CN112182382B - Data processing method, electronic device, and medium - Google Patents

Data processing method, electronic device, and medium Download PDF

Info

Publication number
CN112182382B
CN112182382B CN202011044261.5A CN202011044261A CN112182382B CN 112182382 B CN112182382 B CN 112182382B CN 202011044261 A CN202011044261 A CN 202011044261A CN 112182382 B CN112182382 B CN 112182382B
Authority
CN
China
Prior art keywords
click
channel
information
vector
probability
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
CN202011044261.5A
Other languages
Chinese (zh)
Other versions
CN112182382A (en
Inventor
蔡文渊
骆玮璐
潘翔
张坤坤
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.)
Shanghai Hipu Intelligent Information Technology Co ltd
Original Assignee
Shanghai Hipu Intelligent Information Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shanghai Hipu Intelligent Information Technology Co ltd filed Critical Shanghai Hipu Intelligent Information Technology Co ltd
Priority to CN202011044261.5A priority Critical patent/CN112182382B/en
Publication of CN112182382A publication Critical patent/CN112182382A/en
Application granted granted Critical
Publication of CN112182382B publication Critical patent/CN112182382B/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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • 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
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/50—Network services
    • H04L67/55—Push-based network services

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention relates to a data processing method, an electronic device and a medium, wherein the method comprises the following steps: step S1, acquiring information browsing data; step S2, setting initial values of model parameters for the data processing model; step S3, information coding processing is carried out on the information browsing characteristic vector sequence to obtain a presenting intermediate characteristic sequence
Figure DDA0002707525900000011
Step S4, pair
Figure DDA0002707525900000012
Decoding to obtain the click intermediate characteristic sequence
Figure DDA0002707525900000013
Based on SiObtaining xjCorresponding predicted click probability
Figure DDA0002707525900000014
Based on
Figure DDA0002707525900000015
And actual click probability zjDetermining a first loss function Lc(ii) a Step S5, obtaining the probability of the predicted information push result
Figure DDA0002707525900000016
And a second loss function Lv(ii) a Step S6 according to LcAnd LvJudging whether the model parameters need to be adjusted or not, and if so, based on LcAnd LvAdjusting the model parameters, returning to execute the step S3, otherwise, executing the step S7; step S7, obtaining A corresponding to the current model1、A2Based on A1And A2And determining the corresponding weight of each channel. The method and the device can efficiently and accurately acquire the weights of different channels for the push result data.

