a first low-span passive interest prediction variable, and obtains a first low-span passive interest support degree corresponding to a low-span passive user interest dimension based on the first low-span passive interest prediction variable;

the cloud AI deployment system can take the product of the low-span passive interest influence connectivity coefficient corresponding to each user preference analysis unit and the user preference connectivity variable generated by each user preference analysis unit as the low-span passive weight fusion feature corresponding to each user preference analysis unit, and further accumulate the low-span passive weight fusion features corresponding to each user preference analysis unit to determine the target low-span passive preference feature.

The first low-span passive interest predictor variable is composed of a first low-span passive interest predictor variable value and a second low-span passive interest predictor variable value, the first low-span passive interest predictor variable value can represent that the online behavior activity big data of the candidate user has the interest prediction value of the low-span passive activity, and the second low-span passive interest predictor variable value can represent that the online behavior activity big data of the candidate user does not have the interest prediction value of the low-span passive activity. Therefore, the cloud-based AI deployment system can use the first low-span passive interest predictor variable value as the first low-span passive interest support degree of the candidate user for the online behavior activity big data of the candidate user.

The Process1229 performs interest influence communication processing on user preference connected variables generated by each user preference analysis unit based on high-span passive interest influence communication coefficients corresponding to each user preference analysis unit, determines a target high-span passive preference feature, loads the target high-span passive preference feature to a high-span passive interest prediction unit (i.e., a first high-span passive interest prediction unit), performs interest prediction on the target high-span passive preference feature according to the high-span passive interest prediction unit (i.e., the first high-span passive interest prediction unit), determines a high-span passive interest prediction variable (i.e., a first high-span passive interest prediction variable), and obtains a first high-span passive interest support degree corresponding to the high-span passive user interest dimension based on the high-span passive interest prediction variable (i.e., the first high-span passive interest prediction variable).

The cloud AI deployment system can take the product of the high-span passive interest influence communication coefficient corresponding to each user preference analysis unit and the user preference communication variable generated by each user preference analysis unit as the high-span passive weight fusion feature corresponding to each user preference analysis unit, and further accumulate the high-span passive weight fusion features corresponding to each user preference analysis unit to determine the target high-span passive preference feature.

The first high-span passive interest predictive variable is composed of a first high-span passive interest predictive variable value and a second high-span passive interest predictive variable value, the first high-span passive interest predictive variable value can represent an interest predictive value that the behavior activity big data on the candidate user line has high-span passive activity, and the second high-span passive interest predictive variable value can represent an interest predictive value that the behavior activity big data on the candidate user line does not have high-span passive activity. Therefore, the cloud-based AI deployment system can use the first high-span passive interest predictor variable value as a first high-span passive interest support degree of the candidate user for the online behavior activity big data of the candidate user.

The first active interest prediction unit, the first low-span passive interest prediction unit and the first high-span passive interest prediction unit may be multi-layer perceptrons.

The cloud AI deployment system can collectively refer to the first active interest support, the first low-span passive interest support, and the first high-span passive interest support as a base user interest point thermodynamic diagram.

The first active interest support degree may indicate that the online behavior activity big data of the candidate users generated by the base user interest analysis sub-network has an interest prediction value of the active activities performed by the candidate users, the first low-span passive interest support degree may indicate that the online behavior activity big data of the candidate users generated by the base user interest analysis sub-network has an interest prediction value of the low-span passive activities performed by the candidate users, and the first high-span passive interest support degree may indicate that the online behavior activity big data of the candidate users generated by the base user interest analysis sub-network has an interest prediction value of the high-span passive activities performed by the candidate users.

Therefore, according to the embodiment of the application, when the plurality of user interest dimensions are active user interest dimensions, low-span passive user interest dimensions and high-span passive user interest dimensions, the active interest influence connected variable corresponding to the active user interest dimensions, the low-span passive interest influence connected variable corresponding to the low-span passive user interest dimensions, the high-span passive interest influence connected variable corresponding to the high-span passive user interest dimensions and the plurality of user preference connected variables are output according to the basic user interest analysis sub-network, a first active interest support degree of the candidate user for the candidate user online behavior activity large data is generated based on the active interest influence connected variable and the plurality of user preference connected variables, a first low-span passive interest support degree of the candidate user for the candidate user online behavior activity large data is generated based on the low-span passive interest influence connected variable and the plurality of user preference connected variables, and a first high-span passive interest support degree of the candidate user for the candidate user online behavior activity large data is generated based on the high-span passive interest influence connected variable and the plurality of user preference connected variables, so that the first active interest support degree, the first active interest support degree and the first active interest support degree of the candidate user for the candidate user online behavior large data on the candidate user online behavior activity large data are generated based on the high-span passive user interest support degree. In addition, when the user interest dimension changes, the adaptability of the basic user interest analysis sub-network can be adjusted, the basic user interest analysis sub-network with the similar model structure is compatible with a plurality of user interest dimensions, and basic user interest point thermodynamic diagrams corresponding to different user interest dimensions are generated, so that the training difficulty of the basic user interest analysis sub-network is reduced.

Wherein the plurality of user interest dimensions comprise active user interest dimensions and passive user interest dimensions; the plurality of user interest prediction units comprise a first active user interest prediction unit corresponding to an active user interest dimension and a passive user interest prediction unit corresponding to a passive user interest dimension; the expanded user interest point thermodynamic diagram comprises a second active interest support degree and a second passive interest support degree; the second active interest support degree is user interest prediction data generated by the first active user interest prediction unit; the second passive interest support degree is the user interest prediction data generated by the passive user interest prediction unit.

In some embodiments, the following processes 1311-1314 may be included, and processes 1311-1314 are one specific embodiment of Process103 described above.

A Process1311, which loads model loading source data to a first active user interest prediction unit and a passive user interest prediction unit respectively;

the first active user interest prediction unit comprises a second variable loading unit and a second active interest prediction unit, and the passive user interest prediction unit comprises a third variable loading unit and a second passive interest prediction unit.

The Process1312 is used for respectively performing feature embedding aggregation on the model loading source data according to the second variable loading unit and the third variable loading unit, and determining a second user preference member variable of the candidate user generated by the second variable loading unit for the online behavior activity big data of the candidate user and a third user preference member variable of the candidate user generated by the third variable loading unit for the online behavior activity big data of the candidate user;

the cloud AI deployment system may refer to the description of performing feature embedding aggregation on the model loading source data according to the first variable loading unit, and details are not repeated here.

The Process1313 loads the second user preference member variable to a second active interest prediction unit, performs interest prediction on the second user preference member variable according to the second active interest prediction unit, determines a second active interest prediction variable, and obtains a second active interest support degree corresponding to the active user interest dimension based on the second active interest prediction variable;

the second active interest predictor variable is composed of a third active interest predictor variable value and a fourth active interest predictor variable value, the third active interest predictor variable value can represent that the online behavior activity big data of the candidate user has the interest prediction value of the active activity, and the fourth active interest predictor variable value can represent that the online behavior activity big data of the candidate user does not have the interest prediction value of the active activity. Therefore, the cloud-based AI deployment system may use the third active interest predictor variable value as a second active interest support degree of the candidate user for the online behavior activity big data of the candidate user.

And the Process1314 loads the third user preference member variable to the second passive interest prediction unit, performs interest prediction on the third user preference member variable according to the second passive interest prediction unit, determines a second passive interest prediction variable, and obtains a second passive interest support degree corresponding to the passive user interest dimension based on the second passive interest prediction variable.

