CN115878900A - User online intention analysis method based on artificial intelligence and big data e-commerce platform - Google Patents

User online intention analysis method based on artificial intelligence and big data e-commerce platform Download PDF

Info

Publication number
CN115878900A
CN115878900A CN202211637844.8A CN202211637844A CN115878900A CN 115878900 A CN115878900 A CN 115878900A CN 202211637844 A CN202211637844 A CN 202211637844A CN 115878900 A CN115878900 A CN 115878900A
Authority
CN
China
Prior art keywords
behavior
knowledge point
user online
feedback
information
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Withdrawn
Application number
CN202211637844.8A
Other languages
Chinese (zh)
Inventor
袁广增
王凯
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Kunming Yuanzengbei Information Technology Co ltd
Original Assignee
Kunming Yuanzengbei Information Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Kunming Yuanzengbei Information Technology Co ltd filed Critical Kunming Yuanzengbei Information Technology Co ltd
Priority to CN202211637844.8A priority Critical patent/CN115878900A/en
Publication of CN115878900A publication Critical patent/CN115878900A/en
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The embodiment of the application provides a user online intention analysis method based on artificial intelligence and a big data e-commerce platform, user online behavior data carried in an online intention analysis instruction is loaded into a pre-trained target user online intention prediction model to obtain target user online intention information, the target user online intention information is returned to an e-commerce service system corresponding to the online intention analysis instruction, e-commerce resource pushing data fed back by the e-commerce service system according to the target user online intention information is obtained, pushing processing is carried out on a user online page corresponding to the user online behavior data based on the e-commerce resource pushing data, and therefore the user online intention information can be presumed and relevant e-commerce resource data can be loaded to the user online intention information, and the resource pushing matching degree of the user online e-commerce service is improved.

