CN114969552B - Big data mining method and AI prediction system for personalized information push service - Google Patents

Big data mining method and AI prediction system for personalized information push service Download PDF

Info

Publication number
CN114969552B
CN114969552B CN202210785903.XA CN202210785903A CN114969552B CN 114969552 B CN114969552 B CN 114969552B CN 202210785903 A CN202210785903 A CN 202210785903A CN 114969552 B CN114969552 B CN 114969552B
Authority
CN
China
Prior art keywords
user behavior
event
analyzed
behavior event
events
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.)
Active
Application number
CN202210785903.XA
Other languages
Chinese (zh)
Other versions
CN114969552A (en
Inventor
孟淑君
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.)
Zhongtu Digital Technology Beijing Co ltd
Original Assignee
Zhongtu Digital Technology Beijing 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 Zhongtu Digital Technology Beijing Co ltd filed Critical Zhongtu Digital Technology Beijing Co ltd
Priority to CN202210785903.XA priority Critical patent/CN114969552B/en
Publication of CN114969552A publication Critical patent/CN114969552A/en
Application granted granted Critical
Publication of CN114969552B publication Critical patent/CN114969552B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • 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

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The embodiment of the invention provides a big data mining method and an AI (artificial intelligence) prediction system for personalized information push service, which can perform feature optimization on the behavior vector description of a user behavior event to be analyzed according to the behavior vector description of a user behavior event associated with the user behavior event to be analyzed, and can combine the information of the associated user behavior event in the second behavior vector description of the user behavior event to be analyzed to further determine the second behavior vector description of the user behavior event to be analyzed with higher reliability; because the interest point decision is made according to the second behavior vector description of the user behavior event to be analyzed, the interest point decision can be made according to the accurate behavior vector description of the user behavior event to be analyzed, so that the interest point corresponding to the user behavior event to be analyzed is effectively determined, the reliability of the interest point decision is improved, and the matching degree of the personalized information push service and the target user is finally improved.