Description

Data processing method, electronic device, and medium
Technical Field
The present invention relates to the field of data processing technologies, and in particular, to a data processing method, an electronic device, and a medium.
Background
The field of data processing is an important branch of the computer field. In the field of computers, data may include a variety of text data, image data, audio data, video data, and the like, depending on the manner of presentation; depending on the manner of storage, the data may be stored to a database, text file, a file of a particular format (e.g.,. doc/. xls), etc.; depending on the way the data is formed, static data and dynamic data, especially time-varying data, such as information push data acquired by a network device like a router switch, device LBS data acquired by GPS or beidou, etc. may be included. The data processing may be "forward processing", such as processing the image with an algorithm to make it clearer, or "reverse processing", such as separating the plurality of original images used by the composite image by a computer program given the clarity of the composite image. For another example, in an information push scenario, information may be generally pushed through multiple channels, when a target channel is selected to push information, weights of different channels for push result data need to be obtained reversely according to existing information push result data, and the target channel is selected based on the weight of each channel for the push result data.
"reverse processing" of static data is relatively easy. However, because most of the dynamic data is data changing with time, the influence of the time dimension on the association relationship needs to be considered, not only can more computer storage resources, retrieval resources and operation processing resources be occupied, but also the processing precision is not ideal, and with the arrival of the intelligent era, the information push application is more and more extensive, so how to efficiently and accurately obtain the weights of different channels on the push result data through the 'reverse processing' of the data becomes a technical problem to be solved urgently.
Disclosure of Invention
The invention aims to provide a data processing method, electronic equipment and a medium, which can efficiently and accurately acquire the weights of different channels for pushing result data.
According to a first aspect of the present invention, there is provided a data processing method comprising:
step S1, acquiring n pieces of information browsing data { G ] from preset database1,G2...Gn}，
Figure GDA0003153430060000021
GiTotal miAn information browsing feature vector of
Figure GDA0003153430060000022
Figure GDA0003153430060000023
Arranging according to the corresponding time stamp sequence to form an information browsing characteristic vector sequence, XjRepresents GiJ equals 1i,2i,3i...miThe information browsing feature vector XjIncluding a presentation feature vector xjAnd actual click probability zj,yiIs GiActual information push result probability;
step S2, setting initial values of model parameters for a preset data processing model, wherein the model parameters comprise a channel presenting weight vector A1Channel click weight vector A2Presenting an initial value h of the intermediate feature vector0Clicking the initial value s of the intermediate feature vector0A balance coefficient λ, wherein the channel exhibits a weight vector A1Is used for representing the presentation weight of an information push channel, and a channel click weight vector A2The element(s) of (2) is used to represent the click weight of the information push channel;
step S3 to
Figure GDA0003153430060000024
As input to the model, based on presenting an intermediate feature vector initial value h0Browsing a sequence of feature vectors for information
Figure GDA0003153430060000025
Carrying out information coding processing to obtain a sequence presenting intermediate characteristics
Figure GDA0003153430060000026
Step S4, based on the initial value S of the click middle feature vector0To pair
Figure GDA0003153430060000027
Decoding to obtain the click intermediate characteristic sequence
Figure GDA0003153430060000028
Based on SiObtaining xjCorresponding predicted click probability
Figure GDA0003153430060000029
All the predicted click probabilities corresponding to the n pieces of information browsing data
Figure GDA00031534300600000210
And actual click probability zjDetermining a first loss function Lc；
Step S5, based on the presenting middle characteristic sequence HiChannel presentation weight vector A1Clicking on the intermediate feature sequence SiChannel click weight vector A2Determination of GiCorresponding predicted information push result probability
Figure GDA00031534300600000211
Predicting information push result probability based on the n pieces of information browsing data
Figure GDA00031534300600000212