Description

User online intention analysis method based on artificial intelligence and big data e-commerce platform
Technical Field
The embodiment of the application relates to the technical field of artificial intelligence, in particular to a user online intention analysis method based on artificial intelligence and a big data e-commerce platform.
Background
The current e-commerce page service gradually becomes the subject of the development of information technology in the current era, and an e-commerce recommendation system based on artificial intelligence decision gradually becomes the key point of the research recommendation system service of each large e-commerce platform. The e-commerce recommendation system is widely applied to an e-commerce platform and used for recommending e-commerce resource data according to the preference of an online user so as to improve the information technology driven strategy of the online page experience of the user. Therefore, how to improve the resource pushing matching degree of the online e-commerce service of the user by using artificial intelligence is a direction in which improvement and development are urgently needed in the field.
Disclosure of Invention
In order to overcome at least the above disadvantages in the prior art, an embodiment of the present application is directed to providing an artificial intelligence-based user online intention analysis method and a big data e-commerce platform.
In a first aspect, an embodiment of the present application provides an artificial intelligence-based user online intention analysis method, which is applied to a big data e-commerce platform, and the method includes:
responding to each online intention analysis instruction in the current user online intention analysis task, and acquiring user online behavior data carried in the online intention analysis instruction;
loading the user online behavior data into a pre-trained target user online intention prediction model to obtain target user online intention information, and returning the target user online intention information to the e-commerce service system corresponding to the online intention analysis instruction;
and E-commerce resource pushing data fed back by the E-commerce service system according to the target user online intention information is obtained, and pushing processing is carried out on a user online page corresponding to the user online behavior data based on the E-commerce resource pushing data.
In one possible implementation of the first aspect, the method further comprises:
acquiring staged demand feedback data of the pushed e-commerce resource pushing data in the user online page, performing graph node network generation on the staged demand feedback data, acquiring a corresponding demand feedback graph node network, acquiring a first topic viewpoint feature in a first demand feedback activity in the demand feedback graph node network, determining a corresponding second demand feedback activity based on a first network connection relation of the first demand feedback activity, and acquiring a second topic viewpoint feature in the second demand feedback activity; determining a first prior feedback activity associated with the first topic viewpoint feature and a second prior feedback activity associated with a second topic viewpoint feature based on prior E-commerce behavioral ideographic knowledge points generated by a prior ideographic feature of each prior feedback activity in a prior demand feedback network; establishing a first connected link between the first demand feedback activity and the first prior feedback activity, and establishing a second connected link between the second demand feedback activity and the second prior feedback activity; if the sparse matrix relationship does not exist between the first prior feedback activity and the second prior feedback activity through analysis, constructing a sparse matrix relationship between the first prior feedback activity and the second prior feedback activity based on the first network connection relationship; if the sparse matrix relation exists between the first priori feedback activity and the second priori feedback activity, adjusting the sparse matrix relation based on the first network connection relation; forming an expansion demand feedback graph node network according to the first communication link, the second communication link, the constructed sparse matrix relation and the adjusted sparse matrix relation; and storing viewpoint big data characteristics of the periodic demand feedback data based on the extended demand feedback graph node network, wherein the viewpoint big data characteristics are used as page optimization reference data of the user online page.
For example, in a possible implementation manner of the first aspect, acquiring a first topic viewpoint feature in a first demand feedback activity in a demand feedback graph node network, determining a corresponding second demand feedback activity based on a first network connection relationship of the first demand feedback activity, and acquiring a second topic viewpoint feature in the second demand feedback activity includes: analyzing a first topic viewpoint feature corresponding to a preset topic viewpoint feature mapping table in the first demand feedback activity; determining a corresponding second demand feedback activity based on a first network connection relationship of the first demand feedback activity, wherein the first network connection relationship comprises at least one of a forward trigger relationship, a backward trigger relationship, a cyclic trigger relationship and a cross trigger relationship; and analyzing a second topic viewpoint characteristic corresponding to a preset topic viewpoint characteristic mapping table in the second requirement feedback activity.
For example, in one possible implementation of the first aspect, the topic view feature mapping table is generated by:
acquiring a priori description knowledge set and a priori ideographic feature of each priori feedback activity in a priori demand feedback network, and analyzing the priori ideographic feature to obtain a first e-commerce ideographic field corresponding to the priori ideographic feature; comparing the first e-commerce ideographic field to an e-commerce ideographic field knowledge base to determine a second e-commerce ideographic field associated with a priori feedback activity; and counting the prior description knowledge sets of all the prior feedback activities and the second E-business ideographic fields to generate a first E-business ideographic field cluster, and carrying out deduplication processing on the prior description knowledge sets and/or the second E-business ideographic fields in the first E-business ideographic field cluster to obtain a topic viewpoint feature mapping table.
For example, in one possible implementation of the first aspect, the point of prior e-commerce behavioral ideographic knowledge is generated by: constructing corresponding E-commerce ideographic field distribution for each prior description knowledge set in a first E-commerce ideographic field cluster, and respectively loading second E-commerce ideographic fields associated with each prior description knowledge set into the E-commerce ideographic field distribution; and after loading a corresponding second E-business ideographic field in the E-business ideographic field distribution of each prior description knowledge set, generating corresponding prior E-business behavior ideographic knowledge points.
For example, in one possible implementation of the first aspect, the determining a first prior feedback activity associated with the first topic point characteristic and a second prior feedback activity associated with the second topic point characteristic based on the prior e-commerce behavioral ideographic knowledge points generated by the prior ideographic feature of each prior feedback activity in the prior demand feedback network comprises: comparing the first topic viewpoint features with all prior description knowledge sets and second E-commerce ideographic fields in the prior E-commerce behavior ideographic knowledge points to determine corresponding first prior feedback activities; and comparing the second topic viewpoint characteristics with all the prior description knowledge sets and the second E-commerce ideographic fields in the prior E-commerce behavior ideographic knowledge points to determine corresponding second prior feedback activities.
For example, in one possible implementation of the first aspect, the establishing a first communication link between the first demand feedback activity and the first a priori feedback activity and a second communication link between the second demand feedback activity and the second a priori feedback activity includes: if a first desired feedback activity is resolved to correspond to the determined a priori description knowledge set of the first a priori feedback activity, constructing a direct component connected link between the first desired feedback activity and the first a priori feedback activity; if a second required feedback activity is resolved to correspond to the determined a priori description knowledge set of a second a priori feedback activity, constructing a direct component connected link between the second required feedback activity and the second a priori feedback activity; if the first requirement feedback activity is analyzed to correspond to the determined second e-commerce ideogram field of the first prior feedback activity, an indirect component communication link is constructed between the first requirement feedback activity and the first prior feedback activity; if the second requirement feedback activity is analyzed to correspond to the determined second E-commerce ideogram field of the second prior feedback activity, an indirect component communication link is constructed between the second requirement feedback activity and the second prior feedback activity; and generating a corresponding first communication link or a second communication link according to the direct component communication link and the indirect component communication link of the first requirement feedback activity or the second requirement feedback activity.
For example, in one possible implementation of the first aspect, if it is determined that there is no sparse matrix relationship between the first prior feedback activity and the second prior feedback activity, constructing a sparse matrix relationship between the first prior feedback activity and the second prior feedback activity based on the first network connection relationship includes: and if the incidence relation of the sparse matrix does not exist between the first prior feedback activity and the second prior feedback activity, acquiring inference information of the first network connection relation, and constructing the sparse matrix relation between the first prior feedback activity and the second prior feedback activity according to the inference information.
For example, in a possible implementation manner of the first aspect, if it is determined that a sparse matrix relationship exists between a first a priori feedback activity and a second a priori feedback activity, adjusting the sparse matrix relationship based on the first network connection relationship includes: if the incidence relation of the sparse matrix exists between the first prior feedback activity and the second prior feedback activity, extracting a first activity relation between the first prior feedback activity and the second prior feedback activity; and if the reasoning information of the first activity relation and the first network connection relation is different, constructing a new sparse matrix relation between the first prior feedback activity and the second prior feedback activity according to the reasoning information, and adjusting the previous sparse matrix relation according to the new sparse matrix relation.
For example, in a possible implementation manner of the first aspect, the forming an extended requirement feedback graph node network according to the first connected link, the second connected link, the constructed sparse matrix relationship, and the adjusted sparse matrix relationship includes:
adjusting the prior demand feedback network according to the constructed sparse matrix relationship and the adjusted sparse matrix relationship; obtaining a quantized value of a demand feedback activity in the demand feedback graph node network to obtain a first activity quantized value, obtaining a quantized value of a prior feedback activity in the prior demand feedback network to obtain a second activity quantized value, and performing summation calculation based on the first activity quantized value and the second activity quantized value to obtain a third activity quantized value; obtaining a first link quantized value based on all the direct component connected links and the indirect component connected links, and obtaining a link mean quantized value based on the first link quantized value and the third activity quantized value; determining current position costs of the first network distribution position and the second network distribution position according to the link mean value quantized value, the preset mean value quantized value and the preset position cost, and configuring the first network distribution position and the second network distribution position on a three-dimensional knowledge map framework at intervals according to the current position costs; selecting a first association mode associated with a direct component connected link of the first connected link or the second connected link, and connecting a first requirement feedback activity or a second requirement feedback activity with a corresponding first prior feedback activity or a second prior feedback activity in the first association mode; selecting a second association mode associated with the indirect component connected link of the first connected link or the second connected link, and connecting the first requirement feedback activity or the second requirement feedback activity with the corresponding first prior feedback activity or second prior feedback activity in the second association mode; after the first demand feedback activity or the second demand feedback activity is connected with the corresponding first prior feedback activity or the second prior feedback activity, an extended demand feedback graph node network is generated.
In a second aspect, an embodiment of the present application further provides an artificial intelligence-based user online intention analysis system, where the artificial intelligence-based user online intention analysis system includes a big data e-commerce platform and a plurality of e-commerce service systems communicatively connected to the big data e-commerce platform;
the big data e-commerce platform is used for:
responding to each online intention analysis instruction in the current user online intention analysis task, and acquiring user online behavior data carried in the online intention analysis instruction;
loading the user online behavior data into a pre-trained target user online intention prediction model to obtain target user online intention information, and returning the target user online intention information to the e-commerce service system corresponding to the online intention analysis instruction;
and E-commerce resource pushing data fed back by the E-commerce service system according to the target user online intention information is obtained, and pushing processing is carried out on a user online page corresponding to the user online behavior data based on the E-commerce resource pushing data.
In any aspect, the online intention analysis instruction carries user online behavior data which is loaded into a pre-trained target user online intention prediction model to obtain target user online intention information, the target user online intention information is returned to the e-commerce service system corresponding to the online intention analysis instruction, e-commerce resource pushing data fed back by the e-commerce service system according to the target user online intention information is obtained, and pushing processing is performed on a user online page corresponding to the user online behavior data based on the e-commerce resource pushing data, so that the user online intention information can be presumed and associated e-commerce resource data can be loaded to the user online intention information, and the resource pushing matching degree of the user online e-commerce service is improved.
Drawings
Fig. 1 is a schematic flowchart of a method for analyzing a user online intention based on artificial intelligence according to an embodiment of the present application.
Detailed Description
Referring now to the architecture of the artificial intelligence based user online intention analysis system 10 provided by an embodiment of the present application, the artificial intelligence based user online intention analysis system 10 may include a big data e-commerce platform 100 and an e-commerce service system 200 communicatively connected to the big data e-commerce platform 100. The big data e-commerce platform 100 and the e-commerce service system 200 in the artificial intelligence based user online intention analysis system 10 can cooperatively perform the artificial intelligence based user online intention analysis method described in the following method embodiments, and the detailed description of the method embodiments can be referred to for the specific execution steps of the big data e-commerce platform 100 and the e-commerce service system 200.
The method for analyzing the user online intention based on the artificial intelligence provided by the embodiment can be executed by the big data e-commerce platform 100, and is described in detail below with reference to fig. 1.
The Process100 responds to each online intention analysis instruction in the current user online intention analysis task, and obtains user online behavior data carried in the online intention analysis instruction.
And the Process200 is used for loading the user online behavior data into a pre-trained target user online intention prediction model to obtain target user online intention information, and returning the target user online intention information to the e-commerce service system corresponding to the online intention analysis instruction.
The Process300 obtains e-commerce resource pushing data fed back by the e-commerce service system according to the target user online intention information, and performs pushing processing on a user online page corresponding to the user online behavior data based on the e-commerce resource pushing data.
In this embodiment, after determining the relevant online intention information of the target user, the e-commerce service system may obtain latest conversation topic data relevant to the online intention information of the target user from a current e-commerce resource push data source as the e-commerce resource push data, and if the online intention information of the target user includes attention intention information for a certain component B of a commodity a, may obtain e-commerce live broadcast push data for the certain component B of the commodity a, and perform push processing on a user online page corresponding to the user online behavior data.
Based on the above steps, the online intention analysis instruction carries user online behavior data, the user online behavior data is loaded into a target user online intention prediction model which is trained in advance, the target user online intention information is obtained, the target user online intention information is returned to the e-commerce service system corresponding to the online intention analysis instruction, e-commerce resource pushing data fed back by the e-commerce service system according to the target user online intention information is obtained, pushing processing is performed on a user online page corresponding to the user online behavior data based on the e-commerce resource pushing data, and therefore the online intention information of the user can be inferred and loaded with the associated e-commerce resource data, and the resource pushing matching degree of the user online e-commerce service is improved.
Process200 is illustrated below.
And the Process102 acquires the sample user online behavior data, the sample user online intention information and the sample behavior knowledge point information.
The sample user online behavior data refers to user online behavior data called when a user online intention prediction model is trained, and the user online behavior data is recorded data representing user online behaviors, such as e-commerce live broadcast attention behavior data, sharing behavior data, propagation behavior data and the like of a user. The sample user online intention information is user online intention information used in training a user online intention prediction model and is used as training label information corresponding to sample user online behavior data, and the user online intention information is data information describing user online intention, such as data information describing attention intention of a user for a certain accessory in a digital product. The sample behavior knowledge point information refers to a label corresponding to a behavior knowledge point (used for describing a single online behavior object) in the sample user online behavior data.
For example, the big data e-commerce platform 100 may obtain sample user online behavior data and corresponding sample user online intention information, and then perform behavior knowledge point labeling in a set on the sample user online behavior data to obtain sample behavior knowledge point information.