Description

Big data mining method and AI prediction system for personalized information push service
Technical Field
The invention relates to the technical field of personalized push, in particular to a big data mining method and an AI prediction system for personalized information push service.
Background
With the development and application of internet information technology and computer technology and the openness of internet technology, the behavior big data of users in the internet shows explosive growth, and how to effectively utilize the behavior big data of the users to mine interest points of the related users is convenient to know user requirements and optimize personalized information push services subscribed by the users, so that the method is a direction of key research of each internet service provider at present. However, in the related technical solutions, the user usually inputs keywords to express the interest point requirement, and an active interest point mining solution is lacking, so that the matching degree of the personalized information push service and the target user is not high.
Disclosure of Invention
In order to overcome at least the above-mentioned disadvantages of the prior art, the present invention provides a big data mining method and an AI prediction system for personalized information push service.
In a first aspect, an embodiment of the present invention provides a big data mining method for a personalized information push service, which is applied to an AI prediction system, and the method includes:
extracting a user behavior event to be analyzed matched with the current service to be online from a big data acquisition library of a target user, and searching a plurality of associated user behavior events corresponding to the user behavior event to be analyzed from a target user behavior event sequence;
optimizing a first deep learning network model according to the user behavior event to be analyzed and the plurality of associated user behavior events, and determining a second deep learning network model, wherein the first deep learning network model is an AI decision branch model based on a relational inference network;
according to the first behavior vector description of the user behavior event to be analyzed and the behavior vector description corresponding to each associated user behavior event in the second deep learning network model, performing first traversal feature optimization based on the second deep learning network model, and determining the second behavior vector description of the user behavior event to be analyzed;
according to the second behavior vector description of the user behavior event to be analyzed and the behavior vector descriptions of a plurality of interest points corresponding to the second deep learning network model, outputting interest confidence degrees of the user behavior event to be analyzed on the interest points, wherein the interest points comprise the interest points corresponding to the user behavior events in the target user behavior event sequence in the second deep learning network model;
and outputting the interest points with the interest confidence degrees larger than the first set confidence degree as the interest points corresponding to the user behavior events to be analyzed, and updating the personalized information push service subscribed by the target user based on the interest points corresponding to the user behavior events to be analyzed.
In a second aspect, an embodiment of the present invention further provides a big data mining system for a personalized information push service, where the big data mining system for a personalized information push service includes an AI prediction system and multiple personalized service devices in communication connection with the AI prediction system;
the AI prediction system to:
extracting a user behavior event to be analyzed matched with the current service to be online from a big data acquisition library of a target user, and searching a plurality of associated user behavior events corresponding to the user behavior event to be analyzed from a target user behavior event sequence;
optimizing a first deep learning network model according to the user behavior event to be analyzed and the plurality of associated user behavior events, and determining a second deep learning network model, wherein the first deep learning network model is an AI decision branch model based on a relational inference network;
according to the first behavior vector description of the user behavior event to be analyzed and the behavior vector description corresponding to each associated user behavior event in the second deep learning network model, performing first traversal feature optimization based on the second deep learning network model, and determining the second behavior vector description of the user behavior event to be analyzed;
according to the second behavior vector description of the user behavior event to be analyzed and the behavior vector descriptions of a plurality of interest points corresponding to the second deep learning network model, outputting interest confidence degrees of the user behavior event to be analyzed on the interest points, wherein the interest points comprise the interest points corresponding to the user behavior events in the target user behavior event sequence in the second deep learning network model;
and outputting the interest points with the interest confidence degrees larger than a first set confidence degree as the interest points corresponding to the user behavior events to be analyzed, and updating the personalized information push service subscribed by the target user based on the interest points corresponding to the user behavior events to be analyzed.
Based on the embodiment scheme of any one aspect, feature optimization can be performed on the behavior vector description of the user behavior event to be analyzed according to the behavior vector description of the associated user behavior event of the user behavior event to be analyzed, and the second behavior vector description of the user behavior event to be analyzed with higher reliability can be determined by combining the information of the associated user behavior event in the second behavior vector description of the user behavior event to be analyzed; because the interest point decision is made according to the second behavior vector description of the user behavior event to be analyzed, the interest point decision can be made according to the accurate behavior vector description of the user behavior event to be analyzed, so that the interest point corresponding to the user behavior event to be analyzed is effectively determined, the reliability of the interest point decision is improved, and the matching degree of the personalized information push service and the target user is finally improved.
Drawings
Fig. 1 is a schematic flowchart of a big data mining method for a personalized information push service according to an embodiment of the present invention.
Detailed Description
The architecture of the big data mining system 10 for personalized information push service according to an embodiment of the present invention is described below, and the big data mining system 10 for personalized information push service may include an AI forecast system 100 and a personalized service device 200 communicatively connected to the AI forecast system 100. The AI prediction system 100 and the personalized service device 200 in the big data mining system 10 for personalized information push service can cooperatively perform the big data mining method for personalized information push service described in the following method embodiments, and the detailed description of the method embodiments can be referred to in the following steps of the AI prediction system 100 and the personalized service device 200.
The big data mining method for the personalized information push service provided by the present embodiment may be executed by the AI prediction system 100, and is described in detail below with reference to fig. 1.
In the Process101, a first relationship inference network in the first deep learning network model is determined according to the sample behavior data sequence.
In some of the possible design concepts, the first, prior to AI learning configuration for the first AI decision branch, a first relational inference network (first graph node network) is first determined. The first relational inference network includes: a connection dependency (edge) sequence OR and a network member (entity node) sequence, wherein the connection dependency sequence comprises: according to the sample behavior data sequence, performing pre-learning to obtain a connection dependency relationship sequence ORm between user behavior events, a connection dependency relationship sequence ORm between the user behavior events and corresponding interest points, and a connection dependency relationship sequence ORy between the interest points; the network member sequence includes: a sequence of interest points and a sequence of target user behavior events.
In some possible design concepts, each sample behavior data in the sample behavior data sequence includes: the method comprises the steps of obtaining a user behavior event, a join event node of the user behavior event and a corresponding interest point of the user behavior event in the join event node.
In some possible design ideas, when a first relation inference network is determined according to a sample behavior data sequence, generating a target user behavior event sequence according to a plurality of interest points covered by the sample behavior data sequence and generating an interest point sequence and a connection dependency relationship sequence ORy between the interest points according to the plurality of user behavior events covered by the sample behavior data sequence; and generating a connection dependency relationship sequence ORm, y between the user behavior event and the corresponding interest point according to the interest points corresponding to the plurality of user behavior events covered by the sample behavior data sequence. Therefore, a connection dependency relationship sequence ORm among the user behavior events can be determined, and a first relationship inference network can be constructed according to ORm, ORy and ORm, y.
In some possible design concepts, the associated user behavior event corresponding to the user behavior event may be determined based on the following steps: determining a target user behavior event cluster from the target user behavior event sequence, wherein each target user behavior event in the target user behavior event cluster shares at least one user behavior node data with a connection event node of the user behavior event, and the event magnitude of the target user behavior event covered by the target user behavior event cluster is smaller than the event magnitude of the user behavior event covered by the target user behavior event sequence; determining the matching degree between each target user behavior event in the target user behavior event cluster and the user behavior event; and outputting the target user behavior events at the preset order in the sequencing sequence from large matching degree to small matching degree as the associated user behavior events corresponding to the user behavior events.
In some possible design ideas, firstly, a target user behavior event cluster is determined from a target user behavior event sequence, and the matching degree of mi and mj is calculated for each user behavior event, and secondly, according to a sequence from large to small of the matching degree, the target user behavior events corresponding to the first K matching degrees are selected as the associated user behavior events of the user behavior event mi. The join event node of the target user behavior event and the join event node of the user behavior event at least share one user behavior node data, and (namely, the event magnitude of the target user behavior event is far smaller than that of the target user behavior event), so that the computing resources can be saved.
In some possible design ideas, the matching degree of mi and mj may be a node matching degree.
In some possible design ideas, the matching degree of mi and mj may be the matching degree of interest points. The interest point matching degree of mi and mj can be obtained by the following method: determining the interest confidence of mi on the interest point sequence based on the interest point classification model according to mi and the corresponding connection event node, and obtaining the interest confidence of mj on the interest point sequence in the same way; and taking the matching degree between the interest confidence degrees of mi and mj on the interest point sequence as the interest point matching degree between mi and mj. The interest point classification model is obtained by pre-training according to the same sample behavior data sequence before AI learning configuration is carried out on the first AI decision branch.