The second passive interest predictor variable is composed of a third passive interest predictor variable value and a fourth passive interest predictor variable value, the third passive interest predictor variable value can represent that the large data of the online behavior activity of the candidate user has the interest predictor of the passive activity, and the fourth passive interest predictor variable value can represent that the large data of the online behavior activity of the candidate user does not have the interest predictor of the passive activity. Therefore, the cloud AI deployment system can use the third passive interest prediction variable value as a second passive interest support degree of the candidate user for the online behavior activity big data of the candidate user.

Wherein the second active interest prediction unit and the second passive interest prediction unit may be multi-layer perceptrons.

The cloud AI deployment system can acquire a second active interest prediction variable corresponding to the second user preference member variable based on the second active interest prediction unit, and determine a second active interest support degree based on the second active interest prediction variable; the cloud AI deployment system can obtain a second passive interest prediction variable corresponding to the target passive preference feature based on the second passive interest prediction unit, and determine a second passive interest support degree based on the second passive interest prediction variable. The second active interest support degree and the second passive interest support degree can be collectively referred to as an extended user interest point thermodynamic diagram.

In some embodiments, the cloud AI deployment system may use a product of the second active interest support degree and the second passive interest support degree as the active and passive interest support degree generated by the extended user interest analysis sub-network, where the active and passive interest support degree may indicate that the online behavior activity big data of the candidate user generated by the extended user interest analysis sub-network has the interest prediction values of the active activity and the passive activity performed by the candidate user. For example, the cloud AI deployment system may use the active and passive interest support degree as one of the plurality of user interest dimensions, and accordingly, the cloud AI deployment system may use a product of the first active interest support degree and the first passive interest support degree as the active and passive interest support degree generated by the base user interest analysis sub-network. The second active interest support degree may indicate that the online activity big data of the candidate user generated by the extended user interest analysis sub-network has an interest prediction value of the active activity performed by the candidate user, and the second passive interest support degree may indicate that the online activity big data of the candidate user generated by the extended user interest analysis sub-network has an interest prediction value of the passive activity performed by the candidate user. The second passive interest support degree may represent an interest prediction value from active activity to passive activity, and the active and passive interest support degree may represent an interest prediction value from exposure to active activity to passive activity.

Therefore, when the plurality of user interest dimensions are active user interest dimensions and passive user interest dimensions, the second preference connected variable corresponding to the active user interest dimensions and the third preference connected variable corresponding to the passive user interest dimensions are output according to the expanded user interest analysis sub-network, then the second active interest support degree of the candidate user for the candidate user online behavior activity big data is generated based on the second preference connected variable, and the second passive interest support degree of the candidate user for the candidate user online behavior activity big data is generated based on the third preference connected variable, so that the accuracy of predicting the second active interest support degree and the second passive interest support degree is improved.

The number of the expanded user interest analysis sub-networks is multiple, and the multiple expanded user interest analysis sub-networks comprise expanded user interest analysis sub-networks Mi and expanded user interest analysis sub-networks Mj; wherein i and j are both positive integers less than or equal to the number of expanded user interest analysis subnetworks; the plurality of user interest dimensions include an active user interest dimension, a low span passive user interest dimension, and a high span passive user interest dimension; the plurality of user interest prediction units comprise a second active user interest prediction unit and a first low-span passive user interest prediction unit in the extended user interest analysis sub-network Mi, and a second low-span passive user interest prediction unit and a high-span passive user interest prediction unit in the extended user interest analysis sub-network Mj; the second active user interest prediction unit corresponds to the active user interest dimension, the first low-span passive user interest prediction unit and the second low-span passive user interest prediction unit correspond to the low-span passive user interest dimension, and the high-span passive user interest prediction unit corresponds to the high-span passive user interest dimension; the expanded user interest point thermodynamic diagram comprises a third active interest support degree, a second low-span passive interest support degree, a third low-span passive interest support degree and a second high-span passive interest support degree; the third active interest support degree is user interest prediction data generated by the second active user interest prediction unit; the second low-span passive interest support degree is user interest prediction data generated by the first low-span passive user interest prediction unit; the third low-span passive interest support degree is user interest prediction data generated by the second low-span passive user interest prediction unit; the second high-span passive interest support degree is user interest prediction data generated by the high-span passive user interest prediction unit. In some embodiments, processes 1321-1324 below may be included, and processes 1321-1324 are one specific embodiment of Process103 described above.

The Process1321 is used for loading the model loading source data into a second active user interest prediction unit, a first low-span passive user interest prediction unit, a second low-span passive user interest prediction unit and a high-span passive user interest prediction unit respectively;

the second active user interest prediction unit comprises a fourth variable loading unit, and the first low-span passive user interest prediction unit comprises a fifth variable loading unit; the second low-span passive user interest prediction unit comprises a sixth variable loading unit, and the high-span passive user interest prediction unit comprises a seventh variable loading unit.

A Process1322, which performs feature embedding aggregation on the model loading source data according to a fourth variable loading unit, a fifth variable loading unit, a sixth variable loading unit and a seventh variable loading unit, and determines a fourth user preference member variable of the candidate user for the online behavior activity big data of the candidate user, which is generated by the fourth variable loading unit, a fifth user preference member variable of the candidate user for the online behavior activity big data of the candidate user, which is generated by the fifth variable loading unit, a sixth user preference member variable of the candidate user for the online behavior activity big data of the candidate user, which is generated by the sixth variable loading unit, and a seventh user preference member variable of the candidate user for the online behavior activity big data of the candidate user, which is generated by the seventh variable loading unit;

the implementation manner that the cloud AI deployment system performs feature embedding aggregation on the model loading source data according to the fourth variable loading unit, the fifth variable loading unit, the sixth variable loading unit and the seventh variable loading unit may refer to the description of performing feature embedding aggregation on the model loading source data according to the first variable loading unit, and is not described here again.

A Process1323, obtaining, in a second active user interest prediction unit, a third active interest support corresponding to the active user interest dimension according to a fourth user preference member variable, and obtaining, in a first low-span passive user interest prediction unit, a second low-span passive interest support corresponding to the low-span passive user interest dimension according to a fifth user preference member variable;

for example, the cloud AI deployment system may load a fourth user preference member variable into the third active interest prediction unit, perform interest prediction on the fourth user preference member variable according to the third active interest prediction unit, determine a third active interest prediction variable, and obtain a third active interest support degree corresponding to the active user interest dimension based on the third active interest prediction variable. The second active user interest prediction unit further comprises a third active interest prediction unit, and the first low-span passive user interest prediction unit further comprises a second low-span passive interest prediction unit. In some embodiments, the cloud AI deployment system may load the fifth user preference member variable to the second low-span passive interest prediction unit, perform interest prediction on the fifth user preference member variable according to the second low-span passive interest prediction unit, determine the second low-span passive interest prediction variable, and obtain a second low-span passive interest support corresponding to the low-span passive user interest dimension based on the second low-span passive interest prediction variable.

The third active interest prediction variable is composed of a fifth active interest prediction variable value and a sixth active interest prediction variable value, the fifth active interest prediction variable value can represent that the online behavior activity big data of the candidate user has an interest prediction value of active activity, and the sixth active interest prediction variable value can represent that the online behavior activity big data of the candidate user does not have an interest prediction value of active activity. Therefore, the cloud AI deployment system may use the fifth active interest predictor variable value as a third active interest support degree of the candidate user for the online behavior activity big data of the candidate user.