And actual information push result probability yiDetermining a second loss function Lv；
Step S6 according to LcAnd LvJudging whether the model parameters need to be adjusted or not, and if so, based on LcAnd LvAdjusting the model parameters, returning to execute the step S3, otherwise, executing the step S7;
step S7, obtaining a channel presenting weight vector A corresponding to the current model1Channel click weight vector A2Based on A1And A2And determining the corresponding weight of each channel.
According to a second aspect of the present invention, there is provided an electronic apparatus comprising: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being arranged to perform the method of the first aspect of the invention.
According to a third aspect of the invention, there is provided a computer readable storage medium, the computer instructions being for performing the method of the first aspect of the invention.
Compared with the prior art, the invention has obvious advantages and beneficial effects. By means of the technical scheme, the data processing method, the electronic equipment and the medium provided by the invention can achieve considerable technical progress and practicability, have industrial wide utilization value and at least have the following advantages:
the method can efficiently and accurately acquire the weight of the pushed result data of different channels based on multiple pieces of information browsing data through the reverse processing of the data.
The foregoing description is only an overview of the technical solutions of the present invention, and in order to make the technical means of the present invention more clearly understood, the present invention may be implemented in accordance with the content of the description, and in order to make the above and other objects, features, and advantages of the present invention more clearly understood, the following preferred embodiments are described in detail with reference to the accompanying drawings.
Drawings
Fig. 1 is a schematic diagram of a data processing method according to an embodiment of the present invention.
Detailed Description
To further illustrate the technical means and effects of the present invention adopted to achieve the predetermined objects, the following detailed description of the embodiments and effects of a data processing method, an electronic device and a medium according to the present invention will be provided with reference to the accompanying drawings and preferred embodiments.
An embodiment of the present invention provides a data processing method, as shown in fig. 1, including the following steps:
step S1, acquiring n pieces of information browsing data { G ] from preset database1,G2...GnAs the training set, the training set is selected,
Figure GDA0003153430060000031
Gitotal miAn information browsing feature vector of
Figure GDA0003153430060000032
Arranging according to the corresponding time stamp sequence to form an information browsing characteristic sequence XjRepresents GiJ equals 1i,2i,3i...miThe information browsing feature vector XjIncluding a presentation feature vector xjAnd actual click probability zj,yiIs GiActual information push result probability;
each information browsing characteristic vector corresponds to a channel vector, and the channel is a transmission path for pushing information. The presentation feature vector is used to represent the information feature presented by the information channel vector, and the click probability is used to represent the probability that the information feature presented by the channel is clicked, it can be understood that the actual click probability is 0 or 1, the information presented by the channel is not clicked, the actual probability is 0, the information presented by the channel is clicked, and the actual probability is 1.
Step S2, setting initial values of model parameters for a preset data processing model, wherein the model parameters comprise a channel presenting weight vector A1Channel click weight vector A2Presenting an initial value h of the intermediate feature vector0Clicking the initial value s of the intermediate feature vector0Coefficient of equilibrium lambda0Wherein the channel presents a weight vector A1Is used for representing the presentation weight of an information push channel, and a channel click weight vector A2Is used for representing information push-throughClick weight of lane;
step S3 to
Figure GDA0003153430060000041
As input to the model, based on presenting an intermediate feature vector initial value h0Browsing a sequence of feature vectors for information
Figure GDA0003153430060000042
Carrying out information coding processing to obtain a sequence presenting intermediate characteristics
Figure GDA0003153430060000043
It will be appreciated that the information browsing feature vector X is enteredjIncluding a presentation feature vector xjAnd actual click probability zjHowever, when encoding processing is performed, only the browsing feature vector sequence needs to be processed