The interest point classification model comprises three parts, namely a user behavior node feature construction part, a significance feature attention part and an interest point classification part.
In some possible design ideas, the training steps are performed according to each sample behavior data in the sample behavior data sequence, and may be performed by the processes 401 to 405.
The processes 401 to 405 will be explained below.
In the Process401, feature extraction is performed on the user behavior event and the corresponding join event node, so as to obtain corresponding user behavior node features respectively. The step corresponds to a user behavior node characteristic construction part of the interest point classification model.
In the Process402, the user behavior event and the user behavior node characteristics of the corresponding join event node are respectively calculated based on different attention strategies for the significant characteristics, and the significant behavior vector description of the user behavior event and the significant behavior vector description of the corresponding join event node are respectively obtained. This step corresponds to the salient feature attention strategy calculation part of the interest point classification model.
In the Process403, the two salient behavior vector descriptions are fused, the fused description is used as a feature vector corresponding to the sample behavior data, and the feature vector is transformed to obtain a potential interest point description.
In the Process404, the interest point classification is performed based on the interest point classification model according to the feature vector corresponding to the sample behavior data and the potential interest point description, and the interest confidence of the user behavior event on the interest point sequence is determined. This step corresponds to the point of interest classification part of the point of interest classification model.
In the Process405, the interest confidence of the user behavior event on the interest point sequence and the actual confidence of the actual prior interest point corresponding to the user behavior event are substituted into a preset Loss function to determine a decision cost value, and in the back propagation Process, the parameters of the interest point classification model are updated according to the decision cost value.
The processes 401 to 405 are repeatedly executed until an AI learning termination requirement is met, for example, a preset Loss function is converged or a maximum training time is reached.
In some possible design ideas, after the training process of the interest point classification model is finished, the interest points corresponding to the user behavior events can be classified according to the interest point classification model obtained through learning. In the application stage of the interest point classification model, the step of obtaining the interest confidence of the user behavior event on the interest point sequence is the same as the Process401-404, and is not described herein again.
And determining the associated user behavior events corresponding to the target user behavior events through the two modes, thereby determining the ORm and constructing a first relation inference network according to the ORm, the ORy and the ORm, y.
In some possible design ideas, the associated user behavior event corresponding to the user behavior event mi is m1-m4, the corresponding interest point is y1-y3, and the interest points y2 and y3 are backward interest points of the interest point y1, and the first relationship inference network comprises two different nodes, namely the user behavior event and the interest points, so that the first relationship inference network is a heterogeneous relationship inference network.
In the Process102, according to the first behavior vector description of each user behavior event in the first relational inference network and the behavior vector description of the relational object of each user behavior event in the first relational inference network, the first traversal feature optimization is performed based on the first AI decision branch substituted into the preset network setting coefficient, and the second behavior vector description of each user behavior event is determined.
The first deep learning network model comprises a first relational inference network and a first AI decision branch, and in some possible design ideas, the first AI decision branch comprises L network elements, L is an integer greater than 2, each network element of the first AI decision branch acts on the provided first relational inference network, and each network element of the first AI decision branch outputs a processing result of a current network element through an activation function.
In some possible design ideas, according to the first behavior vector description of each user behavior event in the first relational inference network and the behavior vector description of the relational object of the user behavior event, the first traversal feature optimization is performed based on the first AI decision branch substituted into the preset network setting coefficient, and the second behavior vector description of each user behavior event is determined. The relationship object of each user behavior event comprises: and the associated user behavior events corresponding to the user behavior events and the corresponding interest points.
In some possible design ideas, the relationship object of the user behavior event mi comprises the associated user behavior events m1 to m4 and the corresponding interest points y1 to y3.
In some possible design concepts, the first behavior vector description of the user behavior event is determined by: performing feature extraction on the user behavior event and the linkage event node of the user behavior event, and determining the user behavior event and the user behavior node feature corresponding to the linkage event node of the user behavior event; respectively performing significance feature attention processing on user behavior node features corresponding to user behavior events and linkage event nodes of the user behavior events based on different significance feature attention strategies (such as a soft significance feature attention strategy and a hard significance feature attention strategy), and determining significance behavior vector descriptions corresponding to the user behavior events and the linkage event nodes of the user behavior events; and fusing the significant behavior vector descriptions corresponding to the user behavior event and the connection event node of the user behavior event respectively, and determining the fused description as the first behavior vector description of the user behavior event.
In some possible design approaches, the first traversal feature optimization is implemented by the first AI decision branch, i.e., the first traversal feature optimization is performed by common processing of the various network elements covered by the first AI decision branch.
In some possible design concepts, starting with the 1 st network element of the first AI decision branch, taking the 1 st network element as the current network element, the following operations are performed: performing first feature selection based on penalty terms on the behavior vector description of each associated user behavior event corresponding to the current network unit, and determining first feature selection data corresponding to each associated user behavior event; weighting the first feature selection data corresponding to each associated user behavior event and the fusion description of the significance influence coefficient corresponding to the current network unit to determine a first weighting result; performing third penalty term-based feature selection on the behavior vector description of the interest point corresponding to the user behavior event in the current network unit, determining third feature selection data corresponding to the interest point corresponding to the user behavior event, weighting the third feature selection data corresponding to the interest point corresponding to the user behavior event and the fusion description of the significance influence coefficient corresponding to the current network unit, determining a second weighting result, weighting the first weighting result and the second weighting result, performing second penalty term-based feature selection on the obtained third weighting result, and determining the second feature selection data as the second behavior vector description of the user behavior event in the current network unit.
And after the second behavior vector description of the user behavior event in the 1 st network unit is obtained, taking the layer 2 as the current network unit, taking the second behavior vector description of the user behavior event in the 1 st network unit as the behavior vector description of the current network unit, executing the same processing, and determining the second behavior vector description of the user behavior event in the 2 nd network unit. Based on the foregoing embodiment, the second behavior vector description of the user behavior event in each network element is obtained until the behavior vector description of the user behavior event in the L-th network element of the first AI decision branch is obtained.
In some possible design ideas, when the first traversal feature optimization is performed based on the first AI decision branch substituted into the preset network setting coefficient according to the first behavior vector description of each user behavior event in the first relational inference network and the behavior vector description of the relational object of each user behavior event in the first relational inference network, and when the relational object of each user behavior event in the first relational inference network is the interest point corresponding to the user behavior event, the interest point corresponding to the user behavior event is eliminated with the preset confidence level.
And when the relation object of the user behavior event is the interest point, rejecting the interest point corresponding to the user behavior event by using the preset confidence level. For example, the preset confidence is 0.6, and the actual prior interest points corresponding to the user behavior event are 10, so that each interest point has a probability of 0.6 to be removed, 10 × 0.6, that is, 6 interest points are finally removed, and the behavior vector description of the user behavior event is subjected to feature optimization by using the behavior vector descriptions of the remaining 4 interest points. In this way, in the process of obtaining the second behavior vector description of the user behavior event, only the behavior vector descriptions of the part of interest points corresponding to the user behavior event are used, and in order to determine the eliminated interest points corresponding to the user behavior event, the first AI decision branch learns effective information from the user behavior event associated with the user behavior event, so that the problem of overfitting of the first AI decision branch can be avoided.
In the Process103, according to the first behavior vector description of the multiple interest points in the first relational inference network and the behavior vector description of the relational object of each interest point in the first relational inference network, second traversal feature optimization is performed based on the first AI decision branch substituted into the preset network setting coefficient, and the second behavior vector description of the multiple interest points is determined.
In some possible design concepts, the relationship object of each interest point in the first relationship inference network includes: forward points of interest or backward points of interest for each point of interest.
In some possible design concepts, the first behavior vector description of the point of interest is obtained by random initialization. And performing second iterative processing based on the first AI decision branch substituted into the preset network setting coefficient according to the first behavior vector description of the interest points and the behavior vector description of the corresponding relation object, and determining the second behavior vector description of each interest point.
In some possible design approaches, the second traversal feature optimization is implemented by the first AI decision branch, i.e., the second traversal feature optimization is performed by common processing of the various network elements covered by the first AI decision branch.