The second low-span passive interest predictive variable is composed of a third low-span passive interest predictive variable value and a fourth low-span passive interest predictive variable value, the third low-span passive interest predictive variable value can represent an interest predictive value of the candidate user on-line behavior activity big data with low-span passive activity, and the fourth low-span passive interest predictive variable value can represent an interest predictive value of the candidate user on-line behavior activity big data without low-span passive activity. Therefore, the cloud-based AI deployment system can use the third low-span passive interest predictor variable value as the second low-span passive interest support degree of the candidate user for the online behavior activity big data of the candidate user.

The Process1324, in the second low-span passive user interest prediction unit, obtains a third low-span passive interest support corresponding to the low-span passive user interest dimension according to the sixth user preference member variable, and in the high-span passive user interest prediction unit, obtains a second high-span passive interest support corresponding to the high-span passive user interest dimension according to the seventh user preference member variable.

For example, the cloud AI deployment system may load a sixth user preference member variable to a third low-span passive interest prediction unit, perform interest prediction on the sixth user preference member variable according to the third low-span passive interest prediction unit, determine a third low-span passive interest prediction variable, and obtain a third low-span passive interest support degree corresponding to the low-span passive user interest dimension based on the third low-span passive interest prediction variable. The second low-span passive user interest prediction unit further comprises a third low-span passive interest prediction unit, and the high-span passive user interest prediction unit further comprises a second high-span passive interest prediction unit. In some embodiments, the cloud AI deployment system may load the seventh user preferred member variable to the second high-span passive interest prediction unit, perform interest prediction on the seventh user preferred member variable according to the second high-span passive interest prediction unit, determine the second high-span passive interest prediction variable, and obtain, based on the second high-span passive interest prediction variable, the second high-span passive interest support degree corresponding to the high-span passive user interest dimension.

The third low-span passive interest predictive variable is composed of a fifth low-span passive interest predictive variable value and a sixth low-span passive interest predictive variable value, the fifth low-span passive interest predictive variable value can represent an interest predictive value of the online behavior activity big data of the candidate user with the low-span passive activity, and the sixth low-span passive interest predictive variable value can represent an interest predictive value of the online behavior activity big data of the candidate user without the low-span passive activity. Therefore, the cloud-based AI deployment system may use the fifth low-span passive interest predictor variable value as the third low-span passive interest support degree of the candidate user for the online behavior activity big data of the candidate user.

The second high-span passive interest predictor variable is composed of a third high-span passive interest predictor variable value and a fourth high-span passive interest predictor variable value, the third high-span passive interest predictor variable value can represent that the online behavior activity big data of the candidate user has the interest prediction value of the high-span passive activity, and the fourth high-span passive interest predictor variable value can represent that the online behavior activity big data of the candidate user does not have the interest prediction value of the high-span passive activity. Therefore, the cloud-based AI deployment system may use the third high-span passive interest predictor variable value as the second high-span passive interest support degree of the candidate user for the online behavior activity big data of the candidate user.

The third active interest prediction unit, the second low-span passive interest prediction unit, the third low-span passive interest prediction unit, and the second high-span passive interest prediction unit may be multi-layer perceptrons.

The cloud AI deployment system can collectively refer to the third active interest support degree, the second low-span passive interest support degree, the third low-span passive interest support degree and the second high-span passive interest support degree as the extended user interest point thermodynamic diagram.

The third active interest support degree may represent that the online behavior activity big data of the candidate user generated by the extended user interest analysis sub-network has an interest prediction value of the active activity performed by the candidate user, the second low-span passive interest support degree and the third low-span passive interest support degree may represent that the online behavior activity big data of the candidate user generated by the extended user interest analysis sub-network has an interest prediction value of the low-span passive activity performed by the candidate user, and the second high-span passive interest support degree may represent that the online behavior activity big data of the candidate user generated by the extended user interest analysis sub-network has an interest prediction value of the high-span passive activity performed by the candidate user.

Therefore, according to the embodiment of the application, when the plurality of user interest dimensions are an active user interest dimension, a low-span passive user interest dimension and a high-span passive user interest dimension, a fourth preference connected variable corresponding to the active user interest dimension, a fifth preference connected variable corresponding to the low-span passive user interest dimension, a sixth preference connected variable corresponding to the low-span passive user interest dimension and a seventh preference connected variable corresponding to the high-span passive user interest dimension are output according to the extended user interest analysis sub-network, a third active interest support degree of the candidate user for the candidate user online behavior activity big data is generated based on the fourth preference connected variable, a second low-span passive interest support degree of the candidate user for the candidate user online behavior activity big data is generated based on the fifth preference connected variable, a third low passive interest support degree of the candidate user for the candidate user online behavior activity big data is generated based on the sixth preference connected variable, a second high-span passive interest support degree of the candidate user for the candidate user online behavior big data is generated based on the seventh preference connected variable, and accordingly the third active interest support degree and the third passive interest support degree are increased. In addition, when the user interest dimension changes, the method and the device can adjust the adaptability of the expanded user interest analysis sub-network, the expanded user interest analysis sub-network with the similar model structure is compatible with a plurality of user interest dimensions, and expanded user interest point thermodynamic diagrams corresponding to different user interest dimensions are generated, so that the training difficulty of the expanded user interest analysis sub-network is reduced.

In some embodiments, the training step for calibrating the user interest analysis model described above is described below, and may include the following processes 201 to 204:

the Process201 is used for obtaining a target training interest analysis model and prior user online behavior activity data corresponding to prior users, and determining prior model loading source data based on the prior users and the prior user online behavior activity data;

wherein the target training interest analysis model comprises a first target training interest analysis sub-network and a second target training interest analysis sub-network; the first target training interest analysis sub-network and the second target training interest analysis sub-network are obtained by conducting AI training according to a plurality of user interest dimensions.

The cloud AI deployment system may refer to the description of generating model loading source data based on the candidate user and the online behavior activity big data of the candidate user, which is not described herein again.

The Process202 is used for performing behavior preference variable mining on the priori model loading source data according to a first target training interest analysis sub-network, determining a priori user preference connected variable and a priori interest influence connected variable corresponding to a plurality of user interest dimensions respectively, and determining a first priori user interest point thermodynamic diagram corresponding to the user interest dimensions based on the priori user preference connected variable and the prior interest influence connected variables;

for the implementation that the cloud AI deployment system outputs the first prior user interest point thermodynamic diagram according to the first target training interest analysis subnetwork, reference may be made to the description of outputting the basic user interest point thermodynamic diagram according to the basic user interest analysis subnetwork, which is not described herein again.

The Process203 is used for respectively loading the priori model loading source data to a plurality of target training interest prediction units in a second target training interest analysis sub-network, and obtaining a second priori user interest point thermodynamic diagram corresponding to a plurality of user interest dimensions and generated by the second target training interest analysis sub-network based on the priori user interest prediction data respectively generated by each target training interest prediction unit;

for the implementation that the cloud AI deployment system outputs the second prior user interest point thermodynamic diagram according to the second target training interest analysis sub-network, reference may be made to the description of outputting the extended user interest point thermodynamic diagram according to the extended user interest analysis sub-network, which is not described herein again.