Figure GDA0003153430060000044
Without processing the actual click probability zjAfter proceeding to step S4, the actual click probability z is usedjAnd predicting click probability
Figure GDA0003153430060000045
Determining a first loss function Lc。
Step S4, based on the initial value S of the click middle feature vector0To pair
Figure GDA0003153430060000046
Decoding to obtain the click intermediate characteristic sequence
Figure GDA0003153430060000047
Based on SiObtaining xjCorresponding predicted click probability
Figure GDA0003153430060000048
All the predicted click probabilities corresponding to the n pieces of information browsing data
Figure GDA0003153430060000049
And actual click probability zjDetermining a first loss function Lc；
Step S5, based on the presenting middle characteristic sequence HiChannel presentation weight vector A1Clicking on the intermediate feature sequence SiChannel click weight vector A2Determination of GiCorresponding predicted information push result probability
Figure GDA00031534300600000410
Predicting information push result probability based on the n pieces of information browsing data
Figure GDA00031534300600000411
And actual information push result probability yiDetermining a second loss function Lv；
Wherein a weight vector A is presented1Each element of (1) represents
Figure GDA00031534300600000412
The importance degree of the presentation information corresponding to the channel vector to the information pushing result reaches a preset target; click weight vector A2Each element of (A) represents
Figure GDA00031534300600000413
And the click information corresponding to the channel vector achieves the importance degree of a preset target for the information push result.
Step S6 according to LcAnd LvJudging whether the model parameters need to be adjusted or not, and if so, based on LcAnd LvAdjusting the model parameters, returning to execute the step S3, otherwise, executing the step S7;
model accuracy is adapted to the preset requirements by adjusting model parameters, e.g. by adjusting the rendering weight vector A1Channel click weight vector A2Fitting the real information browsing characteristic sequence to finally obtain a presentation weight vector A close to the actual situation1Channel click weight vector A2. The accuracy of the current model is judged and the model parameters are adjusted through the loss function, so that the method has high reliability and high processing efficiency.
Step S7, obtaining a channel presenting weight vector A corresponding to the current model1Channel click weight vector A2Based on A1And A2And determining the corresponding weight of each channel.
As a variation of the above embodiment, step S6 may alternatively be performed by acquiring q pieces of information browsing data { G } from a preset database1,G2...GnForming a test set, testing the current data processing model to obtain the AUC value of the current model, judging whether the model parameters need to be adjusted according to the AUC (area Under Current) value, and if so, based on LcAnd LvAnd adjusting the model parameters, returning to the step of 3, otherwise, executing the step of 7, wherein the AUC value represents the area size below the ROC curve, and is used for measuring the model accuracy. Therefore, the problem that the accuracy of the model is low due to the fact that the dependence on the data of the test set is too large because the training is carried out only through the test set can be avoided. It is understood that, in order to further improve the model training accuracy, the judgment of the current model accuracy through set detection and the judgment of the current model accuracy through a loss function can be combined.
According to the embodiment of the invention, the weight of different channels to the pushed result data can be obtained based on two characteristics of information presentation and information clicking, and the reliability and accuracy of the obtained result are improved.
As an example, the method further comprises: step S10, constructing the preset database, specifically including:
step S101, information presentation data and information pushing result data of different terminals are obtained, wherein the information presentation data comprise presentation information IDs, presentation equipment IDs, channel IDs, information click data and presentation time stamps, the actual information pushing result probability is 0 or 1, the actual information pushing result probability is 0 to indicate that the preset information pushing target is not reached, and the actual information pushing result probability is 1 to indicate that the preset information pushing target is reached;
the terminal may be physically implemented as a mobile device such as a smart phone, PAD, etc. capable of installing an application (e.g., APP).