The Process104 loads the sample user online behavior data into a sample user online intention prediction model for feature extraction to obtain user behavior tendency features, extracts knowledge of the user behavior tendency features to obtain predicted behavior knowledge points, predicted behavior tendency attributes and predicted behavior knowledge point variables, obtains predicted user online intention information according to the predicted behavior knowledge points, the predicted behavior tendency attributes and the predicted behavior knowledge point variables, and performs behavior knowledge point prediction on the user behavior tendency features to obtain predicted behavior knowledge point information.
The sample user online intention prediction model is a user online intention prediction model for initializing model weight parameter information, and the user online intention prediction model is used for generating corresponding user online intention information for user online behavior data. The user behavior tendency characteristics refer to tendency characteristics corresponding to sample user online behavior data generated by using the initialized model weight parameter information. The predicted behavior knowledge point refers to a behavior knowledge point in sample user online behavior data identified by using the initialized model weight parameter information, and the behavior knowledge point can be a behavior instance for a certain user, a certain live broadcast room, a certain page option and the like. The predicted behavior tendency attribute refers to a behavior knowledge point attribute to be queried in sample user online behavior data identified by using initialized model weight parameter information. The predicted behavior knowledge point variable refers to a behavior knowledge point variable identified using the initialized model weight parameter information, and the behavior knowledge point variable is information for uniquely characterizing the behavior knowledge point. When the behavior knowledge points have a plurality of representation meanings, the specific corresponding behavior knowledge points can be determined through the behavior knowledge point variables. The predicted user online intention information refers to user online intention information corresponding to sample user online behavior data generated by using the initialized model weight parameter information. The predicted behavior knowledge point information refers to behavior knowledge point position labels in sample user online behavior data obtained by using initialized model weight parameter information identification.
Illustratively, the big data e-commerce platform 100 uses an AI neural network to build a framework of a sample user online intention prediction model, and initializes model weight parameter information to obtain the sample user online intention prediction model. And then, the sample user online behavior data can be loaded into a sample user online intention prediction model, and the sample user online intention prediction model can perform feature extraction on the sample user online behavior data, namely extracting feature vectors corresponding to the sample user online behavior data to obtain user behavior tendency features. The online behavior data of the sample user can be divided into behavior response feedback features, feature extraction is carried out on each behavior response feedback feature to obtain a corresponding feature vector, and finally the behavior tendency feature of the user is obtained. And then, simultaneously carrying out knowledge extraction and behavior knowledge point prediction on the behavior tendency characteristics of the user. The method comprises the steps of extracting knowledge of user behavior tendency characteristics, obtaining a predicted behavior knowledge point, a predicted behavior tendency attribute and a predicted behavior knowledge point variable through knowledge extraction, then obtaining an attribute behavior tendency node section corresponding to the predicted behavior tendency attribute through the predicted behavior knowledge point, the predicted behavior tendency attribute and the predicted behavior knowledge point variable, and then obtaining predicted user online intention information based on the predicted behavior knowledge point, the attribute behavior tendency node section and the predicted behavior knowledge point variable. And then, predicting the behavior knowledge points in the user behavior tendency characteristics to obtain predicted behavior knowledge point information.
And the Process106 is used for calculating the intention prediction cost value according to the predicted user online intention information and the sample user online intention information to obtain intention prediction cost value information, and calculating the behavior knowledge point prediction cost value according to the sample behavior knowledge point information and the predicted behavior knowledge point information to obtain behavior knowledge point prediction cost value information.
And the intention prediction cost value information is used for representing a difference value between the predicted user online intention information and the sample user online intention information. The behavior knowledge point prediction cost value information is used for representing a difference value between the predicted behavior knowledge point information and the sample behavior knowledge point information.
Illustratively, the big data e-commerce platform 100 may calculate a difference value between the predicted user online intention information and the sample user online intention information using a cross entropy loss function to obtain the intention prediction cost value information. And meanwhile, calculating a difference value between the sample behavior knowledge point information and the predicted behavior knowledge point information by using a cross entropy loss function to obtain the predicted cost value information of the behavior knowledge point.
And the Process108 optimizes the sample user online intention prediction model according to the intention prediction cost value information and the behavior knowledge point prediction cost value information to obtain an iterative user online intention prediction model, uses the iterative user online intention prediction model as the sample user online intention prediction model, and returns to the step of obtaining the sample user online behavior data, the sample user online intention information and the sample behavior knowledge point information to perform traversal loop processing until the training termination requirement is met, so as to obtain a target user online intention prediction model, wherein the target user online intention prediction model is used for generating corresponding user online intention information for the user online behavior data.
The iterative user online intention prediction model is a user online intention prediction model obtained by optimizing weight parameter information of an initialization model. The training termination requirement refers to a condition for obtaining the target user online intention prediction model through training, and may be that the number of times of traversal reaches a threshold value, or that the model weight parameter information does not change any more, and the like. The target user online intention prediction model refers to a user online intention prediction model after final training is completed.
Illustratively, the big data e-commerce platform 100 calculates weighted parameter values of the intention prediction cost value information and the behavior knowledge point prediction cost value information to obtain target prediction cost value information, and then the big data e-commerce platform 100 judges whether the training termination requirement is met, for example, compares the target prediction cost value information with a set cost value, and when the target prediction cost value information reaches the set cost value, it indicates that the training is completed. And when the target prediction cost value information does not reach the set cost value, the training does not meet the training termination requirement. At this time, the big data e-commerce platform 100 uses the target prediction cost value information to reversely update the initialized model weight parameter information in the sample user online intention prediction model through a gradient descent algorithm, so as to obtain an iterative user online intention prediction model. And taking the iterative user online intention prediction model as a sample user online intention prediction model, and returning to the step of obtaining the sample user online behavior data, the sample user online intention information and the sample behavior knowledge point information for traversing cycle processing until the training termination requirement is met, so as to obtain a target user online intention prediction model, wherein the target user online intention prediction model is used for generating corresponding user online intention information for the user online behavior data.
Based on the steps, the online behavior data of the sample user is loaded into the online intention prediction model of the sample user for feature extraction, user behavior tendency features are obtained, the user behavior tendency feature knowledge is used for extracting and generating predicted behavior knowledge points, meanwhile, behavior knowledge point prediction is carried out, the information of the predicted behavior knowledge points is obtained, then, intention prediction cost value information and behavior knowledge point prediction cost value information are calculated, model weight parameter optimization is carried out by using the intention prediction cost value information and the behavior knowledge point prediction cost value information, therefore, the obtained online intention prediction model of the iterative user is more accurate, after traversal iterative optimization, the prediction performance of the finally generated online intention prediction model of the target user can be improved, and the accuracy of online intention prediction of the user is finally improved.
For some exemplary design ideas, the Process104 loads sample user online behavior data into a sample user online intention prediction model for feature extraction, obtains user behavior tendency features, extracts knowledge of the user behavior tendency features, and obtains predicted behavior knowledge points, predicted behavior tendency attributes, and predicted behavior knowledge point variables, including:
and the Process202 converts the sample user online behavior data into a behavior response feedback feature set, and loads the behavior response feedback feature set into the sample user online intention prediction model.
The behavior response feedback feature set refers to a set formed by behavior response feedback features in the online behavior data of the sample user.
Illustratively, the big data e-commerce platform 100 divides the sample user online behavior data according to the behavior response feedback features to obtain various candidate behavior response feedback features, and splices the various candidate behavior response feedback features into a behavior response feedback feature set based on the sequence of the behavior response feedback features in the sample user online behavior data. The behavior response feedback feature set is then used as input to a sample user online intent prediction model.
The Process204 extracts tendency characteristics corresponding to the behavior response feedback characteristic set through the sample user online intention prediction model to obtain user behavior tendency characteristics;
the Process206 extracts knowledge of the user behavior tendency characteristics through the sample user online intention prediction model to obtain predicted behavior knowledge points, predicted behavior tendency attributes and candidate behavior knowledge point variables, and obtains corresponding description information of each behavior knowledge point according to the predicted behavior knowledge points.
The candidate behavior knowledge point variable refers to a behavior knowledge point variable which needs to be subjected to matching verification.
Illustratively, the big data e-commerce platform 100 extracts the tendency characteristics corresponding to the behavior response feedback characteristic set through the sample user online intention prediction model, that is, each behavior response feedback characteristic in the behavior response feedback characteristic set can be sequentially and respectively encoded, the obtained tendency characteristics corresponding to each behavior response feedback characteristic are obtained, and the tendency characteristics corresponding to all the behavior response feedback characteristics are combined to obtain the user behavior tendency characteristics. Knowledge extraction is carried out on the user behavior tendency characteristics through a sample user online intention prediction model, a predicted behavior knowledge point, a predicted behavior tendency attribute and a candidate behavior knowledge point variable are obtained, and corresponding description information of each behavior knowledge point can be searched from a knowledge base according to the predicted behavior knowledge point.
And the Process208 matches the candidate behavior knowledge point variables with the description information of each behavior knowledge point through the sample user online intention prediction model, and takes the candidate behavior knowledge point variables as predicted behavior knowledge point variables when the matching is successful.
The behavior knowledge point description information is used for explaining different representation meanings corresponding to the behavior knowledge points.
Illustratively, the big data e-commerce platform 100 matches the candidate behavior knowledge point variable with each behavior knowledge point description information through the sample user online intention prediction model, and when the candidate behavior knowledge point variable is not matched in each behavior knowledge point description information, it indicates that the candidate behavior knowledge point variable is generated incorrectly, at this time, the sample user online behavior data needs to be loaded into the sample user online intention prediction model to regenerate the candidate behavior knowledge point variable, and then matching is performed. Until the candidate behavior knowledge point variables are matched in the description information of each behavior knowledge point, the matching is successful, and the candidate behavior knowledge point variables are used as the predicted behavior knowledge point variables.
In the embodiment, the candidate behavior knowledge point variables are matched with the description information of each behavior knowledge point through the sample user online intention prediction model, and when the matching is successful, the candidate behavior knowledge point variables are used as the predicted behavior knowledge point variables, so that the generated predicted behavior knowledge point variables are limited in the range of the description information of each behavior knowledge point, and the accuracy of the generated predicted behavior knowledge point variables can be improved.
For some exemplary design ideas, the Process206 performs knowledge extraction on the user behavior tendency characteristics through a sample user online intention prediction model to obtain predicted behavior knowledge points, predicted behavior tendency attributes and candidate behavior knowledge point variables, and includes the steps of:
the Process302 extracts knowledge by using the prior knowledge point characteristics and the user behavior tendency characteristics through a sample user online intention prediction model to obtain the prediction initial knowledge point characteristics.
The prior knowledge point characteristics refer to characteristics which are preset and used for representing initial behavior response feedback characteristics which frequently appear in the past. The characteristic of the predicted initial knowledge point refers to a characterization vector corresponding to an initial behavior tendency node in information obtained by extracting the weight parameter information knowledge of the initialized model.
Illustratively, the online intent prediction model of the sample user in the big data e-commerce platform 100 splices the prior knowledge point features and the user behavior tendency features to obtain spliced features, and then performs knowledge extraction on the spliced features, which may be performed using an expert knowledge extraction network to obtain output predicted initial knowledge point features.
And the Process304 is used for extracting the knowledge of the predicted initial knowledge point characteristics and the user behavior tendency characteristics through the sample user online intention prediction model to obtain predicted intermediate knowledge point characteristics.
Wherein, the characteristic of predicting the intermediate knowledge point refers to a characterization vector corresponding to an intermediate behavior tendency node in information obtained by extracting the weight parameter information knowledge of the initialization model,
illustratively, the sample user online intention prediction model in the big data e-commerce platform 100 combines the prediction initial knowledge point features and the user behavior tendency features to obtain the splicing features, and then performs knowledge extraction on the splicing features, and may perform knowledge extraction using an expert knowledge extraction network to obtain the output prediction intermediate knowledge point features.
For some exemplary design ideas, a plurality of predicted intermediate knowledge point features can be obtained through knowledge extraction, and in this case, a plurality of intermediate behavior tendency nodes in the information obtained through knowledge extraction are illustrated. The plurality means at least two. And performing knowledge extraction by using the predicted intermediate knowledge point characteristics and the user behavior tendency characteristics obtained by the last knowledge extraction every time the knowledge extraction is performed, so as to obtain the predicted intermediate knowledge point characteristics obtained by the current knowledge extraction.
And the Process306 extracts knowledge of the predicted intermediate knowledge point characteristics and the user behavior tendency characteristics through the sample user online intention prediction model to obtain predicted target knowledge point characteristics.
The characteristic of the prediction target knowledge point refers to a characterization vector corresponding to a termination behavior tendency node in information obtained by extracting the weight parameter information knowledge of the initialization model.
Illustratively, the sample user online intention prediction model in the big data e-commerce platform 100 combines the predicted intermediate knowledge point features and the user behavior tendency features to obtain the splicing features, and then performs knowledge extraction on the splicing features, and may perform knowledge extraction using an expert knowledge extraction network to obtain the output predicted target knowledge point features, at which point the knowledge extraction is completed.
And the Process308 determines a predicted behavior knowledge point, a predicted behavior tendency attribute and a candidate behavior knowledge point variable by using the predicted initial knowledge point feature, the predicted intermediate knowledge point feature and the predicted target knowledge point feature through the sample user online intention prediction model.
Illustratively, the sample user online intention prediction model in the big data e-commerce platform 100 determines corresponding start behavior tendency nodes, intermediate behavior tendency nodes and end behavior tendency nodes using the predicted start knowledge point features, the predicted intermediate knowledge point features and the predicted destination knowledge point features, and then determines predicted behavior knowledge points, predicted behavior tendency attributes and candidate behavior knowledge point variables using the start behavior tendency nodes, the intermediate behavior tendency nodes and the end behavior tendency nodes.
Based on the steps, the priori knowledge point characteristics and the user behavior tendency characteristics are used for knowledge extraction, and the predicted initial knowledge point characteristics are obtained. And performing knowledge extraction on the predicted initial knowledge point characteristics and the user behavior tendency characteristics to obtain predicted intermediate knowledge point characteristics. And finally, determining a predicted behavior knowledge point, a predicted behavior tendency attribute and a candidate behavior knowledge point variable by using the predicted initial knowledge point feature, the predicted intermediate knowledge point feature and the predicted target knowledge point feature, so that each behavior tendency node to be generated can be subjected to knowledge extraction generation in sequence, and the accuracy of knowledge extraction generation is improved.
For some exemplary design ideas, the Process106, which performs intent prediction cost value calculation according to the predicted user online intent information and the sample user online intent information to obtain intent prediction cost value information, includes the steps of:
calculating first training cost value information corresponding to the predicted initial knowledge point characteristics by using the predicted initial knowledge point characteristics and the sample user online intention information; calculating second training cost value information corresponding to the predicted intermediate knowledge point characteristics by using the predicted intermediate knowledge point characteristics and the sample user online intention information; calculating third training cost value information corresponding to the predicted target knowledge point characteristics by using the predicted target knowledge point characteristics and the sample user online intention information; and calculating weighted cost value information of the first training cost value information, the second training cost value information and the third training cost value information to obtain intention prediction cost value information.
The first training cost value information is used for representing a difference value between an initial behavior tendency node generated by decoding the initialized model weight parameter information and a corresponding behavior tendency node in the sample user online intention information. And the second training cost value information is used for representing a difference value between the intermediate behavior tendency node generated by decoding the initialized model weight parameter information and the corresponding behavior tendency node in the sample user online intention information. And the third training cost value information is used for representing a difference value between a termination behavior tendency node generated by decoding the initialized model weight parameter information and a corresponding behavior tendency node in the sample user online intention information.
Illustratively, the big data e-commerce platform 100 calculates a difference value between a feature of a prediction start knowledge point and a feature vector of a corresponding behavior tendency node in the sample user online intention information by using a cross entropy loss function to obtain first training cost value information, calculates a difference value between a feature of a prediction intermediate knowledge point and the feature vector of the corresponding behavior tendency node in the sample user online intention information to obtain second training cost value information, calculates a difference value between a feature of a prediction target knowledge point and the feature vector of the corresponding behavior tendency node in the sample user online intention information to obtain third training cost value information. And finally, calculating weighted cost value information of the first training cost value information, the second training cost value information and the third training cost value information to obtain intention prediction cost value information.
Therefore, the intention prediction cost value information is obtained by calculating the first training cost value information, the second training cost value information and the third training cost value information and then calculating the weighted cost value information, so that the obtained intention prediction cost value information is more accurate.
For some exemplary design considerations, the Process202 converts the sample user online behavior data into a behavior response feedback feature set, and loads the behavior response feedback feature set into the sample user online intention prediction model, including the steps of:
performing behavior response feedback feature extraction on the sample user online behavior data to obtain each target behavior response feedback feature; acquiring a trigger response behavior tendency node section and an end response behavior tendency node section, splicing the trigger response behavior tendency node section, the end response behavior tendency node section and each target behavior response feedback characteristic according to a behavior response feedback sequence of the sample user online behavior data to acquire a behavior response feedback characteristic set, and loading the behavior response feedback characteristic set into a sample user online intention prediction model.
The target behavior response feedback feature refers to a behavior response feedback feature in the online behavior data of the sample user. The triggered response behavior tendency node segment is a response behavior tendency node segment used for representing the beginning of a behavior response feedback characteristic set, and the ended response behavior tendency node segment is a response behavior tendency node segment used for representing the ending of the behavior response feedback characteristic set.
In the embodiment, the triggered response behavior tendency node segment, the end response behavior tendency node segment and each target behavior response feedback feature are spliced according to the behavior response feedback sequence of the sample user online behavior data to obtain the behavior response feedback feature set, and the behavior response feedback feature set is loaded into the sample user online intention prediction model to predict the user online intention, so that the accuracy of the user online intention prediction can be improved.
For some exemplary design ideas, the Process208 matches the candidate behavior knowledge point variables with the description information of each behavior knowledge point through the sample user online intention prediction model, and when the matching is successful, takes the candidate behavior knowledge point variables as the predicted behavior knowledge point variables, including:
generating a behavior knowledge point description map by using the description information of each behavior knowledge point; and inquiring candidate behavior knowledge point variables in the behavior knowledge point description map, and when the candidate behavior knowledge point variables are inquired, taking the candidate behavior knowledge point variables as predicted behavior knowledge point variables.
The behavior knowledge point description map is a knowledge map formed by the behavior knowledge point description information.
Thus, by generating a behavior knowledge point description map; candidate behavior knowledge point variables are inquired in the behavior knowledge point description map, when the candidate behavior knowledge point variables are inquired, the candidate behavior knowledge point variables are used as predicted behavior knowledge point variables, the behavior knowledge point description map is used for matching to obtain the predicted behavior knowledge point variables, and the efficiency of obtaining the behavior knowledge point variables can be improved.
For some exemplary design considerations, the user behavior propensity characteristics include individual behavior response feedback prediction characteristics;
the Process104, which predicts behavior knowledge points for the user behavior tendency characteristics to obtain predicted behavior knowledge point information, includes:
the Process402 selects the current behavior response feedback prediction characteristics from the behavior response feedback prediction characteristics in sequence;
and the Process404 performs standard feature space conversion on the current behavior response feedback prediction features to obtain standard feature space conversion features, and performs behavior knowledge point labeling prediction on the standard feature space conversion features to obtain behavior knowledge point labeling probability.
The current behavior response feedback prediction feature refers to behavior knowledge point annotation information which needs to be subjected to behavior knowledge point prediction at present. The standard feature space conversion feature refers to a vector obtained after standard feature space conversion. The behavior knowledge point annotation probability is a probability value used for representing that the behavior response feedback feature is annotated by the behavior knowledge point, the higher the behavior knowledge point annotation probability is, the more likely the behavior response feedback feature is to be the behavior response feedback feature of the behavior knowledge point, and the lower the behavior knowledge point annotation probability is, the smaller the probability value of the behavior response feedback feature is to be the behavior response feedback feature of the behavior knowledge point.
For example, the big data e-commerce platform 100 may perform behavior knowledge point annotation prediction on each behavior response feedback prediction feature, and obtain a behavior knowledge point annotation probability corresponding to each behavior response feedback prediction feature. The big data e-commerce platform 100 can use a serial processing mode to perform behavior knowledge point annotation prediction and can also use a parallel processing mode to perform behavior knowledge point annotation prediction. The big data e-commerce platform 100 may sequentially select current behavior response feedback prediction features from the behavior response feedback prediction features according to a behavior response feedback sequence in the online behavior data of the user. The big data e-commerce platform 100 may randomly select a current behavioral response feedback prediction feature from the various behavioral response feedback prediction features. And then, performing standard feature space conversion on the current behavior response feedback prediction features, wherein standard feature space conversion parameters obtained by pre-training can be used for performing standard feature space conversion to obtain standard feature space conversion features, the standard feature space conversion parameters can comprise weight parameters and bias parameters, and the standard feature space conversion parameters are provided in a vector form. And then, performing behavior knowledge point annotation prediction on the standard feature space conversion feature through the prediction weight parameters obtained by pre-training, and obtaining behavior knowledge point annotation probability corresponding to the current behavior response feedback prediction feature. The behavior knowledge point marking prediction refers to predicting and marking a behavior knowledge point, and identifying the position of the knowledge point in the user online behavior data, wherein the position can include the behavior response feedback feature of the starting behavior knowledge point, the behavior response feedback feature of the middle behavior knowledge point, the behavior response feedback feature of the ending behavior knowledge point, the behavior response feedback feature of the single-line tendency node, other behavior response feedback features and the like.
And the Process406 determines current prediction labeling information corresponding to the current behavior response prediction features according to the behavior knowledge point labeling probability, returns to the step of sequentially selecting the current behavior response prediction features from the behavior response prediction features for traversal loop processing, obtains prediction labeling information corresponding to the behavior response feedback prediction features until the walk processing in the behavior response feedback prediction features is finished, and obtains the prediction behavior knowledge point information according to the prediction labeling information corresponding to the behavior response feedback prediction features.
The current prediction tagging information refers to target tagging information corresponding to current behavior response feedback characteristics in sample user online behavior data obtained by using initialized model weight parameter information for identification, for example, the current behavior response feedback characteristics may be target tagging information such as start behavior knowledge point behavior response feedback characteristics, intermediate behavior knowledge point behavior response feedback characteristics, end behavior knowledge point behavior response feedback characteristics, behavior knowledge point behavior response feedback characteristics of trend nodes, or other behavior response feedback characteristics.
Illustratively, the big data e-commerce platform 100 determines target labeling information corresponding to the maximum behavior knowledge point labeling probability based on the behavior knowledge point labeling probability, and uses the target labeling information as current prediction labeling information corresponding to the current behavior response feedback prediction feature. For example, the behavior knowledge point annotation probability of identifying and obtaining the current behavior response feedback feature as the beginning behavior knowledge point behavior response feedback feature is 90%, the behavior knowledge point annotation probability of identifying and obtaining the current behavior response feedback feature as the middle behavior knowledge point behavior response feedback feature is 2%, the behavior knowledge point annotation probability of identifying and obtaining the current behavior response feedback feature as the ending behavior knowledge point behavior response feedback feature is 3%, the behavior knowledge point annotation probability of identifying and obtaining the current behavior response feedback feature as the behavior knowledge point behavior response feedback feature with the single row being the tendency node behavior feedback feature is 4%, and the behavior knowledge point annotation probability of identifying and obtaining the current behavior response feedback feature as the other behavior response feedback features is 1%. And marking the behavior knowledge point marking probability of the behavior response feedback feature of the initial behavior knowledge point as the maximum 90%, marking the current behavior response feedback feature as the behavior response feedback feature of the initial behavior knowledge point, and obtaining the current prediction marking information.
And then the big data e-commerce platform 100 returns to step traversal loop processing of sequentially selecting current behavior response feedback prediction characteristics from the behavior response feedback prediction characteristics until the walk processing in the behavior response feedback prediction characteristics is finished, prediction marking information corresponding to the behavior response feedback prediction characteristics is obtained, and prediction behavior knowledge point information is obtained according to the prediction marking information corresponding to the behavior response feedback prediction characteristics.
Based on the steps, the behavior knowledge point marking prediction is carried out on each behavior response feedback prediction characteristic in sequence to obtain the prediction marking information corresponding to each behavior response feedback prediction characteristic, and then the prediction marking information corresponding to each behavior response feedback prediction characteristic is used to obtain the prediction behavior knowledge point information, so that the obtained prediction behavior knowledge point information is more accurate.
For some exemplary design ideas, the Process106 performs behavior knowledge point prediction cost value calculation according to the sample behavior knowledge point information and the predicted behavior knowledge point information to obtain behavior knowledge point prediction cost value information, and includes the steps of:
calculating a difference value between the prediction marking information corresponding to each behavior response prediction characteristic and the corresponding sample behavior response feedback characteristic information in the sample behavior knowledge point information to obtain each loss function value; and calculating the weighting parameter value of each loss function value to obtain the prediction cost value information of the behavior knowledge point.
And the loss function value is used for representing a difference value between the marking information of the behavior response feedback characteristic obtained by training and the sample behavior response feedback characteristic information in the real sample behavior knowledge point information. The sample behavior response feedback characteristic information refers to labeled category information corresponding to the behavior response feedback characteristic.
For example, the big data e-commerce platform 100 may calculate a difference value between prediction labeling information corresponding to each behavior response prediction feature and sample behavior response feedback feature information corresponding to sample behavior knowledge point information using a cross entropy loss function, obtain a loss function value corresponding to each behavior response feedback prediction feature, and then calculate weighted parameter values of all the loss function values to obtain behavior knowledge point prediction cost value information.
Therefore, each loss function value is obtained through calculation, and then the weighting parameter value of each loss function value is calculated to obtain the behavior knowledge point prediction cost value information, so that the accuracy of the obtained behavior knowledge point prediction cost value information is improved.
For some exemplary design considerations, the sample user online intent prediction model includes a sample feature extraction branch, a sample knowledge extraction branch, and a sample prediction branch;
the Process104, which loads the sample user online behavior data into the sample user online intention prediction model, includes:
the Process502 loads the sample user online behavior data into the sample feature extraction branch for feature extraction, and obtains user behavior tendency features.
And the Process504 loads the user behavior tendency characteristics into the sample knowledge extraction branch for knowledge extraction to obtain a predicted behavior knowledge point, a predicted behavior tendency attribute and a predicted behavior knowledge point variable, and obtains the predicted user online intention information according to the predicted behavior knowledge point, the predicted behavior tendency attribute and the predicted behavior knowledge point variable.
And the Process506 loads the user behavior tendency characteristics to the sample prediction branch to perform behavior knowledge point prediction, and obtains the predicted behavior knowledge point information.
The sample feature extraction branch is a feature extraction branch for initializing model weight parameter information and is used for performing feature extraction on input data. The sample knowledge extraction branch refers to a knowledge extraction branch for initializing model weight parameter information and is used for performing knowledge extraction on input features. The sample prediction branch refers to a prediction branch for initializing model weight parameter information and is used for performing information prediction on input features.
Illustratively, the big data e-commerce platform 100 loads sample user online behavior data into a sample feature extraction branch, and performs feature extraction by using an initialized feature extraction parameter to obtain user behavior tendency features. And loading the user behavior tendency characteristics into a sample knowledge extraction branch, performing knowledge extraction by using the predicted knowledge extraction parameters, obtaining predicted behavior knowledge points, predicted behavior tendency attributes and predicted behavior knowledge point variables, and obtaining the predicted user online intention information according to the predicted behavior knowledge points, the predicted behavior tendency attributes and the predicted behavior knowledge point variables. And loading the user behavior tendency characteristics into the sample prediction branch, and performing behavior knowledge point prediction by using the initialized prediction weight parameters to obtain the predicted behavior knowledge point information.
For some exemplary design considerations, the sample user online intent prediction model is a two-branch network architecture. When the sample user online intention prediction model acquires the online behavior data loaded to the sample user, the sample user online behavior data is loaded to the sample feature extraction branch for feature extraction, the output user behavior tendency feature is obtained, and then the user behavior tendency feature is loaded to the two branch networks at the same time, wherein the sample knowledge extraction branch of the first branch and the sample prediction branch of the second branch are included. And performing knowledge extraction generation through the sample knowledge extraction branch of the first branch, and performing prediction through the sample prediction branch of the second branch to obtain the training output of the model.
For some exemplary design considerations, the sample feature extraction branch and the sample knowledge extraction branch may use a long-short term memory neural network model.
The Process108, until the training termination requirement is met, obtains a target user online intention prediction model, including:
and when the training termination requirement is met, extracting branches according to the features meeting the training termination requirement and the knowledge meeting the training termination requirement to obtain the target user online intention prediction model.
Illustratively, the big data e-commerce platform 100 takes the feature extraction branch meeting the training termination requirement and the knowledge extraction branch meeting the training termination requirement as the target user online intention prediction model when judging that the training termination requirement is met. That is, the final target user online intention prediction model does not include a predicted branch, which is a network used in training to assist training.