The Process204 fuses the first prior user interest point thermodynamic diagram and the second prior user interest point thermodynamic diagram in the target training interest analysis model, determines a prior user interest prediction thermodynamic diagram of a prior user for online behavior activity data of the prior user, performs model weight tuning optimization on the target training interest analysis model according to the first prior user interest point thermodynamic diagram, the second prior user interest point thermodynamic diagram, prior interest labeling data of the prior user for online behavior activity data of the prior user, and the prior user interest prediction thermodynamic diagram, and takes the target training interest analysis model after the model weight tuning optimization as a calibrated user interest analysis model.

For example, the cloud AI deployment system may perform interest confidence mean transformation on a first prior user interest point thermodynamic diagram and a second prior user interest point thermodynamic diagram in a target training interest analysis model, and determine a prior user interest prediction thermodynamic diagram of a prior user for prior user online behavior activity data. In some embodiments, the cloud AI deployment system may obtain a first training cost value of the target training interest analysis model according to prior interest annotation data of a prior user for online behavior activity data of the prior user and a prior user interest prediction thermodynamic diagram. In some embodiments, the cloud AI deployment system may obtain a second training cost value of the first target training interest analysis subnetwork according to the first prior user interest point thermodynamic diagram and prior interest labeling data of prior users for the prior user online behavior activity data. In some embodiments, the cloud AI deployment system may obtain a third training cost value of the second target training interest analysis sub-network according to the second prior user interest point thermodynamic diagram and prior interest annotation data of the prior user for the prior user online behavior activity data. In some embodiments, the cloud AI deployment system may obtain a global training cost value of the target training interest analysis model based on the first training cost value, the second training cost value, and the third training cost value, perform model weight tuning optimization on the target training interest analysis model based on the global training cost value, and use the target training interest analysis model after the model weight tuning optimization as the calibrated user interest analysis model. The calibrated user interest analysis model is used for predicting a user interest prediction thermodynamic diagram of the candidate user for the online behavior activity big data of the candidate user.

The first prior user interest point thermodynamic diagram may include a first prior active interest support degree and a first prior passive interest support degree, and the second prior user interest point thermodynamic diagram may include a second prior active interest support degree and a second prior passive interest support degree. The cloud AI deployment system can perform interest confidence coefficient mean conversion on the first prior active interest support degree and the second prior active interest support degree, and determine a prior active user interest prediction thermodynamic diagram of a prior user for online behavior activity data of the prior user; the cloud AI deployment system can perform interest confidence coefficient mean conversion on the first prior passive interest support degree and the second prior passive interest support degree, and determine a prior passive user interest prediction thermodynamic diagram of a prior user for the online behavior activity data of the prior user. In some embodiments, the cloud AI deployment system can obtain an active training cost value of the target training interest analysis model according to prior active interest annotation data of a prior user for online behavior activity data of the prior user and a prior active user interest prediction thermodynamic diagram; the cloud AI deployment system can obtain the passive training cost value of the target training interest analysis model according to the prior passive interest labeling data of the prior user aiming at the online behavior activity data of the prior user and the prior passive user interest prediction thermodynamic diagram. In some embodiments, the cloud-based AI deployment system may obtain a first training cost value of the target training interest analysis model based on the active training cost value and the passive training cost value.

Similarly, the cloud AI deployment system can obtain a first active training cost value of the first target training interest analysis sub-network according to the first priori active interest support degree and the priori active interest marking data; the cloud AI deployment system can obtain a first passive training cost value of the first target training interest analysis sub-network according to the first priori passive interest support degree and the priori passive interest labeling data. In some embodiments, the cloud AI deployment system may obtain a second training cost value for the first target training interest analysis subnetwork based on the first active training cost value and the first passive training cost value. Similarly, the cloud AI deployment system can obtain a second active training cost value of a second target training interest analysis sub-network according to the second priori active interest support degree and the priori active interest marking data; the cloud AI deployment system can obtain a second passive training cost value of a second target training interest analysis sub-network according to the second priori passive interest support degree and the priori passive interest labeling data. In some embodiments, the cloud AI deployment system may obtain a third training cost value for the second target training interest analysis subnetwork based on the second active training cost value and the second passive training cost value.

For example, the first prior user interest point thermodynamic diagram may include a third prior active interest support, a first prior low-span passive interest support, and a first prior high-span passive interest support, and the second prior user interest point thermodynamic diagram may include a fourth prior active interest support, a second prior low-span passive interest support, a third prior low-span passive interest support, and a second prior high-span passive interest support. The implementation manner of obtaining the first training cost value, the second training cost value and the third training cost value according to the third prior active interest support degree, the first prior low-span passive interest support degree, the first prior high-span passive interest support degree, the fourth prior active interest support degree, the second prior low-span passive interest support degree, the third prior low-span passive interest support degree and the second prior high-span passive interest support degree may refer to the description of determining the first training cost value, the second training cost value and the third training cost value according to the first prior active interest support degree, the first prior passive interest support degree, the second prior active interest support degree and the second prior passive interest support degree, which is no longer described herein.

For example, the cloud AI deployment system may also perform model weight tuning optimization on the target training interest analysis model based on the first training cost value, and use the target training interest analysis model after the model weight tuning optimization as the calibrated user interest analysis model. For example, the cloud AI deployment system may further perform model weight tuning optimization on the target training interest analysis model based on the second training cost value and the third training cost value, and use the target training interest analysis model after the model weight tuning optimization as the calibrated user interest analysis model.

For example, the cloud AI deployment system may further perform model weight tuning optimization on the first target training interest analysis sub-network based on the second training cost value, and use the first target training interest analysis sub-network after the model weight tuning optimization as a basic user interest analysis sub-network; and performing model weight tuning optimization on the second target training interest analysis sub-network based on the third training cost value, and taking the second target training interest analysis sub-network after the model weight tuning optimization as an extended user interest analysis sub-network. In some embodiments, the cloud AI deployment system may use the target training interest analysis model after model weight tuning optimization as a calibrated user interest analysis model; the calibration user interest analysis model comprises a basic user interest analysis sub-network and an extended user interest analysis sub-network.

For example, the target training interest analysis model further includes a target training interest fusion unit. In this way, the cloud AI deployment system can obtain a second training cost value of the first target training interest analysis sub-network according to the thermodynamic diagram of the interest points of the first prior user and the prior interest marking data of the prior user for the online behavior activity data of the prior user, perform model weight tuning optimization on the first target training interest analysis sub-network based on the second training cost value, and use the first target training interest analysis sub-network after the model weight tuning optimization as the basic user interest analysis sub-network. In some embodiments, the cloud AI deployment system may obtain a third training cost value of the second target training interest analysis sub-network according to a second prior user interest point thermodynamic diagram and prior interest labeling data of prior users for the online behavior activity data of the prior users, perform model weight tuning optimization on the second target training interest analysis sub-network based on the third training cost value, and use the second target training interest analysis sub-network after the model weight tuning optimization as the extended user interest analysis sub-network. In some embodiments, the cloud AI deployment system may fuse the first prior user interest point thermodynamic diagram and the second prior user interest point thermodynamic diagram to determine a prior fused interest point thermodynamic diagram. In some embodiments, the cloud AI deployment system may load the prior fusion interest point thermodynamic diagrams to the target training interest fusion unit, perform fusion learning on a first prior user interest point thermodynamic diagram and a second prior user interest point thermodynamic diagram in the prior fusion interest point thermodynamic diagrams according to the target training interest fusion unit, and determine prior user interest prediction thermodynamic diagrams of prior users for prior user online behavior activity data. In some embodiments, the cloud AI deployment system may perform model weight tuning optimization on the target training interest fusion unit according to prior interest annotation data of a prior user for online behavior activity data of the prior user and a prior user interest prediction thermodynamic diagram, and use the target training interest fusion unit after the model weight tuning optimization as the interest fusion unit. In some embodiments, the cloud AI deployment system may use the target training interest analysis model after model weight tuning optimization as the calibrated user interest analysis model. The calibrated user interest analysis model comprises an interest fusion unit, a basic user interest analysis sub-network and an extended user interest analysis sub-network.