In some possible design concepts, starting with the 1 st network element of the first AI decision branch, the 1 st network element is taken as the current network element, and the following operations are performed: performing fourth feature selection based on penalty items on the behavior vector description of each interest point in the current network unit, and determining fourth feature selection data corresponding to each interest point; and performing fourth penalty term-based feature selection processing on the behavior vector description of the forward interest point or the backward interest point corresponding to each interest point in the current network unit, weighting the obtained fourth feature selection data corresponding to the forward interest point or the backward interest point corresponding to each interest point and the fusion feature sequence of the significance influence coefficient corresponding to the current network unit as a fourth weighting result, determining a fifth weighting result, performing fifth penalty term-based feature selection on the fifth weighting result, and determining fifth feature selection data as the second behavior vector description of each interest point in the current network unit.
And after the second behavior vector description of each interest point in the 1 st network unit is obtained, taking the layer 2 as the current network unit, taking the second behavior vector description of each interest point in the 1 st network unit as the behavior vector description of the current network unit, executing the same processing, and determining the second behavior vector description of each interest point in the 2 nd network unit. Based on the foregoing embodiment, the second behavior vector description of each point of interest in each network element is obtained until the behavior vector description of each point of interest in the lth network element of the first AI decision branch is obtained.
And the Process104 outputs the interest confidence of each user behavior event on the plurality of interest points according to the second behavior vector description of each user behavior event and the second behavior vector descriptions of the plurality of interest points.
In some possible design ideas, after obtaining the second behavior vector description of each user behavior event and the second behavior vector descriptions of the multiple points of interest, the interest confidence of each user behavior event on the multiple points of interest is output according to the second behavior vector description of each user behavior event and the second behavior vector descriptions of the multiple points of interest.
In the Process105, the interest confidence and the actual confidence of each user behavior event on a plurality of interest points are substituted into a preset Loss function, and a decision cost value corresponding to each user behavior event is determined.
In some possible design ideas, after obtaining the interest confidence of each user behavior event on each interest point, substituting the interest confidence and the corresponding actual confidence of each user behavior event on each interest point into a preset Loss function to calculate the decision cost value corresponding to each user behavior event.
In the Process106, a global decision cost value is determined according to the decision cost value corresponding to each user behavior event, gradient determination of each network setting coefficient is performed in the first AI decision branch substituted into the preset network setting coefficients according to the global decision cost value, so as to optimize the network setting coefficients of the first AI decision branch substituted into the preset network setting coefficients, and the first AI decision branch completing training is output according to the optimized network setting coefficients.
In some possible design ideas, after determining the decision cost value corresponding to each user behavior event, determining a global decision cost value according to the decision cost value corresponding to each user behavior event. For example, the decision cost values corresponding to the user behavior events may be weighted and then averaged to obtain a global decision cost value, which is an event magnitude of the user behavior event included in the global decision cost value.
After the global decision cost value is determined, the gradient determination of each network setting coefficient is carried out in the first AI decision branch substituted into the preset network setting coefficients according to the global decision cost value, so that the network setting coefficients of the first AI decision branch substituted into the preset network setting coefficients are updated, and the first AI decision branch obtained by learning is constructed according to the optimized parameters.
And repeating the updating process until the AI learning termination requirement, such as the global decision cost value, is met to the minimum or the maximum training times are met. After the training is terminated, a first AI decision branch after the training is obtained.
After the trained first AI decision branch is obtained, combining the trained first AI decision branch and the first relational inference network into a gray scale deep learning network model, and verifying the gray scale deep learning network model through the following processes 201 to 205.
The validity verification process of the grayscale AI decision branch is described below.
In the Process201, a plurality of associated user behavior events for validity verification corresponding to the grayscale user behavior event are searched from the target user behavior event sequence.
In some possible design ideas, a plurality of associated user behavior events used for validity verification, which correspond to the grayscale user behavior event, are obtained from a target user behavior event sequence covered by the first relationship inference network. The obtaining mode of the associated user behavior event for validity verification corresponding to the grayscale user behavior event is the same as the obtaining mode of the associated user behavior event corresponding to the user behavior event.
In the Process202, the gray scale deep learning network model is optimized according to the gray scale user behavior event and the plurality of associated user behavior events for validity verification, and the optimized gray scale deep learning network model is determined.
In some possible design ideas, the gray-scale user behavior event is used as a new node to be added into the first relation reasoning network to obtain the optimized first relation reasoning network, and the optimized first relation reasoning network replaces the current first relation reasoning network in the gray-scale deep learning network model to determine the optimized gray-scale deep learning network model.
In the Process203, according to the first behavior vector description of the gray-scale user behavior event and the behavior vector descriptions corresponding to the associated user behavior events for validity verification in the optimized gray-scale deep learning network model, first traversal feature optimization is performed based on the first AI decision branch, and a second behavior vector description of the gray-scale user behavior event is determined.
In some possible design ideas, after the optimized gray scale deep learning network model is obtained, according to a first behavior vector description of a gray scale user behavior event and a behavior vector description of an associated user behavior event corresponding to the gray scale user behavior event and used for validity verification, first-pass feature optimization is performed based on a first AI decision branch, and a second behavior vector description of the gray scale user behavior event is determined.
In some possible design ideas, the first behavior vector description of the gray-scale user behavior event and the behavior vector description of the associated user behavior event corresponding to the gray-scale user behavior event and used for validity verification are subjected to first traversal feature optimization, so that a second behavior vector description of the gray-scale user behavior event is obtained. The obtaining mode of the first behavior vector description of the gray-scale user behavior event is the same as the obtaining mode of the first behavior vector description of the user behavior event in the training process.
Because the interest point relation object corresponding to the gray-scale user behavior event is an empty set, any information from the actual prior interest point is not involved in the process of obtaining the behavior vector description of the gray-scale user behavior event. That is, the behavior vector descriptions of grayscale user behavior events are updated only according to the behavior vector descriptions of the corresponding plurality of associated user behavior events for validity verification.
In the Process204, the interest confidence of the gray-scale user behavior event on the plurality of interest points is output according to the second behavior vector description of the gray-scale user behavior event and the behavior vector descriptions corresponding to the plurality of interest points in the optimized gray-scale deep learning network model.
In some possible design ideas, after the second behavior vector description of the gray-scale user behavior event is obtained, the interest confidence degrees of the gray-scale user behavior event on the interest points are output according to the second behavior vector description of the gray-scale user behavior event and the behavior vector descriptions corresponding to the interest points in the optimized gray-scale deep learning network model.
In some possible design ideas, the interest confidence of the grayscale user behavior event at each interest point may be calculated by using the second behavior vector description of the grayscale user behavior event and the behavior vector descriptions in the second behavior vector description set Y (L) of the interest points (i.e., the behavior vector descriptions of multiple interest points in the optimized grayscale deep learning network model).
In the Process205, when the difference confidence between the interest confidence and the actual confidence of the gray-scale user behavior event at the corresponding interest point is smaller than the second set confidence, it is determined that the validity verification of the optimized gray-scale deep learning network model is passed, and the optimized gray-scale deep learning network model is used as the first deep learning network model.
In some possible design ideas, after the interest confidence of the gray-scale user behavior event on each interest point is determined, whether the optimized gray-scale deep learning network model passes validity verification is judged based on the difference confidence between the interest confidence of the gray-scale user behavior event on each interest point and the actual confidence of the gray-scale user behavior event on the corresponding interest point. For example, when the difference is smaller than the second set confidence level (e.g., 0.1), it may be determined that the optimized gray scale deep learning network model passes the validity verification.
And after judging that the optimized gray scale deep learning network model passes validity verification, taking the optimized gray scale deep learning network model as a first deep learning network model and applying the first deep learning network model.
The application process of the first deep learning network model will be explained below.
In the Process301, a user behavior event to be analyzed matching the current service to be online is extracted from a big data collection library of a target user, and a plurality of associated user behavior events corresponding to the user behavior event to be analyzed are searched from a target user behavior event sequence.
In some possible design ideas, a plurality of associated user behavior events corresponding to the user behavior event to be analyzed are searched from a target user behavior event sequence covered by a first relation inference network of a first deep learning network model.
Process301 may also be implemented by Process3011 and Process 3012.
In the Process3011, a target user behavior event cluster is determined from the target user behavior event sequence.
In some possible design concepts, a target user behavior event cluster is determined from a target user behavior event sequence. Each target user behavior event in the target user behavior event cluster shares at least one user behavior node data with a connection event node of the user behavior event to be analyzed, and the event magnitude of the target user behavior event covered by the target user behavior event cluster is smaller than the event magnitude of the user behavior event covered by the target user behavior event sequence.
In the Process3012, a plurality of associated user behavior events corresponding to the user behavior event to be analyzed are obtained from the target user behavior event cluster.
In some possible design ideas, after a target user behavior event cluster is determined, a plurality of associated user behavior events corresponding to a user behavior event to be analyzed are obtained from the target user behavior event cluster. And the matching degree of each associated user behavior event and the user behavior event to be analyzed is greater than the set matching degree.