Therefore, the sample feature extraction branch, the sample knowledge extraction branch and the sample prediction branch are used for training, when the training termination requirement is met, the target user online intention prediction model is obtained according to the feature extraction branch meeting the training termination requirement and the knowledge extraction branch meeting the training termination requirement, and therefore the prediction performance of the finally generated target user online intention prediction model can be improved.
For some exemplary design ideas, an artificial intelligence-based user online intention analysis method corresponding to a specific embodiment is provided at an application level of the present application, and the method includes the following steps:
and the Process602 responds to each online intention analysis instruction in the current user online intention analysis task, and acquires user online behavior data carried in the online intention analysis instruction.
The online intention analysis instruction refers to a request for analyzing online intention corresponding to the user online behavior data.
Illustratively, the related development user can trigger an online intention analysis operation after the online behavior data of the user is input by the e-commerce service system, the e-commerce service system responds to the online intention analysis operation, generates an online intention analysis instruction based on the online behavior data of the user, and sends the online intention analysis instruction to the big data e-commerce platform 100, and the big data e-commerce platform 100 receives the online intention analysis instruction sent by the e-commerce service system, analyzes the online intention analysis instruction, and obtains the online behavior data of the user.
The Process604 loads the user online behavior data into a target user online intention prediction model for feature extraction to obtain user behavior tendency features, and performs knowledge extraction on the user behavior tendency features to obtain target behavior knowledge points, target behavior tendency attributes and target behavior knowledge point variables; the target user online intention prediction model is characterized in that sample user online behavior data are loaded into a sample user online intention prediction model to perform feature extraction and knowledge extraction, a predicted behavior knowledge point, a predicted behavior tendency attribute and a predicted behavior knowledge point variable are obtained, predicted user online intention information is obtained according to the predicted behavior knowledge point, the predicted behavior tendency attribute and the predicted behavior knowledge point variable, behavior knowledge point prediction is performed on the user behavior tendency characteristic to obtain predicted behavior knowledge point information, intention prediction cost value calculation is performed according to the predicted user online intention information and the sample user online intention information to obtain intention prediction cost value information, behavior knowledge point prediction cost value calculation is performed according to the sample behavior knowledge point information and the predicted behavior knowledge point information to obtain behavior knowledge point prediction cost value information, the sample user online intention prediction model is optimized according to the intention prediction cost value information and the behavior knowledge point prediction cost value information to obtain an iterative user online intention prediction model, the iterative user online intention prediction model is used as the sample user online intention prediction model to perform traversal model optimization until the sample user online intention prediction model meets the training termination requirement.
Illustratively, the big data e-commerce platform 100 acquires a pre-trained target user online intention prediction model, and then deploys the target user online intention prediction model. When the device is required to be used, the user online behavior data are directly loaded into the target user online intention prediction model for feature extraction, user behavior tendency features are obtained, knowledge extraction is carried out on the user behavior tendency features, and output target behavior knowledge points, target behavior tendency attributes and target behavior knowledge point variables are obtained.
And the Process606 is used for obtaining the online intention information of the target user according to the target behavior knowledge point, the target behavior tendency attribute and the target behavior knowledge point variable, and returning the online intention information of the target user to the e-commerce service system corresponding to the online intention analysis instruction.
Illustratively, the big data e-commerce platform 100 finds the attribute field corresponding to the target behavior tendency attribute based on the target behavior knowledge point and the target behavior knowledge point variable, and then generates the target user online intention information based on the target behavior knowledge point and the target behavior knowledge point variable and the attribute field. And then the big data e-commerce platform 100 returns the online intention information of the mark user to the e-commerce service system corresponding to the online intention analysis instruction.
Based on the steps, target user online intention information corresponding to user online behavior data is obtained through a target user online intention prediction model, the target user online intention prediction model is that sample user online behavior data are loaded into the sample user online intention prediction model to conduct feature extraction and knowledge extraction, a predicted behavior knowledge point, a predicted behavior tendency attribute and a predicted behavior knowledge point variable are obtained, predicted user online intention information is obtained according to the predicted behavior knowledge point, the predicted behavior tendency attribute and the predicted behavior knowledge point variable, behavior knowledge point prediction is conducted on user behavior tendency characteristics to obtain predicted behavior knowledge point information, intention prediction cost value information and behavior knowledge point prediction cost value information are obtained through calculation, then the sample user online intention prediction model is optimized through the intention prediction cost value information and the behavior knowledge point prediction value information, an iterative user online intention prediction model is obtained, the iterative user online intention prediction model is used as the sample user online intention prediction model to conduct optimization of the model until the sample user online intention prediction model meets the traversal termination requirement, the prediction performance of the finally generated target user online intention prediction model can be improved, and the intention prediction accuracy of the user online intention prediction can be improved. The target behavior knowledge point, the target behavior tendency attribute and the target behavior knowledge point variable can be directly generated, so that the generation of cascade difference values can be avoided, and the accuracy of the obtained target user online intention information is improved.
For some exemplary design considerations, the target user online intent prediction model includes a target feature extraction branch and a target knowledge extraction branch;
the Process604, which loads the user online behavior data into the target user online intention prediction model, includes the steps of:
loading the user online behavior data into a target feature extraction branch for feature extraction to obtain user behavior tendency features; and loading the user behavior tendency characteristics into a target knowledge extraction branch for knowledge extraction, and obtaining target behavior knowledge points, target behavior tendency attributes and target behavior knowledge point variables.
The target feature extraction branch refers to a trained feature extraction branch and is used for performing feature extraction on input user online behavior data. The target knowledge extraction branch refers to a trained knowledge extraction branch and is used for extracting the knowledge of the input characteristic information.
Illustratively, the big data e-commerce platform 100 performs feature extraction on the user online behavior data by using the target feature extraction parameters in the target feature extraction branch, performs knowledge extraction on the user behavior tendency characteristics by using the target knowledge extraction parameters in the target knowledge extraction branch, and obtains output target behavior knowledge points, target behavior tendency attributes and target behavior knowledge point variables.
Based on the steps, the target behavior knowledge point, the target behavior tendency attribute and the target behavior knowledge point variable can be directly obtained by using the target feature extraction branch and the target knowledge extraction branch to perform feature extraction and knowledge extraction, and the user online intention prediction processing of a plurality of processes is not needed, so that the generation of cascading difference values is avoided, and the efficiency and the accuracy of user online intention prediction are improved.
For some exemplary design ideas, the big data e-commerce platform 100 converts the user online behavior data into a behavior response feedback feature set, and loads the behavior response feedback feature set into a target user online intention prediction model for feature extraction to obtain user behavior tendency features; and loading the user behavior tendency characteristics into a target knowledge extraction branch for knowledge extraction to obtain target behavior knowledge points, target behavior tendency attributes and candidate behavior knowledge point variables, acquiring corresponding target behavior knowledge point description maps by using the target behavior knowledge points, inquiring the candidate behavior knowledge point variables in the target behavior knowledge point description maps, and taking the candidate behavior knowledge point variables as the target behavior knowledge point variables when the candidate behavior knowledge point variables are inquired. The target behavior knowledge point description map is generated by using the behavior knowledge point description information of the target behavior knowledge point, and the behavior knowledge point description information of the target behavior knowledge point can be obtained by inquiring from the knowledge map.
For some exemplary design considerations, another embodiment of the present application specifically includes the following steps:
the Process702 is used for acquiring sample user online behavior data, sample user online intention information and sample behavior knowledge point information;
the Process704 extracts behavior response feedback features of the sample user online behavior data, obtains target behavior response feedback features, obtains a triggered response behavior tendency node section and an end response behavior tendency node section, splices the triggered response behavior tendency node section, the end response behavior tendency node section and the target behavior response feedback features according to a behavior response feedback sequence of the sample user online behavior data to obtain a behavior response feedback feature set, loads the behavior response feedback feature set into a sample user online intention prediction model, loads the behavior response feedback feature set into the sample user online intention prediction model, extracts tendency features corresponding to the behavior response feedback feature set through a sample feature extraction branch, and obtains user behavior tendency features.
The Process706 performs knowledge extraction by using the priori knowledge point feature and the user behavior tendency feature through a sample knowledge extraction branch to obtain a predicted initial knowledge point feature, performs knowledge extraction on the predicted initial knowledge point feature and the user behavior tendency feature to obtain a predicted intermediate knowledge point feature, performs knowledge extraction on the predicted intermediate knowledge point feature and the user behavior tendency feature to obtain a predicted target knowledge point feature, and determines a predicted behavior knowledge point, a predicted behavior tendency attribute and a candidate behavior knowledge point variable by using the predicted initial knowledge point feature, the predicted intermediate knowledge point feature and the predicted target knowledge point feature.
The Process708 acquires corresponding description information of each behavior knowledge point according to the predicted behavior knowledge point, generates a behavior knowledge point description map by using the description information of each behavior knowledge point, queries a candidate behavior knowledge point variable in the behavior knowledge point description map, and takes the candidate behavior knowledge point variable as the predicted behavior knowledge point variable when the candidate behavior knowledge point variable is queried.
And the Process710 obtains the online intention information of the predicted user according to the predicted behavior knowledge point, the predicted behavior tendency attribute and the predicted behavior knowledge point variable, loads each behavior response feedback prediction feature of the behavior tendency feature of the user into the sample prediction branch for predicting the behavior knowledge point, and obtains the prediction marking information corresponding to each behavior response feedback prediction feature.
The Process712 calculates first training cost value information corresponding to the predicted initial knowledge point feature using the predicted initial knowledge point feature and the sample user online intention information, calculates second training cost value information corresponding to the predicted intermediate knowledge point feature using the predicted intermediate knowledge point feature and the sample user online intention information, calculates third training cost value information corresponding to the predicted target knowledge point feature using the predicted target knowledge point feature and the sample user online intention information, calculates weighted cost value information of the first training cost value information, the second training cost value information, and the third training cost value information, and obtains intention predicted cost value information.
The Process714 is used for calculating the difference value between the prediction marking information corresponding to each behavior response feedback prediction characteristic and the corresponding sample behavior response feedback characteristic information in the sample behavior knowledge point information to obtain each loss function value, calculating the weighting parameter value of each loss function value to obtain behavior knowledge point prediction cost value information;
and the Process716 optimizes the sample user online intention prediction model according to the intention prediction cost value information and the behavior knowledge point prediction cost value information to obtain an iterative user online intention prediction model, uses the iterative user online intention prediction model as the sample user online intention prediction model, and returns to the step of obtaining the sample user online behavior data, the sample user online intention information and the sample behavior knowledge point information to perform traversal loop processing until the training termination requirement is met, and extracts branches according to the characteristics meeting the training termination requirement and the knowledge extraction branches meeting the training termination requirement to obtain the target user online intention prediction model.
And the Process718 responds to each online intention analysis instruction in the current user online intention analysis task, the online intention analysis instruction carries user online behavior data, the user online behavior data is loaded into the target user online intention prediction model, the generated target behavior knowledge point, the target behavior tendency attribute and the target behavior knowledge point variable are obtained, the target user online intention information is obtained according to the target behavior knowledge point, the target behavior tendency attribute and the target behavior knowledge point variable, and the target user online intention information is returned to the e-commerce service system corresponding to the online intention analysis instruction.
For some exemplary design considerations, after Process300, the following steps may be further included in the embodiments of the present application.
The STEP110 acquires staged demand feedback data of data pushed by the user for the pushed e-commerce resources in the online page, performs graph node network generation on the staged demand feedback data, acquires a corresponding demand feedback graph node network, acquires a first topic viewpoint feature in a first demand feedback activity in the demand feedback graph node network, determines a corresponding second demand feedback activity based on a first network connection relationship of the first demand feedback activity, and acquires a second topic viewpoint feature in the second demand feedback activity. The demand feedback graph node network may be a knowledge graph in an e-commerce scene, each demand feedback activity has a corresponding event, for example, an event of a first demand feedback activity is a feedback activity for an e-commerce live broadcast room, a first topic viewpoint feature at this time may be a topic viewpoint feature entering a commodity page in the e-commerce live broadcast room, a first network connection relationship at this time may have a plurality of, for example, the first network connection relationship may be a ordering skip relationship, and a second demand feedback activity associated with the first network connection relationship may be an ordering demand feedback activity for a commodity.
A plurality of corresponding second demand feedback activities may be derived based on the first demand feedback activity, and the first demand feedback activity is generally directed to the second demand feedback activity. After the first demand feedback activity and the pointed second demand feedback activity are obtained, the first topic viewpoint feature and the second topic viewpoint feature which are associated with the first demand feedback activity and the second demand feedback activity are respectively extracted.
For some exemplary design considerations, STEP110 in the above embodiments includes:
and analyzing a first topic viewpoint feature corresponding to a preset topic viewpoint feature mapping table in the first demand feedback activity. The first topic viewpoint feature of the first demand feedback activity is determined based on the topic viewpoint feature mapping table, and it can be understood that the first topic viewpoint feature determined in the embodiment of the present application may be a common e-commerce ideographic field in the field, for example, a next ideographic field, a collection ideographic field, a sharing ideographic field, and the like of an e-commerce scene.
And determining a corresponding second demand feedback activity based on a first network connection relationship of the first demand feedback activity, wherein the first network connection relationship comprises at least one of a forward trigger relationship, a backward trigger relationship, a cyclic trigger relationship and a cross trigger relationship. The embodiment of the application determines a corresponding second demand feedback activity by combining the first network connection relationship of the first demand feedback activity.
And analyzing a second topic viewpoint characteristic corresponding to a preset topic viewpoint characteristic mapping table in the second requirement feedback activity. The embodiment of the application determines a second topic viewpoint feature associated with the second demand feedback activity based on the topic viewpoint feature mapping table, where the second topic viewpoint feature may also be a general e-commerce ideographic field of an e-commerce scene.
For some exemplary design considerations, a topic view feature mapping table is generated by the steps comprising:
the method comprises the steps of obtaining a priori description knowledge set and a priori ideographic feature of each priori feedback activity in a priori demand feedback network, and carrying out E-commerce ideographic field analysis on the a priori ideographic feature to obtain a first E-commerce ideographic field corresponding to the a priori ideographic feature. Generally, the feedback network of the prior demand of different application scenarios will have different e-commerce ideographic fields, so the embodiment of the present application will first obtain a prior description knowledge set and a prior ideographic feature of each prior feedback activity in the knowledge graph. The a priori ideographic features may be descriptions of a priori descriptive knowledge sets. It can be understood that the prior description knowledge set of the prior feedback activity is the e-commerce ideogram field of the node, the prior ideogram can be the description of the prior feedback activity, and the first e-commerce ideogram field is the associated e-commerce ideogram field when describing the prior feedback activity.