The number of the first target training interest analysis sub-networks may be P, and the number of the second target training interest analysis sub-networks may be P, where P may be a positive integer greater than 1. The cloud AI deployment system can divide the prior user online behavior activity data into P types, acquire (P-1) type prior user online behavior activity data from the P type prior user online behavior activity data to perform network weight optimization on each first target training interest analysis sub-network, and determine P basic user interest analysis sub-networks after the network weight optimization; and acquiring (P-1) type prior user online behavior activity data from the P type prior user online behavior activity data, performing network weight adjustment on each second target training interest analysis sub-network, and determining P extended user interest analysis sub-networks after the network weight adjustment. The cloud AI deployment system can acquire the prior user online behavior activity data except the x-th type from the P-type prior user online behavior activity data to perform network weight optimization on the x-th first target training interest analysis sub-network; the cloud AI deployment system can acquire the prior user online behavior activity data except the x-th type from the P-type prior user online behavior activity data to perform network weight optimization on the x-th second target training interest analysis sub-network. Wherein x may be a positive integer less than or equal to P. For example, the cloud AI deployment system may obtain the prior user online behavior activity data other than the type 1 from the P-type prior user online behavior activity data to perform network weight tuning on the 1 st first target training interest analysis sub-network; for another example, the cloud AI deployment system may obtain the prior online behavior activity data of the users other than the 2 nd type from the P-type prior online behavior activity data to perform network weight tuning on the 2 nd second target training interest analysis sub-network.

In some embodiments, the cloud AI deployment system may output a first prior user interest point thermodynamic diagram corresponding to a plurality of user interest dimensions according to the P base user interest analysis subnetworks after network weight tuning; the cloud AI deployment system can output a second prior user interest point thermodynamic diagram corresponding to the user interest dimensions according to the P expanded user interest analysis sub-networks after network weight adjustment. The cloud AI deployment system can output a first prior user interest point thermodynamic diagram corresponding to the behavior activity data on the x-th type prior user line according to the x-th basic user interest analysis sub-network, further splice the P-type first prior user interest point thermodynamic diagrams generated by the P-th basic user interest analysis sub-networks, and determine the first prior user interest point thermodynamic diagram corresponding to the behavior activity data on the P-type prior user line (namely, the first prior user interest point thermodynamic diagrams corresponding to a plurality of user interest dimensions); the cloud AI deployment system can output a second prior user interest point thermodynamic diagram corresponding to the behavior activity data on the x-th type prior user line according to the x-th extended user interest analysis sub-network, and further splices the P-type second prior user interest point thermodynamic diagrams generated by the P-type extended user interest analysis sub-networks to determine the second prior user interest point thermodynamic diagram corresponding to the behavior activity data on the P-type prior user line (namely, the second prior user interest point thermodynamic diagrams corresponding to the user interest dimensions). For example, the cloud AI deployment system may output a first prior user interest point thermodynamic diagram corresponding to the online behavior activity data of the type 1 prior user according to the 1 st base user interest analysis subnetwork; for another example, the cloud AI deployment system may output a second prior user interest point thermodynamic diagram corresponding to the online behavior activity data of the type 2 prior user according to the 2 nd expanded user interest analysis subnetwork.

The network labels of the P basic user interest analysis sub-networks are the same, and the learning data of the P basic user interest analysis sub-networks are different, so that the network weights trained by the P basic user interest analysis sub-networks are different; the network labels of the P extended user interest analysis subnetworks are the same, and the network weights trained by the P extended user interest analysis subnetworks are different because the learning data of the P extended user interest analysis subnetworks are different.

The cloud AI deployment system can fuse first prior user interest point thermodynamic diagrams of the P basic user interest analysis sub-networks, determine a first prior fusion interest point thermodynamic diagram, fuse second prior user interest point thermodynamic diagrams of the P extended user interest analysis sub-networks, and determine a second prior fusion interest point thermodynamic diagram. In some embodiments, the cloud AI deployment system may load the first prior fusion interest point thermodynamic diagram and the second prior fusion interest point thermodynamic diagram to the target training interest fusion unit, and perform network weight optimization on the target training interest fusion unit according to the first prior fusion interest point thermodynamic diagram, the second prior fusion interest point thermodynamic diagram and prior interest labeling data of prior users for prior user online behavior activity data, so as to determine the interest fusion unit.

The cloud AI deployment system can conduct network weight optimization on the target training interest fusion units according to the first priori user interest point thermodynamic diagrams and the second priori user interest point thermodynamic diagrams. For example, the number of the target training interest fusion units may be multiple, and the cloud AI deployment system may perform network weight optimization on the multiple target training interest fusion units according to the first prior user interest point thermodynamic diagram and the second prior user interest point thermodynamic diagram. The implementation manner of the cloud AI deployment system performing network weight optimization on the multiple target training interest fusion units according to the first prior user interest point thermodynamic diagram and the second prior user interest point thermodynamic diagram may refer to the description of performing network weight optimization on the first target training interest analysis sub-network or the second target training interest analysis sub-network by loading the source data according to the prior model, which is not described herein again.

Correspondingly, when a user interest prediction thermodynamic diagram of a candidate user for big data of online behavior activities of the candidate user is predicted through a calibration user interest analysis model comprising an interest fusion unit, P basic user interest analysis sub-networks and P extended user interest analysis sub-networks, the cloud AI deployment system needs to perform interest confidence coefficient mean transformation on basic user interest point thermodynamic diagrams respectively generated by the P basic user interest analysis sub-networks and determine a first user average interest point thermodynamic diagram corresponding to the P basic user interest analysis sub-networks; the cloud AI deployment system needs to perform interest confidence mean value conversion on the extended user interest point thermodynamic diagrams respectively generated by the P extended user interest analysis sub-networks, and determine second user average interest point thermodynamic diagrams corresponding to the P extended user interest analysis sub-networks.

Therefore, in the embodiment of the application, the prior model can load source data to perform network weight optimization on the target training interest analysis model, that is, a first prior user interest point thermodynamic diagram is output according to a first target training interest analysis sub-network in the target training interest analysis model, a second prior user interest point thermodynamic diagram is output according to a second target training interest analysis sub-network in the target training interest analysis model, and then the target training interest analysis model is subjected to network weight optimization based on the first prior user interest point thermodynamic diagram and the second prior user interest point thermodynamic diagram to determine the calibrated user interest analysis model. The calibrated user interest analysis model can be used for determining a user interest prediction thermodynamic diagram of a candidate user for the online behavior activity big data of the candidate user, and the accuracy of the predicted user interest prediction thermodynamic diagram can be improved according to the calibrated user interest analysis model obtained by adjusting the network weight of the target training interest analysis model.

In some embodiments, cloud AI deployment 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 suitable actions and processes based on a program stored in the machine-readable storage medium 120, such as program instructions related to the big data mining method for serving user interest 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 when the computer-executable instructions are executed by a processor, the computer-readable storage medium is used for implementing the big data mining method for serving user interest analysis according to any one of the above embodiments.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solution of the present invention, and not to limit the same; while the invention has been described in detail and with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and these modifications or substitutions do not depart from the spirit of the corresponding technical solutions of the embodiments of the present invention.