Step S102, acquiring information presentation data and information push result data corresponding to each user ID according to the incidence relation between the user ID and the equipment ID;
it is understood that one user ID may correspond to a plurality of device IDs, and the information browsed by the unified user on different devices may be collected through step S102, and corresponding information browsing data may be obtained through step S103
And step S103, forming a time sequence by the information presentation data corresponding to each information pushing result according to the presentation time stamp according to the time sequence, storing the time sequence in the record of the database, and constructing the preset database.
In the model training process, the proportion setting of the positive sample and the negative sample has a direct influence on the model accuracy, and the proportion is too high or too low, which can reduce the model accuracy and affect the data processing result, so that the proportion of the positive sample and the negative sample is set in a reasonable range to improve the model training accuracy.
In the above modified embodiment, the selected number of the test set data may affect the accuracy and the training efficiency of the model training, and if the test set selection data is too much, the model training efficiency may be reduced, and if the test set selection data is too little, the accuracy of the model training may be reduced, so the ratio of the training set data to the test set data may be set to (3:1, 5:1), preferably, the ratio of the number of the training set data to the number of the test set data is 4:1, it should be noted that the training set data and the test set data are different information browsing data, so that the dependency of the model on a large number of the same data may be avoided, the accuracy of the model training may be improved, and the accuracy of the data processing result may be improved.
As an example, the step S3 includes:
step S301, based on the presenting middle characteristic vector h of the moment before the information browsing characteristic vector sequencej-1And the present feature vector x at the current timejCarrying out information coding processing, and determining the presenting intermediate characteristic vector at the current moment:
hj＝fe(xj,hj-1)
wherein f ise() For a predetermined coding function, as an example, fe() Is an encoding function of a long short term memory network (LSTM).
Step S302, based on presenting the intermediate feature vector hjDetermining a sequence of presenting intermediate features
Figure GDA0003153430060000071
As an embodiment, in step S4, the initial value S is based on the click middle feature vector0To pair
Figure GDA0003153430060000072
Decoding to obtain the click intermediate characteristic sequence
Figure GDA0003153430060000073
The method comprises the following steps:
step S401, based on the click intermediate feature vector S of the previous moment of the information browsing feature vector sequencej-1And predicting click probability
Figure GDA0003153430060000074
And
Figure GDA0003153430060000075
determining a click middle feature vector at the current moment:
Figure GDA0003153430060000076
wherein f isd() For a predetermined decoding function, fd() Is a decoding function of a long short term memory network (LSTM).
Step S402, based on the click intermediate feature vector SjDetermining a click middle feature sequence
Figure GDA0003153430060000077
In step S4, based on SiObtaining xjCorresponding predicted click probability
Figure GDA0003153430060000078
The method comprises the following steps:
step S411 based on SiAnd
Figure GDA0003153430060000079
obtaining xjCorresponding predicted click probability:
Figure GDA00031534300600000710
wherein, g () is a preset perceptron model function, and the core formula is that the activation equation is
Figure GDA00031534300600000711
In the process, the coding function adopts the presenting intermediate characteristic vector h at the last momentj-1And the present feature vector x at the current timejDetermining a present intermediate feature vector h for a current time instantj(ii) a Decoding function is based on click intermediate characteristic vector s of one moment on information browsing characteristic vector sequencej-1And predicting click probability
Figure GDA00031534300600000712
And
Figure GDA00031534300600000713
determining a click intermediate feature vector at the current moment; based on SiAnd
Figure GDA00031534300600000714
obtaining xjCorresponding predicted click probability
Figure GDA00031534300600000715
The relation between two information browsing characteristic vectors in adjacent neighbors can be dynamically adjusted to improve the efficiency and the precision of model training, thereby improving the efficiency and the precision of data processing
As an embodiment, in step S4, a first loss function L is determined based on all the predicted click probabilities and all the actual click probabilities corresponding to the n pieces of information browsing datacThe method comprises the following steps:
Figure GDA00031534300600000716
as an embodiment, in the step S5, the intermediate feature sequence H based on the presentation is describediChannel presentation weight vector A1Clicking on the intermediate feature sequence SiChannel click weight vector A2Determination of GiCorresponding predicted information push result probability
Figure GDA00031534300600000717
The method comprises the following steps:
step S501, based on presenting the intermediate characteristic sequence HiChannel presentation weight vector A1Acquiring a presentation parameter:
Figure GDA0003153430060000081
step S502, based on the click intermediate characteristic sequence SiChannel click weight vector A2Acquiring click parameters:
Figure GDA0003153430060000082
step S503, obtaining the probability of the predicted information push result based on the presentation parameter and the click parameter
Figure GDA0003153430060000083
Figure GDA0003153430060000084
As an embodiment, the probability of the result of the predicted information pushing based on the n pieces of information browsing data
Figure GDA0003153430060000085
And actual information push result probability yiDetermining a second loss function LvThe method comprises the following steps:
Figure GDA0003153430060000086
as an embodiment, the step S6 includes:
step S601, continuously acquiring M LcAnd LvIf M consecutive LcAnd LvIf the model parameters are gradually reduced and the variation amplitude is smaller than a preset variation threshold value, judging that the model parameters do not need to be adjusted, otherwise, entering the step S602 to adjust the model parameters;
step S602 based on LcAnd LvThe adjustment magnitude of the model parameter is determined, the model parameter is adjusted based on the adjustment magnitude, and then the execution returns to step S3.
Based on LcAnd LvThe judgment of the model parameters can quickly and accurately judge whether the current model needs to be adjusted or not, and the adjustment range of the model parameters is determined, so that the accuracy and efficiency of model training are improved, and the accuracy and efficiency of data processing are improved.
In the above modified embodiment, the step S6 includes:
step S611, inputting the information browsing characteristic vector sequence in each piece of information browsing data in the test set into the current data processing model to obtain the corresponding predicted information pushing result probability;
step S612, determining a model AUC value corresponding to each piece of information browsing data based on the predicted information pushing result probability and the actual information pushing result probability corresponding to each piece of information browsing data;
step S613, obtaining variation amplitude of AUC values of the R continuous models, and if the variation amplitude is smaller than a preset variation amplitude threshold, determining that the variation amplitude is based on LcAnd LvAnd adjusting the model parameters, returning to execute the step S3, and otherwise, executing the step S7.
The accuracy of the current model is judged by constructing the test set, so that the problem that the accuracy of the model is low due to overlarge data dependence on the test set caused by training only through the test set can be avoided, the accuracy of the model training result is improved, and the accuracy of the data processing result is improved.
As an embodiment, in step S7, a channel presentation weight vector a corresponding to the current model is obtained1Channel click weight vector A2Based on A1And A2Determining the weight corresponding to each channel:
Attrk＝(1-λd)A1k+λdA2k
wherein, AttrkIs the weight of the k channel, A1kIs A1The channel of the k-th channel presents a weight value, A2kIs represented by A2Channel click weight value, lambda, of the kth channeldK is the balance coefficient of the current model, K is 1,2,3.
An embodiment of the present invention further provides an electronic device, including: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions configured to perform a method according to an embodiment of the invention.
The embodiment of the invention also provides a computer-readable storage medium, and the computer instructions are used for executing the method of the embodiment of the invention.
In a specific application scenario, the information can be advertisement information, and the weight of different information channels for the pushed advertisement information to reach the preset target can be obtained through the embodiment of the invention, so that powerful reference is provided for the next advertisement pushing.
Although the present invention has been described with reference to a preferred embodiment, it should be understood that various changes, substitutions and alterations can be made herein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (9)