Comparing the first e-commerce ideographic field to an e-commerce ideographic field knowledge base to determine a second e-commerce ideographic field associated with a priori feedback activity. After the first e-commerce ideographic field describing the prior feedback activity is obtained, at the moment, whether the first e-commerce ideographic field belongs to an e-commerce scene cannot be judged, so that the first e-commerce ideographic field needs to be compared with an e-commerce ideographic field knowledge base, and the first e-commerce ideographic field belonging to the e-commerce scene is used as a second e-commerce ideographic field.
And counting the prior description knowledge sets of all the prior feedback activities and the second E-business ideographic fields to generate a first E-business ideographic field cluster, and carrying out deduplication processing on the prior description knowledge sets and/or the second E-business ideographic fields in the first E-business ideographic field cluster to obtain a topic viewpoint feature mapping table. According to the embodiment of the application, a priori description knowledge set of all prior feedback activities and a second E-business ideogram cluster are counted to generate a first E-business ideogram cluster, and the E-business ideogram in the first E-business ideogram cluster belongs to an E-business scene, a priori description knowledge set of a certain prior feedback activity or has a relationship with the prior feedback activity. After obtaining the prior description knowledge set and/or the second e-commerce ideographic field associated with all the prior feedback activities, there may be a repeated prior description knowledge set and/or a second e-commerce ideographic field in the first e-commerce ideographic field cluster at this time, so that the prior description knowledge set and/or the second e-commerce ideographic field need to be deduplicated at this time.
When the prior description knowledge set and/or the second e-commerce ideographic fields are subjected to de-duplication processing, if the second e-commerce ideographic fields identical to the prior description knowledge set exist, all the repeated second e-commerce ideographic fields are deleted, and if a plurality of repeated second e-commerce ideographic fields exist, one of the second e-commerce ideographic fields is randomly reserved. Generally, no repeated a priori description knowledge sets exist, and if the repeated a priori description knowledge sets exist, one of the repeated a priori description knowledge sets is randomly reserved.
For some exemplary design considerations, a priori e-commerce behavioral ideographic knowledge points are generated by the following steps, including:
and constructing a corresponding E-business ideogram field distribution for each a priori description knowledge set in the first E-business ideogram field cluster, and loading second E-business ideogram fields associated with each a priori description knowledge set into the E-business ideogram field distribution respectively. The method and the device for the electric business ideographic field distribution can set corresponding electric business ideographic field distribution for each prior description knowledge set, and further can obtain the corresponding relation between each prior description knowledge set and the second electric business ideographic field.
And after loading a corresponding second E-commerce ideographic field in the E-commerce ideographic field distribution of each prior description knowledge set, generating corresponding prior E-commerce behavior ideographic knowledge points. According to the embodiment of the application, the second E-commerce ideographic fields associated with each prior description knowledge set can be counted through the prior E-commerce behavior ideographic knowledge points, so that the prior E-commerce behavior ideographic knowledge points have the corresponding relation between each prior description knowledge set and other second E-commerce ideographic fields, and the relation of each prior feedback activity is easy to count subsequently.
The STEP120 determines a first prior feedback activity associated with the topic point characteristics and a second prior feedback activity associated with the topic point characteristics based on prior E-commerce behavioral ideographic knowledge points generated by prior ideographic features of each prior feedback activity in the prior demand feedback network. According to the embodiment of the application, a first priori feedback activity associated with a first topic viewpoint characteristic and a second priori feedback activity associated with a second topic viewpoint characteristic are respectively obtained by combining the prior E-commerce behavior ideographic knowledge points.
For some exemplary design considerations, STEP120 includes:
STEP1201, comparing the first topic viewpoint features with all prior description knowledge sets and second E-commerce ideographic fields in the prior E-commerce behavior ideographic knowledge points, and determining corresponding first prior feedback activities.
STEP1202, comparing the second topic viewpoint features with all prior description knowledge sets and second E-business ideographic fields in the prior E-business behavior ideographic knowledge points to determine corresponding second prior feedback activities.
The method includes the steps that a first topic viewpoint feature and a second topic viewpoint feature are compared with a priori E-commerce behavior ideographic knowledge point respectively, a corresponding priori description knowledge set and a corresponding second E-commerce ideographic field are determined, and if the first topic viewpoint feature and the second topic viewpoint feature are judged to correspond to any one of the priori description knowledge set in the priori E-commerce behavior ideographic knowledge point or the second E-commerce ideographic field, corresponding first priori feedback activity and corresponding second priori feedback activity are obtained.
STEP130, constructing a first communication link between the first demand feedback activity and the first a priori feedback activity, and constructing a second communication link between the second demand feedback activity and the second a priori feedback activity. It can be understood that in the embodiment of the present application, a first connection link is established between the first requirement feedback activity and the first prior feedback activity, and a second connection link is established between the second requirement feedback activity and the second prior feedback activity, so that the feedback activities in the graph node networks with two different dimensions are connected.
For some exemplary design considerations, STEP130 includes:
if a first desired feedback activity is resolved to correspond to the determined a priori description knowledge set of the first a priori feedback activity, constructing a direct component connected link between the first desired feedback activity and the first a priori feedback activity; if a second desired feedback activity is resolved to correspond to the determined a priori descriptive knowledge set of a second a priori feedback activity, then a direct component connected link is constructed between the second desired feedback activity and the second a priori feedback activity. Taking the example of establishing the direct component connection link between the first demand feedback activity and the first prior feedback activity, the first demand feedback activity has the E-commerce ideographic field for ordering the commodity, and at this time, the first prior feedback activity exists for ordering the commodity, so the direct component connection link can be established between the first demand feedback activity and the first prior feedback activity, that is, the first demand feedback activity and the first prior feedback activity are directly related.
And if the first demand feedback activity is analyzed to correspond to the determined second e-commerce ideogram field of the first prior feedback activity, constructing an indirect component communication link between the first demand feedback activity and the first prior feedback activity. And if the second requirement feedback activity is analyzed to correspond to the determined second E-commerce ideogram field of the second prior feedback activity, constructing an indirect component communication link between the second requirement feedback activity and the second prior feedback activity. Taking the example of establishing an indirect component communication link between a first demand feedback activity and a first prior feedback activity, the first demand feedback activity has an e-commerce ideographic field for ordering commodities, and at this time, a second e-commerce ideographic field for a certain first prior feedback activity is used for ordering commodities, so that an indirect component communication link can be established between the first demand feedback activity and the first prior feedback activity, namely, the first demand feedback activity and the first prior feedback activity are indirectly related. The certain first prior feedback activity is, for example, live E-commerce feedback activity, description information of a second live E-commerce ideographic field related to the live E-commerce feedback activity belongs to a part of commodity ordering, and the live E-commerce feedback activity of commodity ordering needs to be replaced when overvoltage occurs. In this case, there is an indirect connection between the first demand feedback activity and the first a priori feedback activity.
And generating a corresponding first communication link or a second communication link according to the direct component communication link and the indirect component communication link of the first demand feedback activity or the second demand feedback activity. According to the embodiment of the application, all direct component connected links and indirect component connected links of the demand feedback activities are aggregated, and then corresponding first connected links or second connected links are generated.
STEP140, if it is analyzed that a sparse matrix relationship does not exist between the first prior feedback activity and the second prior feedback activity, constructing a sparse matrix relationship between the first prior feedback activity and the second prior feedback activity based on the first network connection relationship. If the sparse matrix relationship does not exist between the first prior feedback activity and the second prior feedback activity, it is proved that the prior feedback graph node network may not have the incidence relationship between the first prior feedback activity and the second prior feedback activity, so that the sparse matrix relationship needs to be established between the first prior feedback activity and the second prior feedback activity based on the first network connection relationship.
For some exemplary design considerations, STEP140 includes:
and if the incidence relation of the sparse matrix does not exist between the first prior feedback activity and the second prior feedback activity. At this time, the prior demand feedback network does not have a relationship between the first prior feedback activity and the second prior feedback activity, for example, the incidence relationship of a sparse matrix does not exist between the two feedback activities of the commodity ordering and the live broadcast feedback activity of the e-commerce, and the incidence relationship may not be configured when the graph node network is constructed. And acquiring inference information of the first network connection relation, and constructing a sparse matrix relation between the first prior feedback activity and the second prior feedback activity according to the inference information. At this time, the embodiment of the application obtains inference information of the first network connection relationship, and the inference information corresponds to logic between commodity ordering and E-commerce live broadcast feedback activities, so that a sparse matrix relationship can be constructed between the first prior feedback activity and the second prior feedback activity based on the inference information. The aggregated prior demand feedback network has knowledge point relationships with more multidimensional dimensions.
STEP150, if it is analyzed that a sparse matrix relationship exists between the first prior feedback activity and the second prior feedback activity, updating the sparse matrix relationship based on the first network connection relationship. If a sparse matrix relationship exists between the first prior feedback activity and the second prior feedback activity, the sparse matrix relationship needs to be updated based on the first network connection relationship, and a new link can be constructed between the first prior feedback activity and the second prior feedback activity by the updating mode.
For some exemplary design considerations, STEP150 includes:
and if the incidence relation of the sparse matrix exists between the first prior feedback activity and the second prior feedback activity, extracting a first activity relation between the first prior feedback activity and the second prior feedback activity. The method includes the steps that a first activity relation is extracted firstly, for example, a first activity relation exists between a commodity order (first prior feedback activity) and an e-commerce live broadcast feedback activity (second prior feedback activity), and a triple of the first activity relation can be that the commodity order includes the e-commerce live broadcast feedback activity.
And if the reasoning information analyzed that the first activity relation is different from the first network connection relation is analyzed, constructing a new sparse matrix relation between the first priori feedback activity and the second priori feedback activity according to the reasoning information, and updating the previous sparse matrix relation according to the new sparse matrix relation. At this time, the embodiment of the application compares the inference information of the first activity relationship with the inference information of the first network connection relationship, and if the inference information of the first activity relationship is different from the inference information of the first network connection relationship, a new sparse matrix relationship is constructed between the first priori feedback activity and the second priori feedback activity based on the inference information.
And STEP160, forming an expansion demand feedback graph node network according to the first communication link, the second communication link, the constructed sparse matrix relation and the adjusted sparse matrix relation. Through the method, the demand feedback activities and the prior feedback activities in the demand feedback graph node network and the prior demand feedback network can be aggregated, and the sparse matrix relation in the prior demand feedback network is adjusted to obtain the expanded demand feedback graph node network.
For some exemplary design considerations, STEP160 comprises:
and adjusting the prior demand feedback network according to the constructed sparse matrix relation and the adjusted sparse matrix relation. The embodiment of the application firstly obtains the updated prior demand feedback network, and compared with the prior demand feedback network before updating, the prior demand feedback network has more knowledge relation of prior feedback activity.
The method comprises the steps of constructing a first network distribution position associated with a demand feedback graph node network, and constructing a second network distribution position associated with an updated prior demand feedback network, wherein the first network distribution position and the second network distribution position are arranged in parallel. In the embodiment of the application, at least two parallel network distribution positions are constructed, the demand feedback graph node network corresponds to a first network distribution position, and the prior demand feedback network corresponds to a second network distribution position.
Selecting a first association mode associated with a direct component connected link of the first connected link or the second connected link, and connecting the first requirement feedback activity or the second requirement feedback activity with the corresponding first prior feedback activity or the second prior feedback activity in the first association mode. The demand feedback activity and the a priori feedback activity having the first pattern of correlation may be viewed as having a direct relationship.
Selecting a second association mode associated with an indirect component connected link of the first connected link or the second connected link, and connecting the first requirement feedback activity or the second requirement feedback activity with a corresponding first prior feedback activity or a second prior feedback activity in the second association mode. The demand feedback activity and the a priori feedback activity having the second pattern of correlation may be viewed as having an indirect relationship.
After the first demand feedback activity or the second demand feedback activity is connected with the corresponding first prior feedback activity or the second prior feedback activity, an extended demand feedback graph node network is generated. According to the embodiment of the application, after the corresponding demand feedback activities and the prior feedback activities are connected through the direct component communication link and the indirect component communication link, the corresponding extended demand feedback graph node network is obtained.
For some exemplary design considerations, the constructing a first network distribution location associated with the demand feedback graph node network, and constructing a second network distribution location associated with the updated a priori demand feedback network, where the first network distribution location and the second network distribution location are arranged in parallel, includes: and obtaining a quantized value of the demand feedback activity in the demand feedback graph node network to obtain a first activity quantized value, and obtaining the activity quantity of the demand feedback graph node network and the prior demand feedback network which can be reflected by the prior feedback activity in the prior demand feedback network through the first activity quantized value, the second activity quantized value and the third activity quantized value. And obtaining a first link quantized value based on all the direct component connected links and the indirect component connected links, and obtaining a link mean quantized value based on the first link quantized value and the third activity quantized value. The link mean value quantization value is obtained in the embodiment of the application, if the link mean value quantization value is larger, it is proved that the link between each prior feedback activity and the demand feedback activity is more, and if the link mean value quantization value is smaller, it is proved that the link between each prior feedback activity and the demand feedback activity is less.
Determining current position costs of the first network distribution position and the second network distribution position according to the link mean value quantized value, the preset mean value quantized value and the preset position cost, and configuring the first network distribution position and the second network distribution position on a three-dimensional knowledge map framework at intervals according to the current position costs.
The STEP170, storing the viewpoint big data characteristic of the periodic demand feedback data based on the extended demand feedback graph node network, as the page optimization reference data of the user online page.
Based on the steps, when the demand feedback graph node network and the prior demand feedback network are fused, the demand feedback activity is connected with the corresponding prior feedback activity, the potential relation of different prior feedback activities in the prior demand feedback network can be determined based on the relation of different demand feedback activities in the demand feedback graph node network, and then the sparse matrix relation of the prior feedback activities in the prior demand feedback network is adjusted, so that when the demand feedback activity is connected with the prior feedback activity, the demand feedback graph node network and the prior demand feedback network can be established to expand the demand feedback graph node network, the prior demand feedback network can be adjusted based on the demand feedback graph node network, and the reliability of graph node network expansion is improved.
In some embodiments, big data ecommerce platform 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 through a program stored in the machine-readable storage medium 120, such as program instructions related to the artificial intelligence based user online intent analysis method 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 executed by a processor, the computer-executable instructions are used to implement the method for analyzing user online intention based on artificial intelligence according to any one of the above embodiments.
Yet another embodiment of the present invention further provides a computer program product, which includes a computer program, and the computer program, when executed by a processor, implements the artificial intelligence based user online intention analysis method according to any of the above embodiments.
The foregoing is only an alternative implementation of some implementation scenarios in this application, and it should be noted that a person having ordinary skill in the art can also use other similar implementation means based on the technical idea of this application without departing from the technical idea of the present application, and the scope of protection of the embodiments of this application also belongs to this application.