Claims

1. A user interest analysis method for big data mining, wherein the method is executed by the cloud AI deployment system, and the method comprises the following steps:

acquiring a target training interest analysis model and prior user online behavior activity data corresponding to a prior user, and determining prior model loading source data based on the prior user and the prior user online behavior activity data; the target training interest analysis model comprises a first target training interest analysis sub-network and a second target training interest analysis sub-network; the first target training interest analysis sub-network and the second target training interest analysis sub-network are obtained by AI training according to a plurality of user interest dimensions respectively;

performing behavior preference variable mining on the prior model loading source data according to the first target training interest analysis sub-network, determining prior user preference connected variables and prior interest influence connected variables corresponding to the multiple user interest dimensions respectively, and determining a first prior user interest point thermodynamic diagram corresponding to the multiple user interest dimensions based on the prior user preference connected variables and the multiple prior interest influence connected variables;

loading the prior model loading source data to a plurality of target training interest prediction units in the second target training interest analysis sub-network respectively, and obtaining a second prior user interest point thermodynamic diagram corresponding to the user interest dimensions generated by the second target training interest analysis sub-network based on prior user interest prediction data generated by each target training interest prediction unit respectively;

in the target training interest analysis model, fusing the first prior user interest point thermodynamic diagram and the second prior user interest point thermodynamic diagram, determining a prior user interest prediction thermodynamic diagram of the prior user for the prior user online behavior activity data, performing model weight tuning optimization on the target training interest analysis model according to the first prior user interest point thermodynamic diagram, the second prior user interest point thermodynamic diagram, prior interest labeling data of the prior user for the prior user online behavior activity data and the prior user interest prediction thermodynamic diagram, and taking the target training interest analysis model after the model weight tuning optimization as a calibrated user interest analysis model; the calibrated user interest analysis model is used for predicting a user interest prediction thermodynamic diagram of the candidate user for the online behavior activity big data of the candidate user.

2. The user interest analysis method for big data mining according to claim 1, wherein the fusing the first prior user interest point thermodynamic diagram and the second prior user interest point thermodynamic diagram in the target training interest analysis model, determining a prior user interest prediction thermodynamic diagram of the prior user for the prior user online behavior activity data, performing model weight tuning optimization on the target training interest analysis model according to the first prior user interest point thermodynamic diagram, the second prior user interest point thermodynamic diagram, the prior interest annotation data of the prior user for the prior user online behavior activity data, and the prior user interest prediction thermodynamic diagram, and taking the model weight tuning optimized target training interest analysis model as a calibration user interest analysis model, includes:

performing interest confidence mean conversion on the first prior user interest point thermodynamic diagram and the second prior user interest point thermodynamic diagram in the target training interest analysis model, and determining a prior user interest prediction thermodynamic diagram of the prior user for the prior user online behavior activity data;

obtaining a first training cost value of the target training interest analysis model according to the prior interest marking data of the prior user aiming at the online behavior activity data of the prior user and the prior user interest prediction thermodynamic diagram;

obtaining a second training cost value of the first target training interest analysis sub-network according to the first prior user interest point thermodynamic diagram and prior interest marking data of the prior user for the prior user online behavior activity data;

obtaining a third training cost value of the second target training interest analysis sub-network according to the second prior user interest point thermodynamic diagram and prior interest marking data of the prior user for the prior user online behavior activity data;

obtaining a global training cost value of the target training interest analysis model based on the first training cost value, the second training cost value and the third training cost value, performing model weight tuning optimization on the target training interest analysis model based on the global training cost value, and taking the target training interest analysis model after the model weight tuning optimization as a calibrated user interest analysis model.

3. The user interest analysis method for big data mining according to claim 1, wherein the target training interest analysis model further comprises a target training interest fusion unit;

the target training interest analysis model is subjected to model weight tuning optimization according to the first prior user interest point thermodynamic diagram, the second prior user interest point thermodynamic diagram, the prior interest labeling data of the prior user for the prior user online behavior activity data and the prior user interest prediction thermodynamic diagram, and the target training interest analysis model after model weight tuning optimization is used as a calibrated user interest analysis model, and the method comprises the following steps of:

obtaining a second training cost value of the first target training interest analysis sub-network according to the first prior user interest point thermodynamic diagram and prior interest labeling data of the prior user for the online behavior activity data of the prior user, performing model weight tuning optimization on the first target training interest analysis sub-network based on the second training cost value, and taking the first target training interest analysis sub-network after the model weight tuning optimization as a basic user interest analysis sub-network;

obtaining a third training cost value of the second target training interest analysis sub-network according to the second prior user interest point thermodynamic diagram and prior interest labeling data of the prior user for the prior user online behavior activity data, performing model weight tuning optimization on the second target training interest analysis sub-network based on the third training cost value, and taking the second target training interest analysis sub-network after the model weight tuning optimization as an extended user interest analysis sub-network;

fusing the first prior user interest point thermodynamic diagram and the second prior user interest point thermodynamic diagram to determine a prior fused interest point thermodynamic diagram;

loading the priori fusion interest point thermodynamic diagrams to the target training interest fusion unit, performing fusion learning on the first and second priori user interest point thermodynamic diagrams in the priori fusion interest point thermodynamic diagrams according to the target training interest fusion unit, and determining a priori user interest prediction thermodynamic diagram of the priori user for the prior user online behavior activity data;

performing model weight tuning optimization on the target training interest fusion unit according to prior interest labeling data of the prior user aiming at the online behavior activity data of the prior user and the prior user interest prediction thermodynamic diagram, and taking the target training interest fusion unit subjected to model weight tuning optimization as an interest fusion unit;

taking the target training interest analysis model after model weight tuning optimization as a calibrated user interest analysis model; the calibrated user interest analysis model comprises the interest fusion unit, the basic user interest analysis sub-network and the extended user interest analysis sub-network.

4. The user interest analysis method for big data mining according to claim 1, wherein the method further comprises:

obtaining a calibrated user interest analysis model and a candidate user corresponding to the candidate user online behavior activity big data, and generating model loading source data based on the candidate user and the candidate user online behavior activity big data; the calibrated user interest analysis model comprises a basic user interest analysis sub-network and an extended user interest analysis sub-network; the basic user interest analysis sub-network and the extended user interest analysis sub-network are obtained by AI training according to a plurality of user interest dimensions respectively, wherein the cloud AI deployment system acquires a candidate user corresponding to the online behavior activity big data of the candidate user, and takes a user tag of the candidate user, a data tag of the online behavior activity big data of the candidate user and an online behavior activity data set of the user corresponding to the candidate user as model loading source data for calibrating a user interest analysis model, wherein the user tag is used for representing the candidate user, and the data tag is used for representing the online behavior activity data of the user;

performing behavior preference variable mining on the model loading source data according to the basic user interest analysis sub-network, determining user preference connected variables and interest influence connected variables corresponding to the user interest dimensions respectively, and determining a basic user interest point thermodynamic diagram corresponding to the user interest dimensions based on the user preference connected variables and the interest influence connected variables;

respectively loading the model loading source data to a plurality of user interest prediction units in the extended user interest analysis sub-network, and obtaining an extended user interest point thermodynamic diagram corresponding to a plurality of user interest dimensions generated by the extended user interest analysis sub-network based on user interest prediction data respectively generated by each user interest prediction unit;

in the calibration user interest analysis model, fusing the basic user interest point thermodynamic diagrams and the expanded user interest point thermodynamic diagrams, and determining user interest prediction thermodynamic diagrams of the candidate users for the candidate user online behavior activity big data;

determining a credible user interest point sequence of the candidate user based on the user interest prediction thermodynamic diagram;