1. A data processing method, comprising:
step S1, acquiring n pieces of information browsing data { G ] from preset database1,G2...Gn}，
Figure FDA0003153430050000011
GiTotal miAn information browsing feature vector of
Figure FDA0003153430050000012
Figure FDA0003153430050000013
Figure FDA0003153430050000014
Arranging according to the corresponding time stamp sequence to form information browsing featuresSyndrome vector sequence, XjRepresents GiJ equals 1i,2i,3i...miThe information browsing feature vector XjIncluding a presentation feature vector xjAnd actual click probability zj,yiIs GiActual information push result probability;
step S2, setting initial values of model parameters for a preset data processing model, wherein the model parameters comprise a channel presenting weight vector A1Channel click weight vector A2Presenting an initial value h of the intermediate feature vector0Clicking the initial value s of the intermediate feature vector0A balance coefficient λ, wherein the channel exhibits a weight vector A1Is used for representing the presentation weight of an information push channel, and a channel click weight vector A2The element(s) of (2) is used to represent the click weight of the information push channel;
step S3 to
Figure FDA0003153430050000015
As input to the model, based on presenting an intermediate feature vector initial value h0Browsing a sequence of feature vectors for information
Figure FDA0003153430050000016
Carrying out information coding processing to obtain a sequence presenting intermediate characteristics
Figure FDA0003153430050000017
Step S4, based on the initial value S of the click middle feature vector0To pair
Figure FDA0003153430050000018
Decoding to obtain the click intermediate characteristic sequence
Figure FDA0003153430050000019
Based on SiObtaining xjCorresponding predicted click probability
Figure FDA00031534300500000110
All the predicted click probabilities corresponding to the n pieces of information browsing data
Figure FDA00031534300500000111
And actual click probability zjDetermining a first loss function Lc；
In step S4, the initial value S is based on the click middle feature vector0To pair
Figure FDA00031534300500000112
Decoding to obtain the click intermediate characteristic sequence
Figure FDA00031534300500000113
The method comprises the following steps:
step S401, based on the click intermediate feature vector S of the previous moment of the information browsing feature vector sequencej-1And predicting click probability
Figure FDA00031534300500000114
And
Figure FDA00031534300500000115
determining click middle feature vector s of current momentj：
Figure FDA00031534300500000116
Wherein f isd() Is a preset decoding function;
step S402, based on the click intermediate feature vector SjDetermining a click middle feature sequence
Figure FDA00031534300500000117
Step S5, based on the presenting middle characteristic sequence HiChannel presentation weight vector A1Clicking on the intermediate feature sequence SiChannel click weight vector A2Determination of GiCorresponding predicted information push result probability
Figure FDA00031534300500000118
Predicting information push result probability based on the n pieces of information browsing data
Figure FDA00031534300500000119
And actual information push result probability yiDetermining a second loss function Lv；
Step S6 according to LcAnd LvJudging whether the model parameters need to be adjusted or not, and if so, based on LcAnd LvAdjusting the model parameters, returning to execute the step S3, otherwise, executing the step S7;
step S7, obtaining a channel presenting weight vector A corresponding to the current model1Channel click weight vector A2Based on A1And A2And determining the corresponding weight of each channel.
2. The method of claim 1,
the step S3 includes:
step S301, based on the presenting middle characteristic vector h of the moment before the information browsing characteristic vector sequencej-1And the present feature vector x at the current timejPerforming information coding processing to determine the present intermediate characteristic vector h at the current momentj：
hj＝fe(xj,hj-1)
Wherein f ise() Is a preset coding function;
step S302, based on presenting the intermediate feature vector hjDetermining a sequence of presenting intermediate features
Figure FDA0003153430050000021
3. The method of claim 2,
in step S4, based on SiObtaining xjCorresponding predicted click probability
Figure FDA0003153430050000022
The method comprises the following steps:
step S411 based on SiAnd
Figure FDA0003153430050000023
obtaining xjCorresponding predicted click probability:
Figure FDA0003153430050000024
wherein g () is a preset perceptron model function.
4. The method of claim 3,
in step S4, a first loss function L is determined based on all the predicted click probabilities and the actual click probabilities corresponding to the n pieces of information browsing datacThe method comprises the following steps:
Figure FDA0003153430050000025
5. the method of claim 1,
in the step S5, the intermediate feature sequence H based on the presentationiChannel presentation weight vector A1Clicking on the intermediate feature sequence SiChannel click weight vector A2Determination of GiCorresponding predicted information push result probability
Figure FDA0003153430050000026
The method comprises the following steps:
step S501, based on presenting the intermediate characteristic sequence HiChannel presentation weight vector A1Obtaining a presentation parameter C1：
Figure FDA0003153430050000027
Step S502, based on the click intermediate characteristic sequence SiChannel click weight vector A2Obtaining click parameter C2：
Figure FDA0003153430050000031
Step S503, obtaining the probability of the predicted information push result based on the presentation parameter and the click parameter
Figure FDA0003153430050000032
Figure FDA0003153430050000033
6. The method of claim 5,
the probability of the result of the predicted information push based on the n pieces of information browsing data
Figure FDA0003153430050000034
And actual information push result probability yiDetermining a second loss function LvThe method comprises the following steps:
Figure FDA0003153430050000035
7. the method of claim 5,
the step S6 includes:
step S601, continuously acquiring M LcAnd LvIf M consecutive LcAnd LvIf the model parameters are gradually reduced and the variation amplitude is smaller than a preset variation threshold value, judging that the model parameters do not need to be adjusted, otherwise, entering the step S602 to adjust the model parameters;
step S602 based on LcAnd LvThe adjustment magnitude of the model parameter is determined, the model parameter is adjusted based on the adjustment magnitude, and then the execution returns to step S3.
8. An electronic device, comprising:
at least one processor;
and a memory communicatively coupled to the at least one processor;
wherein the memory stores instructions executable by the at least one processor, the instructions being arranged to perform the method of any of the preceding claims 1-7.
9. A computer-readable storage medium having stored thereon computer-executable instructions for performing the method of any of the preceding claims 1-7.
CN202011044261.5A 2020-09-28 2020-09-28 Data processing method, electronic device, and medium Active CN112182382B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011044261.5A CN112182382B (en) 2020-09-28 2020-09-28 Data processing method, electronic device, and medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011044261.5A CN112182382B (en) 2020-09-28 2020-09-28 Data processing method, electronic device, and medium