Claims (10)

1. An artificial intelligence-based user online intention analysis method, characterized by comprising:
responding to each online intention analysis instruction in the current user online intention analysis task, and acquiring user online behavior data carried in the online intention analysis instruction;
loading the user online behavior data into a pre-trained target user online intention prediction model to obtain target user online intention information, and returning the target user online intention information to the e-commerce service system corresponding to the online intention analysis instruction;
and E-commerce resource pushing data fed back by the E-commerce service system aiming at the target user online intention information is obtained, and pushing processing is carried out on a user online page corresponding to the user online behavior data based on the E-commerce resource pushing data.
2. The method for analyzing the user online intention based on the artificial intelligence of claim 1, wherein the step of loading the user online behavior data into a pre-trained target user online intention prediction model to obtain the target user online intention information comprises:
loading the user online behavior data into a pre-trained target user online intention prediction model for feature extraction to obtain user behavior tendency features, extracting knowledge of the user behavior tendency features, and obtaining target behavior knowledge points, target behavior tendency attributes and target behavior knowledge point variables;
obtaining target user online intention information according to the target behavior knowledge point, the target behavior tendency attribute and the target behavior knowledge point variable, and returning the target user online intention information to the e-commerce service system corresponding to the online intention analysis instruction;
acquiring e-commerce resource pushing data fed back by the e-commerce service system aiming at the target user online intention information, and pushing a user online page corresponding to the user online behavior data based on the e-commerce resource pushing data;
wherein the target user online intention prediction model comprises a target feature extraction branch and a target knowledge extraction branch;
the loading the user online behavior data into a pre-trained target user online intention prediction model for feature extraction to obtain user behavior tendency features, extracting knowledge of the user behavior tendency features, and obtaining target behavior knowledge points, target behavior tendency attributes and target behavior knowledge point variables comprises:
loading the user online behavior data into the target feature extraction branch for feature extraction to obtain user behavior tendency features;
and loading the user behavior tendency characteristics into the target knowledge extraction branch for knowledge extraction, and obtaining target behavior knowledge points, target behavior tendency attributes and target behavior knowledge point variables.
3. The artificial intelligence based user online intent analysis method according to claim 1, further comprising:
acquiring sample user online behavior data, sample user online intention information and sample behavior knowledge point information;
converting the sample user online behavior data into a behavior response feedback feature set, and loading the behavior response feedback feature set into a sample user online intention prediction model;
analyzing the tendency characteristics corresponding to the behavior response feedback characteristic set based on the sample user online intention prediction model to obtain user behavior tendency characteristics;
extracting knowledge by using a priori knowledge point characteristic and the user behavior tendency characteristic based on the sample user online intention prediction model to obtain a prediction initial knowledge point characteristic;
performing knowledge extraction on the prediction initial knowledge point characteristics and the user behavior tendency characteristics based on the sample user online intention prediction model to obtain prediction intermediate knowledge point characteristics;
performing knowledge extraction on the predicted intermediate knowledge point characteristics and the user behavior tendency characteristics based on the sample user online intention prediction model to obtain predicted target knowledge point characteristics;
determining a predicted behavior knowledge point, a predicted behavior tendency attribute and a candidate behavior knowledge point variable based on the predicted initial knowledge point feature, the predicted intermediate knowledge point feature and the predicted target knowledge point feature based on the sample user online intention prediction model, and acquiring corresponding description information of each behavior knowledge point according to the predicted behavior knowledge point;
matching the candidate behavior knowledge point variables with the description information of each behavior knowledge point based on the sample user online intention prediction model, and taking the candidate behavior knowledge point variables as predicted behavior knowledge point variables when matching is successful;
obtaining on-line intention information of a predicted user according to the predicted behavior knowledge point, the predicted behavior tendency attribute and the predicted behavior knowledge point variable, and predicting the behavior knowledge point of the user behavior tendency characteristic to obtain predicted behavior knowledge point information;
determining first training cost value information corresponding to the predicted initial knowledge point characteristic based on the predicted initial knowledge point characteristic and the sample user online intention information;
determining second training cost value information corresponding to the predicted intermediate knowledge point characteristics based on the predicted intermediate knowledge point characteristics and the sample user online intention information;
determining third training cost value information corresponding to the predicted target knowledge point characteristics based on the predicted target knowledge point characteristics and the sample user online intention information;
determining weighted cost value information of the first training cost value information, the second training cost value information and the third training cost value information to obtain intention prediction cost value information, and performing behavior knowledge point prediction cost value calculation according to the sample behavior knowledge point information and the predicted behavior knowledge point information to obtain behavior knowledge point prediction cost value information;
and iteratively optimizing the sample user online intention prediction model according to the intention prediction cost value information and the behavior knowledge point prediction cost value information to obtain an iterative user online intention prediction model, using the iterative user online intention prediction model as the sample user online intention prediction model, and returning to the step of obtaining the sample user online behavior data, the sample user online intention information and the sample behavior knowledge point information for traversal loop processing until the training termination requirement is met, so as to obtain a target user online intention prediction model, wherein the target user online intention prediction model is used for generating corresponding user online intention information for the user online behavior data.
4. The method for analyzing the user online intention based on the artificial intelligence of claim 3, wherein the converting the sample user online behavior data into a behavior response feedback feature set, and the loading the behavior response feedback feature set into a sample user online intention prediction model comprises:
performing behavior response feedback feature extraction on the sample user online behavior data to obtain each target behavior response feedback feature;
acquiring a trigger response behavior tendency node segment and an end response behavior tendency node segment, splicing the trigger response behavior tendency node segment, the end response behavior tendency node segment and each target behavior response feedback characteristic according to a behavior response feedback sequence of the sample user online behavior data to acquire a behavior response feedback characteristic set, and loading the behavior response feedback characteristic set into a sample user online intention prediction model.
5. The method for analyzing the user's online intention based on the artificial intelligence of claim 3, wherein the matching the candidate behavior knowledge point variables with the description information of each behavior knowledge point based on the sample user online intention prediction model, and when the matching is successful, taking the candidate behavior knowledge point variables as the predicted behavior knowledge point variables comprises:
generating a behavior knowledge point description map based on the description information of each behavior knowledge point;
and inquiring the candidate behavior knowledge point variable in the behavior knowledge point description map, and when the candidate behavior knowledge point variable is inquired, taking the candidate behavior knowledge point variable as the predicted behavior knowledge point variable.
6. The artificial intelligence based user online intent analysis method of claim 3, wherein the user behavioral propensity characteristics include individual behavioral response feedback prediction characteristics;
the predicting the behavior knowledge point of the user behavior tendency characteristics to obtain the predicted behavior knowledge point information comprises the following steps:
sequentially selecting current behavior response feedback prediction characteristics from the behavior response feedback prediction characteristics;
performing standard feature space conversion on the current behavior response feedback prediction features to obtain standard feature space conversion features, and performing behavior knowledge point annotation prediction on the standard feature space conversion features to obtain behavior knowledge point annotation probability;
and determining current prediction marking information corresponding to the current behavior response prediction characteristics according to the behavior knowledge point marking probability, returning to step traversal loop processing of sequentially selecting the current behavior response prediction characteristics from the behavior response prediction characteristics, obtaining the prediction marking information corresponding to the behavior response feedback prediction characteristics until the walk processing in the behavior response feedback prediction characteristics is finished, and obtaining the prediction behavior knowledge point information according to the prediction marking information corresponding to the behavior response feedback prediction characteristics.
7. The method for analyzing the online intention of the user based on the artificial intelligence as claimed in claim 6, wherein the performing the behavior knowledge point prediction cost value calculation according to the sample behavior knowledge point information and the predicted behavior knowledge point information to obtain the behavior knowledge point prediction cost value information comprises:
determining a difference value between the prediction marking information corresponding to each behavior response prediction characteristic and the corresponding sample behavior response feedback characteristic information in the sample behavior knowledge point information to obtain each loss function value;
and determining the weighting parameter value of each loss function value to obtain the information of the behavior knowledge point prediction cost value.
8. The artificial intelligence based user online intention analysis method of claim 3, wherein the sample user online intention prediction model comprises a sample feature extraction branch, a sample knowledge extraction branch and a sample prediction branch;
the loading the sample user online behavior data into a sample user online intention prediction model comprises:
loading the sample user online behavior data into the sample feature extraction branch for feature extraction to obtain user behavior tendency features;
loading the user behavior tendency characteristics into the sample knowledge extraction branch for knowledge extraction to obtain a predicted behavior knowledge point, a predicted behavior tendency attribute and a predicted behavior knowledge point variable, and obtaining predicted user online intention information according to the predicted behavior knowledge point, the predicted behavior tendency attribute and the predicted behavior knowledge point variable;
loading the user behavior tendency characteristics to the sample prediction branch to predict behavior knowledge points, and obtaining predicted behavior knowledge point information;
when the training termination requirement is met, obtaining an online intention prediction model of the target user, including:
and when the training termination requirement is met, obtaining the target user online intention prediction model according to the feature extraction branch meeting the training termination requirement and the knowledge extraction branch meeting the training termination requirement.
9. The artificial intelligence based user online intention analysis method of any one of claims 1-8, wherein the method further comprises:
acquiring staged demand feedback data of the pushed e-commerce resource pushing data in the user online page, performing graph node network generation on the staged demand feedback data, acquiring a corresponding demand feedback graph node network, acquiring a first topic viewpoint feature in a first demand feedback activity in the demand feedback graph node network, determining a corresponding second demand feedback activity based on a first network connection relation of the first demand feedback activity, and acquiring a second topic viewpoint feature in the second demand feedback activity;
determining a first prior feedback activity associated with the first topic viewpoint feature and a second prior feedback activity associated with the second topic viewpoint feature based on prior E-commerce behavioral ideographic knowledge points generated by the prior ideographic feature of each prior feedback activity in the prior demand feedback network;
establishing a first connected link between the first demand feedback activity and the first prior feedback activity, and establishing a second connected link between the second demand feedback activity and the second prior feedback activity;
if the sparse matrix relationship does not exist between the first prior feedback activity and the second prior feedback activity through analysis, constructing a sparse matrix relationship between the first prior feedback activity and the second prior feedback activity based on the first network connection relationship;
if the sparse matrix relation exists between the first priori feedback activity and the second priori feedback activity, adjusting the sparse matrix relation based on the first network connection relation;
forming an expansion demand feedback graph node network according to the first communication link, the second communication link, the constructed sparse matrix relation and the adjusted sparse matrix relation;
and storing the viewpoint big data characteristic of the periodic demand feedback data based on the node network of the extended demand feedback graph as page optimization reference data of the user online page.
10. A big data e-commerce platform, comprising a processor and a memory for storing a computer program capable of running on the processor, the processor being configured to execute the artificial intelligence based user online intent analysis method of any one of claims 1 to 9 when running the computer program.
CN202211637844.8A 2022-12-20 2022-12-20 User online intention analysis method based on artificial intelligence and big data e-commerce platform Withdrawn CN115878900A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211637844.8A CN115878900A (en) 2022-12-20 2022-12-20 User online intention analysis method based on artificial intelligence and big data e-commerce platform