The associated user behavior event is determined based on the embodiment, and the associated user behavior event corresponding to the user behavior event to be analyzed is determined from the target user behavior event cluster with smaller magnitude, so that the calculation amount in the process of determining the associated user behavior event can be greatly reduced, and the time complexity of the calculation process is reduced.
Process3012 can also be implemented by Process30121 and Process30122, and Process30121 and Process30122 will be described below.
In the Process30121, a matching degree between each target user behavior event and a user behavior event to be analyzed is determined.
In some possible design concepts, a matching degree between a user behavior event to be analyzed and each target user behavior event in the target user behavior event cluster is determined.
In the Process30122, the target user behavior event with the preset rank in the sequence from large to small of the matching degree is output as a related user behavior event corresponding to the user behavior event to be analyzed.
In some possible design ideas, after the matching degree between the user behavior event to be analyzed and each target user behavior event is determined, the matching degrees are arranged from large to small, and the first K partial target user behavior events in the arrangement result from large to small are used as the associated user behavior events corresponding to the user behavior event to be analyzed.
For example, 50 target user behavior events exist in the target user behavior event cluster, after the matching degrees between the 50 target user behavior events and the user behavior events to be analyzed are obtained through calculation, the target user behavior events corresponding to the first 10 matching degrees in the arrangement results from large to small are selected according to the arrangement from large to small of the obtained 50 matching degrees, and are used as the associated user behavior events corresponding to the user behavior events to be analyzed.
By the method for determining the associated user behavior event according to the matching degree, the matching degree of the determined associated user behavior event and the user behavior event to be analyzed can meet a certain condition, and therefore the associated user behavior event corresponding to the user behavior event to be analyzed can be accurately determined.
Process30121 can be implemented by Process30121A1-30121A 2. As will be explained below.
In the Process30121A1, feature extraction is performed on each target user behavior event and the user behavior event to be analyzed, and user behavior node features corresponding to each target user behavior event and the user behavior event to be analyzed are determined.
In some possible design ideas, when the matching degree is the node matching degree, feature extraction can be performed on each target user behavior event and the user behavior event to be analyzed, so that corresponding user behavior node features are obtained respectively.
In the Process30121A2, the matching degree between each target user behavior event and the user behavior node feature corresponding to the user behavior event to be analyzed is output as the node matching degree between each target user behavior event and the user behavior event to be analyzed.
In some possible design ideas, after user behavior node features corresponding to each target user behavior event and a user behavior event to be analyzed are obtained, matching degrees between the user behavior node features corresponding to each target user behavior event and the user behavior node features corresponding to the user behavior event to be analyzed are calculated, and the calculated matching degrees are used as the node matching degrees between each target user behavior event and the user behavior event to be analyzed.
Based on the embodiment, the node matching degree between each target user behavior event and the user behavior event to be analyzed can be accurately determined, so that the associated user behavior event corresponding to the user behavior event to be analyzed can be accurately determined according to the node matching degree.
Process30121 can also be implemented by Process301021B1-30121B 2. As will be explained below.
In the Process30121B1, according to each target user behavior event and the user behavior event to be analyzed, performing interest point classification based on the interest point classification model, and determining interest point classification information corresponding to each target user behavior event and the user behavior event to be analyzed.
In some possible design ideas, when the matching degree is the interest point matching degree, classifying the interest points according to the target user behavior events based on the interest point classification model, and determining interest point classification information corresponding to the target user behavior events, where the interest point classification information is interest confidence of the target user behavior events at multiple interest points (i.e., classification attribution degrees corresponding to the target user behavior events).
And determining the interest confidence degrees of the user behavior event to be analyzed on the plurality of interest points (namely, the classification attribution degree corresponding to the user behavior event to be analyzed) in the same way.
In some possible design ideas, the interest point classification model is the trained model, and the interest point classification model is obtained by learning according to each user behavior event in the target user behavior event sequence covered by the first relational inference network and the interest point corresponding to each user behavior event. The training process is the same as the training process of the interest point classification model provided above, and is not described herein again.
In the Process30121B2, the matching degree between the interest point classification information corresponding to each target user behavior event and the user behavior event to be analyzed is output as the interest point matching degree between each target user behavior event and the user behavior event to be analyzed.
In some possible design ideas, after the interest point classification information (i.e., the classification attribution degree) corresponding to each target user behavior event and the user behavior event to be analyzed is determined, the matching degree between the classification attribution degree corresponding to each target user behavior event and the classification attribution degree corresponding to the user behavior event to be analyzed is calculated, and the calculated matching degree is used as the interest point matching degree between each target user behavior event and the user behavior event to be analyzed.
Based on the embodiment, the interest point matching degree between each target user behavior event and the user behavior event to be analyzed can be accurately determined, so that the associated user behavior event corresponding to the user behavior event to be analyzed can be accurately determined according to the interest point matching degree.
In the Process302, a first deep learning network model is optimized according to a user behavior event to be analyzed and a plurality of associated user behavior events, and a second deep learning network model is determined, wherein the first deep learning network model is an AI decision branch model based on a relational inference network, such as a graph neural network model.
In some possible design ideas, after the associated user behavior event corresponding to the user behavior event to be analyzed is obtained, the first deep learning network model is optimized according to the user behavior event to be analyzed and the corresponding associated user behavior event, and the second deep learning network model is determined.
Process302 may also be implemented by Process3021 and Process 3022. The following describes Process3021 and Process 3022.
In the Process3021, the user behavior event to be analyzed is loaded into the first relational inference network, and the user behavior event to be analyzed is connected with each associated user behavior event in the first relational inference network, so as to determine the second relational inference network.
In some possible design ideas, user behavior events to be analyzed are loaded into a first relation inference network of a first deep learning network model, the user behavior events to be analyzed are connected with corresponding associated user behavior events in the first relation inference network, and connection dependency relations between the user behavior events to be analyzed and the associated user behavior events are added into a connection dependency relation sequence between the user behavior events covered by the first relation inference network, so that the user behavior events to be analyzed are connected with the associated user behavior events.
And after the user behavior event to be analyzed is connected with each associated user behavior event in the first relational inference network, determining a second relational inference network.
In the Process3022, the second deep learning network model is determined according to the second relationship inference network instead of the current first relationship inference network in the first deep learning network model.
In some possible design ideas, after the second relationship inference network is obtained, the first relationship inference network in the first deep learning network model is replaced by the second relationship inference network, so that the second deep learning network model is obtained.
Since the first deep learning network model comprises the first relational inference network and the first AI decision branch, after the second deep learning network model is obtained by replacing the first relational inference network in the first deep learning network model with the second relational inference network, the second deep learning network model comprises the second relational inference network and the first AI decision branch. The second deep learning network model can be accurately determined based on the foregoing embodiments.
In the Process303, according to the first behavior vector description of the user behavior event to be analyzed and the behavior vector descriptions corresponding to each associated user behavior event in the second deep learning network model, the first traversal feature optimization is performed based on the second deep learning network model, and the second behavior vector description of the user behavior event to be analyzed is determined.
In this embodiment, the behavior vector description may be used to perform feature expression on a related user behavior event, for example, if a user behavior event that a corresponding user performs a user behavior event for a certain e-commerce live broadcast activity is expressed in the user behavior event, such as a series of behaviors of paying attention to, collecting, forwarding, and sharing, the corresponding behavior vector description may express specific field variables (such as field variables of an attention object, an attention reason, and the like) of paying attention to, collecting, forwarding, and sharing.
In some possible design ideas, after the second deep learning network model is determined, first traversal feature optimization is performed based on the second deep learning network model according to the first behavior vector description of the user behavior event to be analyzed and the behavior vector description of each associated user behavior event in the second relation inference network of the second deep learning network model.
In some possible design approaches, the first pass feature optimization is performed through a first AI decision branch based on coverage of the second deep-learning network model.
The determination method of the second behavior vector description of the user behavior event to be analyzed is the same as the determination method of the second behavior vector description of each user behavior event in the Process104, and is not described herein again.
Process303 can also be implemented by Process3031 and Process 3032. Process3031 and Process3032 will be described below.
In Process3031, the following operations are performed by each network element of the first AI decision branch: and performing first feature selection based on penalty items on the behavior vector description of each associated user behavior event corresponding to the current network unit, and determining first feature selection data corresponding to each associated user behavior event.