5. The user interest analysis method for big data mining according to claim 4, wherein the performing behavior preference variable mining on the model loading source data according to the basic user interest analysis sub-network, determining user preference connected variables and interest influence connected variables corresponding to the user interest dimensions respectively, and determining a basic user interest point thermodynamic diagram corresponding to the user interest dimensions based on the user preference connected variables and the interest influence connected variables comprises:

loading the model loading source data to the base user interest analysis sub-network; the basic user interest analysis sub-network comprises a first variable loading unit, a plurality of user preference analysis units and preference influence analysis units corresponding to the user interest dimensions respectively;

performing feature embedding aggregation on the model loading source data according to the first variable loading unit, and determining a first user preference member variable of the candidate user generated by the first variable loading unit for the online behavior activity big data of the candidate user;

loading the first user preference member variable to the plurality of user preference analysis units respectively, performing connected node analysis on the first user preference member variable according to the plurality of user preference analysis units respectively, and determining user preference connected variables generated by the user preference analysis units respectively;

loading the first user preference member variable to a plurality of preference influence analysis units respectively, carrying out preference influence analysis on the first user preference member variable according to the preference influence analysis units respectively, and determining interest influence connected variables generated by the preference influence analysis units respectively;

determining a basic user interest point thermodynamic diagram corresponding to the user interest dimensions based on a plurality of user preference connected variables and a plurality of interest influence connected variables;

wherein the plurality of user interest dimensions comprises an active user interest dimension and a passive user interest dimension; the plurality of preference impact analysis units comprise an active preference impact analysis unit indicated by the active user interest dimension and a passive preference impact analysis unit indicated by the passive user interest dimension; the plurality of interest influence connected variables comprise active interest influence connected variables and passive interest influence connected variables;

the loading the first user preference member variable to a plurality of preference influence analysis units respectively, performing preference influence analysis on the first user preference member variable according to the preference influence analysis units respectively, and determining interest influence connected variables generated by the preference influence analysis units respectively, includes:

loading the first user preference member variable to the active preference influence analysis unit, performing preference influence analysis on the first user preference member variable according to the active preference influence analysis unit, and determining the active interest influence connected variable generated by the active preference influence analysis unit;

loading the first user preference member variable to the passive preference influence analysis unit, performing preference influence analysis on the first user preference member variable according to the passive preference influence analysis unit, and determining the passive interest influence connected variable generated by the passive preference influence analysis unit;

the active interest influence communication variable comprises active interest influence communication coefficients respectively corresponding to the user preference analysis units; the passive interest influence connected variable comprises passive interest influence connected coefficients respectively corresponding to the user preference analysis units; the basic user interest analysis sub-network further comprises a first active interest prediction unit and a first passive interest prediction unit; the basic user interest point thermodynamic diagram comprises a first active interest support degree and a first passive interest support degree; the determining a basic user interest point thermodynamic diagram corresponding to the user interest dimensions based on the user preference connected variables and the interest influence connected variables comprises:

based on the active interest influence communication coefficients respectively corresponding to the user preference analysis units, carrying out interest influence communication processing on user preference communication variables respectively generated by the user preference analysis units to determine target active preference characteristics;

loading the target active preference feature to the first active interest prediction unit, performing interest prediction on the target active preference feature according to the first active interest prediction unit, determining a first active interest prediction variable, and obtaining the first active interest support degree corresponding to the active user interest dimension based on the first active interest prediction variable;

based on the passive interest influence communication coefficients respectively corresponding to the user preference analysis units, performing interest influence communication processing on user preference communication variables respectively generated by the user preference analysis units, determining target passive preference features, loading the target passive preference features to the first passive interest prediction unit, performing interest prediction on the target passive preference features according to the first passive interest prediction unit, determining first passive interest prediction variables, and based on the first passive interest prediction variables, obtaining the first passive interest support degrees corresponding to the passive user interest dimensions;

wherein the passive preference impact analysis unit comprises a low span passive preference impact analysis unit and a high span passive preference impact analysis unit; the passive interest influence connected variables comprise low-span passive interest influence connected variables and high-span passive interest influence connected variables;

the loading the first user preference member variable to the passive preference influence analysis unit, performing preference influence analysis on the first user preference member variable according to the passive preference influence analysis unit, and determining the passive interest influence connected variable generated by the passive preference influence analysis unit includes:

loading the first user preference member variable to the low-span passive preference influence analysis unit, performing preference influence analysis on the first user preference member variable according to the low-span passive preference influence analysis unit, and determining the low-span passive interest influence connected variable generated by the low-span passive preference influence analysis unit;

loading the first user preference member variable to the high-span passive preference influence analysis unit, performing preference influence analysis on the first user preference member variable according to the high-span passive preference influence analysis unit, and determining the high-span passive interest influence connected variable generated by the high-span passive preference influence analysis unit;

the low-span passive interest influence connected variable comprises low-span passive interest influence connected coefficients respectively corresponding to the user preference analysis units; the high-span passive interest influence connected variable comprises high-span passive interest influence connected coefficients respectively corresponding to the user preference analysis units; the target passive preference features comprise target low-span passive preference features and target high-span passive preference features; the first passive interest prediction unit comprises a first low-span passive interest prediction unit and a high-span passive interest prediction unit; the first passive interest predictor comprises a first low span passive interest predictor and a high span passive interest predictor; the first passive interest support degree comprises a first low-span passive interest support degree and a first high-span passive interest support degree; the passive user interest dimensions comprise a low-span passive user interest dimension and a high-span passive user interest dimension;

the method comprises the steps of performing interest influence communication processing on user preference connected variables generated by the user preference analysis units respectively based on passive interest influence communication coefficients corresponding to the user preference analysis units respectively, determining target passive preference characteristics, loading the target passive preference characteristics to the first passive interest prediction unit, performing interest prediction on the target passive preference characteristics according to the first passive interest prediction unit, determining first passive interest prediction variables, and obtaining the first passive interest support corresponding to the passive user interest dimension based on the first passive interest prediction variables, and comprises the following steps:

based on the low-span passive interest influence communication coefficients respectively corresponding to the user preference analysis units, carrying out interest influence communication processing on user preference communication variables respectively generated by the user preference analysis units, and determining the target low-span passive preference characteristics;

loading the target low-span passive preference feature to the first low-span passive interest prediction unit, performing interest prediction on the target low-span passive preference feature according to the first low-span passive interest prediction unit, determining a first low-span passive interest prediction variable, and obtaining the first low-span passive interest support degree corresponding to the low-span passive user interest dimension based on the first low-span passive interest prediction variable;

based on the high-span passive interest influence communication coefficients respectively corresponding to the user preference analysis units, carrying out interest influence communication processing on user preference communication variables respectively generated by the user preference analysis units, and determining the target high-span passive preference characteristics;

loading the target high-span passive preference feature to the high-span passive interest prediction unit, performing interest prediction on the target high-span passive preference feature according to the high-span passive interest prediction unit, determining the high-span passive interest prediction variable, and obtaining the first high-span passive interest support degree corresponding to the high-span passive user interest dimension based on the high-span passive interest prediction variable.