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211637844.8A CN115878900A (en) 2022-12-20 2022-12-20 User online intention analysis method based on artificial intelligence and big data e-commerce platform

Publications (1)

Publication Number Publication Date
CN115878900A true CN115878900A (en) 2023-03-31

Family

ID=85755250

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211637844.8A Withdrawn CN115878900A (en) 2022-12-20 2022-12-20 User online intention analysis method based on artificial intelligence and big data e-commerce platform

Country Status (1)

Country Link
CN (1) CN115878900A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117114741A (en) * 2023-10-20 2023-11-24 成都乐超人科技有限公司 Information decision method and system based on merchant portrait analysis

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117114741A (en) * 2023-10-20 2023-11-24 成都乐超人科技有限公司 Information decision method and system based on merchant portrait analysis
CN117114741B (en) * 2023-10-20 2024-03-22 成都乐超人科技有限公司 Information decision method and system based on merchant portrait analysis

Similar Documents

Publication Publication Date Title
CN109033107B (en) Image retrieval method and apparatus, computer device, and storage medium
CN111311107A (en) Risk assessment method and device based on user relationship and computer equipment
CN111368789B (en) Image recognition method, device, computer equipment and storage medium
CN114117240B (en) Internet content pushing method based on big data demand analysis and AI system
CN115048370B (en) Artificial intelligence processing method for big data cleaning and big data cleaning system
CN114090663B (en) User demand prediction method applying artificial intelligence and big data optimization system
CN113628005A (en) E-commerce session big data based pushing and updating method and big data AI system
CN114117235A (en) E-commerce content pushing method adopting artificial intelligence analysis and E-commerce big data system
CN113360349A (en) Information optimization method based on big data and cloud service and artificial intelligence monitoring system
CN115878900A (en) User online intention analysis method based on artificial intelligence and big data e-commerce platform
CN114817747B (en) User behavior analysis method based on internet big data and cloud computing service system
US20240095529A1 (en) Neural Network Optimization Method and Apparatus
CN114564648A (en) Personalized service content optimization method based on big data and artificial intelligence cloud system
CN117094535B (en) Artificial intelligence-based energy supply management method and system
CN116415647A (en) Method, device, equipment and storage medium for searching neural network architecture
CN112541556A (en) Model construction optimization method, device, medium, and computer program product
CN114896502B (en) User demand decision method applying AI and big data analysis and Internet system
CN112839102A (en) Business processing method applied to big data image pushing and machine learning server
CN111736774A (en) Redundant data processing method and device, server and storage medium
CN115563069B (en) Data sharing processing method and system based on artificial intelligence and cloud platform
CN114978765B (en) Big data processing method for information attack defense and AI attack defense system
CN114049161B (en) E-commerce big data feedback-based push optimization method and E-commerce big data system
CN114581139A (en) Content updating method based on big data intention mining and deep learning service system
CN114757721A (en) Service prediction analysis method and AI (Artificial Intelligence) mining system for joint big data mining
JPWO2019167240A1 (en) Information processing equipment, control methods, and programs

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WW01 Invention patent application withdrawn after publication
WW01 Invention patent application withdrawn after publication

Application publication date: 20230331