In some possible design concepts, the feature of the behavior vector description of the user behavior event to be analyzed is optimized by each network element of the first AI decision branch. In each network element of the first AI decision branch, first, a behavior vector description of each associated user behavior event in the current network element is subjected to first feature selection based on a penalty term, and first feature selection data corresponding to each associated user behavior event is determined.
In the Process3032, second feature selection based on a penalty term is performed on the first feature selection data corresponding to each associated user behavior event and the fusion feature sequence of the significance influence coefficient corresponding to the current network element, and the second feature selection data is determined as a second behavior vector description of the user behavior event to be analyzed in the current network element.
In some possible design ideas, after first feature selection data corresponding to each associated user behavior event is determined, a significance influence coefficient corresponding to each associated user behavior event in a current network unit is determined, and the significance influence coefficient corresponding to each associated user behavior event in the current network unit is determined according to behavior vector description of a user behavior event to be analyzed in the current network unit and behavior vector description of the associated user behavior event in the current network unit.
After the significance influence coefficient of each associated user behavior event in the current network unit is determined, calculating first feature selection data corresponding to each associated user behavior event and fusion description of each associated user behavior event in the significance influence coefficient of the current network unit, weighting the fusion description corresponding to each associated user behavior event, and then performing second feature selection based on penalty terms on the weighting result, wherein the determined second feature selection data is second behavior vector description of the user behavior event to be analyzed in the current network unit.
The method can update the behavior vector description of the user behavior event to be analyzed according to the behavior vector description of the associated user behavior event, so that the optimized behavior vector description of the user behavior event to be analyzed can be accurately determined.
Process306-Process308 are also included prior to Process 303. The processes 306-308 will be described below.
In the Process306, feature extraction is performed on the user behavior event to be analyzed and the join event node of the user behavior event to be analyzed, and user behavior node features corresponding to the user behavior event to be analyzed and the join event node of the user behavior event to be analyzed are determined.
In some possible design ideas, feature extraction is firstly carried out on the user behavior event to be analyzed and the connection event node of the user behavior event to be analyzed, and user behavior node features corresponding to the user behavior event to be analyzed and the connection event node of the user behavior event to be analyzed are determined.
In the Process307, significant feature attention processing is performed on the user behavior node features corresponding to the user behavior event to be analyzed and the join event nodes of the user behavior event to be analyzed respectively based on different significant feature attention strategies, and significant behavior vector descriptions corresponding to the user behavior event to be analyzed and the join event nodes of the user behavior event to be analyzed are determined respectively.
In some possible design ideas, after determining user behavior node features corresponding to a user behavior event to be analyzed and a join event node of the user behavior event to be analyzed, respectively performing saliency feature attention processing on the user behavior node features of the user behavior event to be analyzed and the user behavior node features corresponding to the join event node of the user behavior event to be analyzed based on different saliency feature attention strategies, and determining saliency behavior vector description corresponding to the user behavior event to be analyzed and saliency behavior vector description corresponding to the join event node of the user behavior event to be analyzed.
In the Process308, the significant behavior vector descriptions corresponding to the user behavior event to be analyzed and the join event node of the user behavior event to be analyzed are fused, and the fused description is determined as the first behavior vector description of the user behavior event to be analyzed.
In some possible design ideas, after obtaining the significant behavior vector description corresponding to the user behavior event to be analyzed and the significant behavior vector description corresponding to the join event node of the user behavior event to be analyzed, fusing the significant behavior vector description corresponding to the user behavior event to be analyzed and the significant behavior vector description corresponding to the join event node of the user behavior event to be analyzed, and taking the fused description as the first behavior vector description of the user behavior event to be analyzed.
In the above embodiment, the first behavior vector description of the user behavior event to be analyzed is obtained according to the significant behavior vector description corresponding to the join event node of the user behavior event to be analyzed, so that the first behavior vector description of the user behavior event to be analyzed can capture semantic features from the corresponding join event node.
In the Process304, the interest confidence degrees of the user behavior event to be analyzed on the interest points are output according to the second behavior vector description of the user behavior event to be analyzed and the behavior vector descriptions of the interest points corresponding to the second deep learning network model.
In some possible design ideas, after determining a second behavior vector description of the user behavior event to be analyzed, performing interest point decision on the second behavior vector description of the user behavior event to be analyzed and behavior vector descriptions corresponding to multiple interest points in a second relational inference network, and calculating interest confidence of the user behavior event to be analyzed on each interest point.
The plurality of interest points comprise interest points corresponding to user behavior events in a target user behavior event sequence of the second relational inference network.
Process304 may also be implemented by Process3041-Process 3043. The following describes processes 3041 to 3043.
In the Process3041, a fused description of the behavior vector description of each interest point, the first setting coefficient, and the second behavior vector description of the user behavior event to be analyzed is output as a first fused description.
In some possible design ideas, a fusion description among behavior vector descriptions of the interest points, a first setting coefficient and a second behavior vector description of a user behavior event to be analyzed is calculated, and the first fusion description is determined.
In the Process3042, a fusion description of the second setting coefficient and a second behavior vector description of the user behavior event to be analyzed is output as a second fusion description.
In some possible design ideas, after the first fusion description is calculated, a fusion description between a second setting coefficient and a second behavior vector description of the user behavior event to be analyzed is calculated, and the second fusion description is determined.
In the Process3043, an interest decision is performed on the sum of the first fusion description and the second fusion description through a Sigmoid function, and an interest confidence of a user behavior event to be analyzed on each interest point is determined.
In some possible design ideas, after the first fusion description and the second fusion description are obtained, the first fusion description and the second fusion description are weighted, and after a weighting result is determined, the weighting result is processed through a Sigmoid function, so that the interest confidence of the user behavior event to be analyzed on each interest point is obtained.
Based on the embodiment, the interest confidence of the user behavior event to be analyzed on each interest point can be accurately determined, so that the user behavior event to be analyzed can be accurately classified according to the interest confidence.
In the Process305, the interest point with the interest confidence greater than the first set confidence is output as the interest point corresponding to the user behavior event to be analyzed, and the personalized information push service subscribed by the target user is updated based on the interest point corresponding to the user behavior event to be analyzed.
In some possible design ideas, after the interest confidence of the user behavior event to be analyzed on each interest point is determined, the interest point with the interest confidence greater than a first set confidence (for example, 0.7) is used as the interest point corresponding to the user behavior event to be analyzed. On this basis, the related key push words of the interest points corresponding to the user behavior events to be analyzed can be added to the personalized information push service subscribed by the target user, so that the personalized information push of the target user can be performed subsequently based on the updated personalized information push service.
Based on the steps, the embodiment of the application can perform feature optimization on the behavior vector description of the user behavior event to be analyzed according to the behavior vector description of the associated user behavior event of the user behavior event to be analyzed, and can combine the information of the associated user behavior event in the second behavior vector description of the user behavior event to be analyzed to further determine the second behavior vector description of the user behavior event to be analyzed with higher reliability; because the interest point decision is carried out according to the second behavior vector description of the user behavior event to be analyzed, the interest point decision can be carried out according to the accurate behavior vector description of the user behavior event to be analyzed, so that the interest point corresponding to the user behavior event to be analyzed is effectively determined, the reliability of the interest point decision is improved, and finally the matching degree of the personalized information push service and the target user is improved.
In some embodiments, the AI prediction system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.
The processor 110 may perform various suitable actions and processes through a program stored in the machine-readable storage medium 120, such as program instructions related to the big data mining method for personalized information push services 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 by 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 for implementing the big data mining method for personalized information push service 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 big data mining method for personalized information push service according to any of the above embodiments.
It should be understood that, although each operation step is indicated by an arrow in the flowchart of the embodiment of the present application, the implementation order of the steps is not limited to the order indicated by the arrow. In some implementation scenarios of the embodiments of the present application, the implementation steps in the flowcharts may be performed in other sequences as desired, unless explicitly stated otherwise herein. In addition, some or all of the steps in each flowchart may include multiple sub-steps or multiple stages based on an actual implementation scenario. Some or all of these sub-steps or stages may be performed at the same time, or each of these sub-steps or stages may be performed at different times. In a scenario where execution times are different, an execution sequence of the sub-steps or the phases may be flexibly configured according to requirements, which is not limited in the embodiment of the present application.
The foregoing is only an optional implementation manner of a part of implementation scenarios in the present application, and it should be noted that, for those skilled in the art, other similar implementation means based on the technical idea of the present application are also within the protection scope of the embodiments of the present application without departing from the technical idea of the present application.