6. The user interest analysis method for big data mining of claim 4, wherein the plurality of user interest dimensions comprise an active user interest dimension and a passive user interest dimension; the plurality of user interest prediction units comprise a first active user interest prediction unit corresponding to the active user interest dimension and a passive user interest prediction unit corresponding to the passive user interest dimension; the expanded user interest point thermodynamic diagram comprises a second active interest support degree and a second passive interest support degree; the second active interest support degree is user interest prediction data generated by the first active user interest prediction unit; the second passive interest support degree is user interest prediction data generated by the passive user interest prediction unit;

the step of respectively loading the model loading source data to a plurality of user interest prediction units in the extended user interest analysis sub-network, and obtaining an extended user interest point thermodynamic diagram corresponding to the user interest dimensions generated by the extended user interest analysis sub-network based on user interest prediction data respectively generated by each user interest prediction unit, includes:

loading the model loading source data to the first active user interest prediction unit and the passive user interest prediction unit respectively; the first active user interest prediction unit comprises a second variable loading unit and a second active interest prediction unit, and the passive user interest prediction unit comprises a third variable loading unit and a second passive interest prediction unit;

respectively performing feature embedding aggregation on the model loading source data according to the second variable loading unit and the third variable loading unit, and determining a second user preference member variable of the candidate user for the online behavior activity big data of the candidate user generated by the second variable loading unit and a third user preference member variable of the candidate user for the online behavior activity big data of the candidate user generated by the third variable loading unit;

loading the second user preference member variable to the second active interest prediction unit, performing interest prediction on the second user preference member variable according to the second active interest prediction unit, determining a second active interest prediction variable, and obtaining a second active interest support degree corresponding to the active user interest dimension based on the second active interest prediction variable;

and loading the third user preference member variable to the second passive interest prediction unit, performing interest prediction on the third user preference member variable according to the second passive interest prediction unit, determining a second passive interest prediction variable, and obtaining a second passive interest support degree corresponding to the passive user interest dimension based on the second passive interest prediction variable.

7. The user interest analysis method for big data mining according to claim 4, wherein the number of the extended user interest analysis sub-networks is plural, and the plural extended user interest analysis sub-networks include an extended user interest analysis sub-network Mi and an extended user interest analysis sub-network Mj; the i and the j are both positive integers which are less than or equal to the number of the extended user interest analysis sub-networks; the plurality of user interest dimensions include an active user interest dimension, a low-span passive user interest dimension, and a high-span passive user interest dimension; the plurality of user interest prediction units comprise a second active user interest prediction unit and a first low-span passive user interest prediction unit in the extended user interest analysis sub-network Mi, and a second low-span passive user interest prediction unit and a high-span passive user interest prediction unit in the extended user interest analysis sub-network Mj; the second active user interest prediction unit corresponds to the active user interest dimension, the first low-span passive user interest prediction unit and the second low-span passive user interest prediction unit correspond to the low-span passive user interest dimension, and the high-span passive user interest prediction unit corresponds to the high-span passive user interest dimension; the expanded user interest point thermodynamic diagram comprises a third active interest support degree, a second low-span passive interest support degree, a third low-span passive interest support degree and a second high-span passive interest support degree; the third active interest support degree is user interest prediction data generated by the second active user interest prediction unit; the second low-span passive interest support degree is user interest prediction data generated by the first low-span passive user interest prediction unit; the third low-span passive interest support degree is user interest prediction data generated by the second low-span passive user interest prediction unit; the second high-span passive interest support degree is user interest prediction data generated by the high-span passive user interest prediction unit;

loading the model loading source data to the second active user interest prediction unit, the first low-span passive user interest prediction unit, the second low-span passive user interest prediction unit and the high-span passive user interest prediction unit respectively;

the second active user interest prediction unit comprises a fourth variable loading unit, and the first low-span passive user interest prediction unit comprises a fifth variable loading unit; the second low-span passive user interest prediction unit comprises a sixth variable loading unit, and the high-span passive user interest prediction unit comprises a seventh variable loading unit;

feature embedding aggregation is performed on the model loading source data according to the fourth variable loading unit, the fifth variable loading unit, the sixth variable loading unit and the seventh variable loading unit, a fourth user preference member variable of the candidate user for the online behavior activity big data of the candidate user generated by the fourth variable loading unit is determined, a fifth user preference member variable of the candidate user for the online behavior activity big data of the candidate user generated by the fifth variable loading unit is determined, a sixth user preference member variable of the candidate user for the online behavior activity big data of the candidate user generated by the sixth variable loading unit is determined, and a seventh user preference member variable of the candidate user for the online behavior activity big data of the candidate user generated by the seventh variable loading unit is determined;

in the second active user interest prediction unit, obtaining the third active interest support degree corresponding to the active user interest dimension according to the fourth user preference member variable, and in the first low-span passive user interest prediction unit, obtaining the second low-span passive interest support degree corresponding to the low-span passive user interest dimension according to the fifth user preference member variable;

in the second low-span passive user interest prediction unit, the third low-span passive interest support degree corresponding to the low-span passive user interest dimension is obtained according to the sixth user preference member variable, and in the high-span passive user interest prediction unit, the second high-span passive interest support degree corresponding to the high-span passive user interest dimension is obtained according to the seventh user preference member variable.

8. The user interest analysis method for big data mining according to claim 7, wherein the second active user interest prediction unit further comprises a third active interest prediction unit, and the first low-span passive user interest prediction unit further comprises a second low-span passive interest prediction unit;

the obtaining, in the second active user interest prediction unit, the third active interest support degree corresponding to the active user interest dimension according to the fourth user preference member variable, and obtaining, in the first low-span passive user interest prediction unit, the second low-span passive interest support degree corresponding to the low-span passive user interest dimension according to the fifth user preference member variable, include:

loading the fourth user preference member variable to the third active interest prediction unit, performing interest prediction on the fourth user preference member variable according to the third active interest prediction unit, determining a third active interest prediction variable, and obtaining a third active interest support degree corresponding to the active user interest dimension based on the third active interest prediction variable;

loading the fifth user preference member variable to the second low-span passive interest prediction unit, performing interest prediction on the fifth user preference member variable according to the second low-span passive interest prediction unit, determining a second low-span passive interest prediction variable, and obtaining a second low-span passive interest support degree corresponding to the low-span passive user interest dimension based on the second low-span passive interest prediction variable.

9. The user interest analysis method for big data mining of claim 4, wherein the base user interest point thermodynamic diagram includes a first active interest support degree and a first passive interest support degree; the expanded user interest point thermodynamic diagram comprises a second active interest support degree and a second passive interest support degree; the user interest predictive thermodynamic diagrams comprise an active user interest predictive thermodynamic diagram and a passive user interest predictive thermodynamic diagram;

the fusing the basic user interest point thermodynamic diagram and the expanded user interest point thermodynamic diagram in the calibrated user interest analysis model to determine the user interest prediction thermodynamic diagram of the candidate user for the candidate user online behavior activity big data comprises:

performing interest confidence mean conversion on the first active interest support degree and the second active interest support degree in the calibrated user interest analysis model, and determining the active user interest prediction thermodynamic diagram of the candidate user for the candidate user online behavior activity big data;

and performing interest confidence coefficient mean conversion on the first passive interest support degree and the second passive interest support degree, and determining the passive user interest prediction thermodynamic diagram of the candidate user for the online behavior activity big data of the candidate user.

10. A cloud AI deployment 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 user interest analysis method for big data mining according to any one of claims 1 to 9 when running the computer program.