Publications (2)

Publication Number Publication Date
CN112182382A CN112182382A (en) 2021-01-05
CN112182382B true CN112182382B (en) 2021-08-24

Family

ID=73946414

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011044261.5A Active CN112182382B (en) 2020-09-28 2020-09-28 Data processing method, electronic device, and medium

Country Status (1)

Country Link
CN (1) CN112182382B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114227701B (en) * 2022-02-25 2022-05-10 科大智能物联技术股份有限公司 Robot fault prediction method based on production data

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103714084B (en) * 2012-10-08 2018-04-03 腾讯科技（深圳）有限公司 The method and apparatus of recommendation information
CN110378434A (en) * 2019-07-24 2019-10-25 腾讯科技（深圳）有限公司 Training method, recommended method, device and the electronic equipment of clicking rate prediction model
CN110991742A (en) * 2019-12-04 2020-04-10 清华大学 Social network information forwarding probability prediction method and system

Also Published As

Publication number Publication date
CN112182382A (en) 2021-01-05

Similar Documents

Publication Publication Date Title
US11531867B2 (en) User behavior prediction method and apparatus, and behavior prediction model training method and apparatus
CN109840589B (en) Method and device for operating convolutional neural network on FPGA
CN110909182B (en) Multimedia resource searching method, device, computer equipment and storage medium
EP2657884B1 (en) Identifying multimedia objects based on multimedia fingerprint
WO2023138188A1 (en) Feature fusion model training method and apparatus, sample retrieval method and apparatus, and computer device
US20230004608A1 (en) Method for content recommendation and device
CN110390056B (en) Big data processing method, device and equipment and readable storage medium
US10129504B2 (en) Method and system for measuring quality of video call
WO2021179631A1 (en) Convolutional neural network model compression method, apparatus and device, and storage medium
EP3912099A1 (en) Compound model scaling for neural networks
US20170169330A1 (en) Method and Electronic Device for Displaying Play Content in Smart Television
CN112231516B (en) Training method of video abstract generation model, video abstract generation method and device
CN112182382B (en) Data processing method, electronic device, and medium
CN111814759A (en) Method and device for acquiring face quality label value, server and storage medium
CN112182379B (en) Data processing method, electronic device, and medium
KR20210090706A (en) Sort
CN112446461A (en) Neural network model training method and device
CN117688390A (en) Content matching method, apparatus, computer device, storage medium, and program product
CN111612783B (en) Data quality assessment method and system
CN107508705A (en) The resource tree constructing method and computing device of a kind of HTTP elements
CN113298083A (en) Data processing method and device
CN113111273A (en) Information recommendation method and device, electronic equipment and storage medium
CN112182381B (en) Data processing method, electronic device, and medium
CN109871487B (en) News recall method and system
CN112532692B (en) Information pushing method and device and storage medium

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
CB02 Change of applicant information

Address after: Room 401, 2-6 / F, No.5 Lane 541, Wenshui East Road, Hongkou District, Shanghai 200434

Applicant after: Shanghai hipu Intelligent Information Technology Co.,Ltd.

Address before: Room 401, 2-6 / F, No.5 Lane 541, Wenshui East Road, Hongkou District, Shanghai 200434

Applicant before: Shanghai Honglu Data Technology Co.,Ltd.

CB02 Change of applicant information
GR01 Patent grant
GR01 Patent grant
PE01 Entry into force of the registration of the contract for pledge of patent right

Denomination of invention: Data processing methods, electronic equipment and media

Effective date of registration: 20230210

Granted publication date: 20210824

Pledgee: Industrial Bank Co.,Ltd. Shanghai Hongkou sub branch

Pledgor: Shanghai hipu Intelligent Information Technology Co.,Ltd.

Registration number: Y2023310000027

PE01 Entry into force of the registration of the contract for pledge of patent right
PC01 Cancellation of the registration of the contract for pledge of patent right

Granted publication date: 20210824

Pledgee: Industrial Bank Co.,Ltd. Shanghai Hongkou sub branch

Pledgor: Shanghai hipu Intelligent Information Technology Co.,Ltd.

Registration number: Y2023310000027

PC01 Cancellation of the registration of the contract for pledge of patent right
PE01 Entry into force of the registration of the contract for pledge of patent right

Denomination of invention: Data processing methods, electronic devices and media

Granted publication date: 20210824

Pledgee: Industrial Bank Co.,Ltd. Shanghai Hongkou sub branch

Pledgor: Shanghai hipu Intelligent Information Technology Co.,Ltd.

Registration number: Y2024310000213

PE01 Entry into force of the registration of the contract for pledge of patent right