Claims (7)

1. A big data mining method for personalized information push services, the method comprising:
extracting a user behavior event to be analyzed matched with the current service to be online from a big data acquisition library of a target user, and searching a plurality of associated user behavior events corresponding to the user behavior event to be analyzed from a target user behavior event sequence;
optimizing a first deep learning network model according to the user behavior event to be analyzed and the plurality of associated user behavior events, and determining a second deep learning network model, wherein the first deep learning network model is an AI decision branch model based on a relational inference network;
according to the first behavior vector description of the user behavior event to be analyzed and the behavior vector description corresponding to each associated user behavior event in the second deep learning network model, performing first traversal feature optimization based on the second deep learning network model, and determining the second behavior vector description of the user behavior event to be analyzed;
according to the second behavior vector description of the user behavior event to be analyzed and the behavior vector descriptions of a plurality of interest points corresponding to the second deep learning network model, outputting interest confidence degrees of the user behavior event to be analyzed on the interest points, wherein the interest points comprise the interest points corresponding to the user behavior events in the target user behavior event sequence in the second deep learning network model;
outputting the interest points with the interest confidence degrees larger than a first set confidence degree as the interest points corresponding to the user behavior events to be analyzed, and updating the personalized information push service subscribed by the target user based on the interest points corresponding to the user behavior events to be analyzed;
the behavior vector description is used for carrying out feature expression on related user behavior events;
prior to the performing a first traversal feature optimization based on the second deep-learning network model, the method further comprises:
performing feature extraction on the user behavior event to be analyzed and the linkage event node of the user behavior event to be analyzed, and determining the user behavior node features corresponding to the user behavior event to be analyzed and the linkage event node of the user behavior event to be analyzed;
respectively performing significance characteristic attention processing on the user behavior event to be analyzed and the user behavior node characteristics corresponding to the join event nodes of the user behavior event to be analyzed based on different significance characteristic attention strategies, and determining significance behavior vector descriptions respectively corresponding to the user behavior event to be analyzed and the join event nodes of the user behavior event to be analyzed;
respectively corresponding significant behavior vector descriptions of the user behavior event to be analyzed and the connection event node of the user behavior event to be analyzed are fused, and the fused description is determined as a first behavior vector description of the user behavior event to be analyzed;
the first traversal feature optimization is performed through a first AI decision branch based on coverage of the second deep-learning network model, the first deep-learning network model and the second deep-learning network model both comprising the first AI decision branch;
before performing first traversal feature optimization based on the second deep learning network model according to the first behavior vector description of the user behavior event to be analyzed and the behavior vector description corresponding to each associated user behavior event in the second deep learning network model, the method further includes:
processing and outputting the first AI decision branch based on:
determining a first relational inference network of the first deep learning network model according to the sample behavior data sequence, wherein the first relational inference network comprises the target user behavior event sequence;
according to the first behavior vector description of each user behavior event in the first relational inference network and the behavior vector description of a relational object of each user behavior event in the first relational inference network, performing first traversal feature optimization based on the first AI decision branch substituted into a preset network setting coefficient, and determining a second behavior vector description of each user behavior event;
according to the first behavior vector description of the interest points in the first relational inference network and the behavior vector description of the relational object of each interest point in the first relational inference network, performing second traversal feature optimization based on the first AI decision branch substituted into a preset network setting coefficient, and determining second behavior vector description of the interest points;
wherein the relationship object of each user behavior event in the first relationship inference network comprises:
the associated user behavior events corresponding to the user behavior events and the corresponding interest points; the relationship object of each interest point in the first relationship inference network comprises: forward interest points or backward interest points of each interest point;
outputting interest confidence of each user behavior event on the plurality of interest points according to the second behavior vector description of each user behavior event and the second behavior vector descriptions of the plurality of interest points;
substituting the interest confidence degrees and the actual confidence degrees of the user behavior events on the plurality of interest points into a preset Loss function to determine a decision cost value corresponding to each user behavior event;
determining a global decision cost value according to the decision cost value corresponding to each user behavior event, performing gradient determination of each network setting coefficient in the first AI decision branch substituted into preset network setting coefficients according to the global decision cost value, so as to optimize the network setting coefficients of the first AI decision branch substituted into the preset network setting coefficients, and outputting the first AI decision branch completing training according to the optimized network setting coefficients;
when the first traversal feature optimization is performed based on the first AI decision branch substituted into a preset network setting coefficient according to the first behavior vector description of each user behavior event in the first relational inference network and the behavior vector description of the relational object of each user behavior event in the first relational inference network, the method includes:
when the relation object of each user behavior event in the first relation inference network is the interest point corresponding to the user behavior event, rejecting the interest point corresponding to the user behavior event with preset confidence;
wherein after outputting the first AI decision branch in the first deep-learning network model, the method further comprises:
combining the first AI decision branch and the first relational inference network into a gray scale deep learning network model, verifying the effectiveness of the gray scale deep learning network model based on the following steps:
searching a plurality of associated user behavior events for validity verification corresponding to the gray-scale user behavior events from the target user behavior event sequence;
optimizing the gray scale deep learning network model according to the gray scale user behavior events and the plurality of associated user behavior events for validity verification, and determining the optimized gray scale deep learning network model;
performing first traversal feature optimization based on the first AI decision branch according to the first behavior vector description of the gray-scale user behavior event and the behavior vector description of each associated user behavior event for validity verification in the optimized gray-scale deep learning network model, and determining a second behavior vector description of the gray-scale user behavior event;
outputting interest confidence degrees of the gray-scale user behavior events on the interest points according to the second behavior vector description of the gray-scale user behavior events and the behavior vector descriptions corresponding to the interest points in the optimized gray-scale deep learning network model;
when the difference confidence between the interest confidence and the actual confidence of the gray-scale user behavior event on the corresponding interest point is smaller than a second set confidence, determining that the optimized gray-scale deep learning network model passes validity verification, and taking the optimized gray-scale deep learning network model as the first deep learning network model;
the step of performing first traversal feature optimization based on the second deep learning network model according to the first behavior vector description of the user behavior event to be analyzed and the behavior vector description corresponding to each associated user behavior event in the second deep learning network model, and determining the second behavior vector description of the user behavior event to be analyzed specifically includes:
performing, by each network element of the first AI decision branch:
performing first feature selection based on a penalty term on the behavior vector description of each associated user behavior event corresponding to the current network unit, and determining first feature selection data corresponding to each associated user behavior event;
performing second penalty item-based feature selection on the first feature selection data corresponding to each associated user behavior event and the fusion feature sequence of the significance influence coefficient corresponding to the current network element, determining second feature selection data as a second behavior vector description of the user behavior event to be analyzed in the current network element, and determining the significance influence coefficient corresponding to each associated user behavior event in the current network element according to the behavior vector description of the user behavior event to be analyzed in the current network element and the behavior vector description of the associated user behavior event in the current network element;
the step of outputting interest confidence levels of the user behavior event to be analyzed on the interest points according to the second behavior vector description of the user behavior event to be analyzed and the behavior vector descriptions of the interest points corresponding to the interest points in the second deep learning network model specifically includes:
outputting the behavior vector description of each interest point, a first setting coefficient and the fusion description of the second behavior vector description of the user behavior event to be analyzed as a first fusion description;
outputting a second fusion description of a second set coefficient and a second behavior vector description of the user behavior event to be analyzed as a second fusion description;
and performing interest decision on the sum of the first fusion description and the second fusion description through a Sigmoid function, and determining the interest confidence of the user behavior event to be analyzed on each interest point.
2. The big data mining method for personalized information push services according to claim 1, wherein the step of finding a plurality of associated user behavior events corresponding to a user behavior event to be analyzed from a sequence of target user behavior events specifically comprises:
determining a target user behavior event cluster from the target user behavior event sequence, wherein each target user behavior event in the target user behavior event cluster and a connection event node of the user behavior event to be analyzed share at least one user behavior node data, and the event magnitude of the target user behavior event covered by the target user behavior event cluster is smaller than the event magnitude of the user behavior event covered by the target user behavior event sequence;
and acquiring a plurality of associated user behavior events corresponding to the user behavior event to be analyzed from the target user behavior event cluster, wherein the matching degree of each associated user behavior event and the user behavior event to be analyzed is greater than the set matching degree.
3. The big data mining method for personalized information push service according to claim 2, wherein the step of obtaining a plurality of associated user behavior events corresponding to the user behavior event to be analyzed from the target user behavior event cluster specifically comprises:
determining the matching degree between each target user behavior event and the user behavior event to be analyzed;
and outputting the target user behavior event with the preset order in the sequencing sequence from large to small of the matching degree as the associated user behavior event corresponding to the user behavior event to be analyzed.
4. The big data mining method for the personalized information push service according to claim 3, wherein when the matching degree is a node matching degree, the step of determining the matching degree between each of the target user behavior events and the user behavior event to be analyzed specifically comprises:
respectively extracting characteristics of each target user behavior event and the user behavior event to be analyzed, and determining user behavior node characteristics corresponding to each target user behavior event and the user behavior event to be analyzed;
and outputting the matching degree between the user behavior node characteristics corresponding to each target user behavior event and the user behavior event to be analyzed as the node matching degree between each target user behavior event and the user behavior event to be analyzed.
5. The big data mining method for personalized information push service according to claim 3, wherein when the matching degree is a point of interest matching degree, the step of determining the matching degree between each target user behavior event and the user behavior event to be analyzed specifically comprises:
according to the target user behavior events and the user behavior events to be analyzed, respectively classifying interest points based on an interest point classification model, and determining interest point classification information corresponding to the target user behavior events and the user behavior events to be analyzed, wherein the interest point classification model is obtained by learning according to the user behavior events in the target user behavior event sequence and the interest points corresponding to the user behavior events;
and outputting the matching degree between the target user behavior events and the interest point classification information respectively corresponding to the user behavior events to be analyzed as the interest point matching degree between the target user behavior events and the user behavior events to be analyzed.
6. The big data mining method for personalized information push services according to claim 1, characterized in that the first deep learning network model comprises a first relational inference network comprising:
according to a sample behavior data sequence, pre-learning is carried out to obtain a connection dependency relationship sequence between the user behavior events, a connection dependency relationship sequence between the user behavior events and corresponding interest points, a connection dependency relationship sequence between the interest points, an interest point sequence and the target user behavior event sequence;
the step of optimizing the first deep learning network model according to the user behavior event to be analyzed and the plurality of associated user behavior events and determining the second deep learning network model specifically includes:
loading the user behavior events to be analyzed into the first relational inference network, connecting the user behavior events to be analyzed with the associated user behavior events in the first relational inference network, and determining a second relational inference network;
determining a second deep learning network model according to the second relation inference network to replace the current first relation inference network in the first deep learning network model;
wherein each sample behavior data in the sample behavior data sequence comprises: the user behavior event, a joint event node of the user behavior event and a corresponding interest point of the user behavior event in the joint event node;
before the optimizing the first deep-learning network model as a function of the user behavioral events to be analyzed and the plurality of associated user behavioral events, the method further comprises:
generating the target user behavior event sequence according to a plurality of user behavior events covered by the sample behavior data sequence;
generating the interest point sequence and a connection dependency relationship sequence between the interest points according to the interest points covered by the sample behavior data sequence;
and generating a connection dependency relationship sequence between the user behavior event and the corresponding interest point according to the interest points corresponding to the plurality of user behavior events covered by the sample behavior data sequence.
7. An AI prediction system comprising a processor and a memory for storing a computer program operable on the processor, the processor being configured to perform the big data mining method for personalized information push services of any one of claims 1 to 6 when running the computer program.
CN202210785903.XA 2022-07-06 2022-07-06 Big data mining method and AI prediction system for personalized information push service Active CN114969552B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210785903.XA CN114969552B (en) 2022-07-06 2022-07-06 Big data mining method and AI prediction system for personalized information push service

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210785903.XA CN114969552B (en) 2022-07-06 2022-07-06 Big data mining method and AI prediction system for personalized information push service

Publications (2)

Publication Number Publication Date
CN114969552A CN114969552A (en) 2022-08-30
CN114969552B true CN114969552B (en) 2023-04-18

Family

ID=82966778

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210785903.XA Active CN114969552B (en) 2022-07-06 2022-07-06 Big data mining method and AI prediction system for personalized information push service

Country Status (1)

Country Link
CN (1) CN114969552B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116881541A (en) * 2023-05-05 2023-10-13 厦门亚瑟网络科技有限公司 AI processing method for online searching activity and online service big data system
CN116578334B (en) * 2023-07-12 2023-09-22 苏州盈天地资讯科技有限公司 User online dynamic docking method and system based on configuration

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113821720A (en) * 2021-07-14 2021-12-21 腾讯科技（深圳）有限公司 Behavior prediction method and device and related product

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113472640B (en) * 2021-09-02 2021-12-14 广州市威士丹利智能科技有限公司 Intelligent gateway information processing method and system
CN114117235A (en) * 2021-12-07 2022-03-01 绥化市纯互联网商务有限公司 E-commerce content pushing method adopting artificial intelligence analysis and E-commerce big data system
CN114564648A (en) * 2022-03-03 2022-05-31 赵尚益 Personalized service content optimization method based on big data and artificial intelligence cloud system
CN115329204A (en) * 2022-03-08 2022-11-11 张伟斌 Cloud business service pushing method and pushing processing system based on big data mining
CN114661785A (en) * 2022-03-09 2022-06-24 王体润 User image determining method based on big data mining and cloud computing service system

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113821720A (en) * 2021-07-14 2021-12-21 腾讯科技（深圳）有限公司 Behavior prediction method and device and related product

Also Published As

Publication number Publication date
CN114969552A (en) 2022-08-30

Similar Documents

Publication Publication Date Title
CN114969552B (en) Big data mining method and AI prediction system for personalized information push service
CN112199580B (en) Service processing method and artificial intelligence platform for cloud computing and big data positioning
CN113590900A (en) Sequence recommendation method fusing dynamic knowledge maps
CN111967594A (en) Neural network compression method, device, equipment and storage medium
CN111459988B (en) Automatic design method for machine learning assembly line
US11681922B2 (en) Performing inference and training using sparse neural network
US20220156508A1 (en) Method For Automatically Designing Efficient Hardware-Aware Neural Networks For Visual Recognition Using Knowledge Distillation
CN113361794A (en) Information pushing method and AI (Artificial Intelligence) pushing system based on Internet e-commerce big data
CN113628005A (en) E-commerce session big data based pushing and updating method and big data AI system
CN113343092A (en) Content source recommendation updating method based on big data mining and cloud computing service system
CN114817747B (en) User behavior analysis method based on internet big data and cloud computing service system
CN116362261A (en) User session information analysis method and software product for dealing with digital service items
CN112508177A (en) Network structure searching method and device, electronic equipment and storage medium
CN113642652A (en) Method, device and equipment for generating fusion model
CN114221991B (en) Session recommendation feedback processing method based on big data and deep learning service system
CN113705402A (en) Video behavior prediction method, system, electronic device and storage medium
CN114896502B (en) User demand decision method applying AI and big data analysis and Internet system
CN112508178A (en) Neural network structure searching method and device, electronic equipment and storage medium
CN114978765B (en) Big data processing method for information attack defense and AI attack defense system
CN115329205B (en) Big data mining method and AI recommendation system for service of personalized push service
CN116186016A (en) Training data cleaning method and system for AI training task
CN114625961A (en) Intelligent online service pushing method applied to big data and big data server
CN114443970A (en) Artificial intelligence and big data based digital content pushing method and AI system
CN115720203A (en) Internet application analysis method and internet application system based on artificial intelligence
Din et al. Gnn approach for software reliability

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
TA01 Transfer of patent application right

Effective date of registration: 20221209

Address after: 517 Liaocheng Road, Huaiyin District, Jinan, Shandong 250000

Applicant after: Meng Shujun

Address before: Room 1011, North District, Huaiyin District, Jinan City, Shandong Province, 250000

Applicant before: Jinan Bangjie Electronic Technology Co.,Ltd.

TA01 Transfer of patent application right
TA01 Transfer of patent application right

Effective date of registration: 20230309

Address after: 101300 8 Jixiang street, houshayu Jixiang Industrial Park, Shunyi District, Beijing

Applicant after: ZHONGTU DIGITAL TECHNOLOGY (BEIJING) Co.,Ltd.

Address before: 517 Liaocheng Road, Huaiyin District, Jinan, Shandong 250000

Applicant before: Meng Shujun

TA01 Transfer of patent application right
GR01 Patent grant
GR01 Patent grant